Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
I
A HYBRID APPROACH FOR EFFICIENT
COMPRESSION OF MULTIMODAL MEDICAL
IMAGES
THESIS
Submitted by
B. PERUMAL (Reg. No. 201111204)
In partial fulfillment for the award of the degree
of
DOCTOR OF PHILOSOPHY
IN ELECTRONICS & INSTRUMENTATION ENGINEERING
DEPARTMENT OF ELECTRONICS & INSTRUMENTATION
ENGINEERING
KALASALINGAM UNIVERSITY (KALASALINGAM ACADEMY OF RESEARCH AND EDUCATION)
ANAND NAGAR, KRISHNANKOIL – 626 126 JULY 2016
CERTIFICATE
This is to certify that all corrections and suggestions pointed out by the Indian/Foreign
Examiner(s) are incorporated in the Thesis titled “A Hybrid Approach for Efficient
Compression of Multimodal Medical Images” submitted by Mr.B.Perumal (Reg.No.
201111204)
SUPERVISOR
Place: Krishnankoil
Date: 17.09.2016
II
KALASALINGAM UNIVERSITY
(Kalasalingam Academy of Research and Education)
ANAND NAGAR, KRISHNANKOIL-626 126
BONAFIDE CERTIFICATE
Certified that this thesis titled, “A HYBRID APPROACH FOR
EFFICIENT COMPRESSION OF MULTIMODAL MEDICAL
IMAGES” is the bonafide work of Mr. PERUMAL.B, who carried out the
research under my supervision. Certified further that, to the best of my
knowledge the work reported herein does not form a part of any other thesis or
dissertation on the basis of which a degree or award was conferred on an earlier
occasion on this or any other scholar.
Dr. M. PALLIKONDA RAJASEKARAN SUPERVISOR, Professor, Department of Electronics and Communication Engineering, Kalasalingam University, Anand Nagar, Krishnankoil – 626 126
III
ABSTRACT
Medical imaging plays a vital role in giving medical assistance that
gives diagnostic information for clinical management of the patients and offering
suitable treatment. Every year, terabytes of medical image data’s square measure
is used through progressive imaging modalities like Positron Emission
Tomography (PET), Magnetic Resonance Imaging (MRI), Computed
Tomography (CT), and lots of additional new methodology of medical imaging.
Improvements in technology have given the chance for radiology systems to use
intricate compression algorithms to scale back the file size of every image in an
endeavor to raise the knowledge volume shaped by new or additional intricate
modalities. In General, various compression strategies like Discrete Cosine
Transform (DCT), Discrete Wavelet Transform (DWT), Fractal Compression,
Set Partitioning In Hierarchical Trees (SPIHT), Neural Network Back
Propagation (NNBP) and Radial Basis Function Neural Network (RBFNN) are
applied to medical images. Currently, evolving hybrid schemes for effective
image compression has gained immense admiration among the researchers. The
hybrid technique affords well-organized and precise coding of the medical
images. An efficient compression technique like Hybrid DWT with BPNN and
Hybrid Fractal with NNRBF for compression of the medical data is able to
resolve the complications with storage and transmission. The latest compression
schemes bring better compression rates if the loss of quality is affordable.
Medicine cannot afford insufficiency in diagnostically significant Region of
Interest (ROI). An approach that brings high compression rate with good quality
in the ROI is required. A hybrid coding scheme seems to be the only solution to
this twofold problem. The objective of this thesis is to compare a few basic
compression techniques with Hybrid DWT with BPNN and Hybrid Fractal with
NNRBF for compression. There are different parameters for analyzing the image
compression methods, which include Compression Ratio (CR), Peak Signal to
Noise Ratio (PSNR), Bits per pixel (Bpp) and Mean Square Error (MSE). The
IV
quality of any compressed image can be assessed using this set of parameters.
The result clearly shows that hybrid image compression using Hybrid Fractal
with NNRBF provides better CR and PSNR. The mentioned analyses are carried
out in MATLAB simulations.
V
ACKNOWLEDGEMENT
Though only my name appears on the cover of this thesis, a great
many people have contributed to its production. I owe my gratitude to all those
people who have made this thesis possible and because of whom my research
experience has been one that I will cherish forever.
My deep felt gratitude goes to our respected and honorable Chairman
(Late) “Kalvivallal” Thiru. T. Kalasalingam, B.Com., for providing the
technical environment to complete my project successfully.
I am highly indebted to express my token of thanks to our beloved
Chancellor “Ilayavallal” Dr. K. Sridharan, M.Com., M.B.A., M.Phil., P.hD.,
for allowing me to do the project work.
I thank our Director Dr.S.Shasi Anand, Ph.D., for being the beacon
light in guiding and infusing the strength and enthusiasm to work over
successful.
I express my sincere thanks to our Vice-Chancellor Dr.S.Saravana
Sankar, Ph.D., for his valuable suggestions and continuous encouragement in
the completion of the project work.
I am extremely grateful to my Supervisor Dr.M.Pallikonda
Rajasekaran, Ph.D., Professor, Department of Electronics and Communication
Engineering for his many thoughtful comments and valuable suggestions.
Last but not the least I thank all, my parents, my wife, my son,
teaching staff of our University, non-teaching staff, R&D Department and my
friends for their moral support.
PERUMAL B
VI
TABLE OF CONTENTS
CHAPTER
NO.
TITLE PAGE
NO.
ABSTRACT
LIST OF TABLES
LIST OF FIGURES
LIST OF SYMBOLS AND ABBREVIATIONS
III
XI
XIII
XVI
1. INTRODUCTION
1.1 INTRODUCTION 1
1.2 DATA COMPRESSION 1
1.2.1 Image Compression 2
1.2.2 Compression Techniques 2
1.2.2.1 Lossless Compression 2
1.2.2.2 Lossy Compression 5
1.3 IMAGING TECHNIQUES 7
1.3.1 Computer Tomography (CT) 7
1.3.2 Magnetic Resonance Imaging (MRI) 7
1.3.3 Positron Emission Tomography (PET) 9
1.4 IMAGE COMPRESSION PERFORMANCE
METRICS
10
1.4.1 Image quality 11
1.4.1.1 Distortion 11
1.4.1.2 Fidelity or Quality 12
1.4.1.3 Compression Ratio (CR) 12
1.4.1.4 Bits per pixel (Bpp) 12
1.4.1.5 Speed of Compression 13
1.5 THE COMPRESSION SYSTEM 13
1.6 OVERVIEW OF METHODOLOGY 16
VII
1.7 RESEARCH MOTIVATION 16
1.8 PROBLEM STATEMENT 18
1.9 OBJECTIVES OF THE RESEARCH 18
1.10 CONTRIBUTIONS OF THE THESIS 19
1.11 ORGANIZATION OF THE THESIS 19
2. LITERATURE SURVEY
2.1 INTRODUCTION 22
2.1.1 Description of image compression block
diagram
23
2.1.2 Image format 24
2.2 LITERATURE REVIEW 24
2.3 CONCLUSION 40
3. METHODOLOGIES
3.1 DISCRETE COSINE TRANSFORM (DCT) 41
3.1.1 Image Compression in DCT 42
3.1.2 DCT Encoding 43
3.1.3 Compression Steps in DCT 44
3.1.4 Quantization Steps 44
3.1.5 Entropy Encoding 45
3.2 DISCRETE WAVELET TRANSFORM (DWT) 46
3.2.1 Advantages of DWT 48
3.2.2 Wavelets used in Image Compression 48
3.2.3 Aspects of Wavelets 49
3.3 FRACTAL ALGORITHM 50
3.3.1 Presentation about Fractal Algorithm 51
3.3.2 Features of Fractal Algorithm 52
3.3.3 Fractal Image Compression 52
3.4 SET PARTITIONING IN HIERARCHICAL
TREES (SPIHT)
53
VIII
3.4.1 Haar Wavelet 53
3.4.2 Formation of Cells 54
3.4.3 Zero Tree Encoding 55
3.4.4 SPIHT Algorithm 55
3.5 INTRODUCTION TO NEURAL NETWORKS 57
3.5.1 Back Propagation Neural Networks
(BPNN)
58
3.5.2 Image Compression using Back
Propagation
60
3.5.3 Use of Image Compression in Back
Propagation
61
3.5.4 Neural Network Radial Basis Function
(NNRBF)
63
3.5.4.1 Radial basis function operation 63
3.5.4.2 Output nodes 66
3.5.4.3 Training of RBF neural
networks
66
4. COMPRESSION TECHNIQUES FOR MEDICAL
IMAGES USING FRACTAL, SPIHT AND DCT
ALGORITHMS
4.1 INTRODUCTION 67
4.2 FRACTAL, SPIHT AND DCT METHODS 68
4.2.1 Fractal 68
4.2.2 Set Partitioning in Hierarchical Trees 69
4.2.3 Discrete Cosine Transform 70
4.3 IMAGE QUALITY PARAMETER
EVALUATION
71
4.3.1 Performance Parameter 72
4.4 RESULTS AND COMPARISON 73
IX
4.5 CONCLUSION 83
5. EFFICIENT IMAGE COMPRESSION
TECHNIQUES FOR COMPRESSING
MULTIMODAL MEDICAL IMAGES USING
NEURAL NETWORK RADIAL BASIS
FUNCTION APPROACH
5.1 INTRODUCTION 84
5.2 ALGORITHMS FOR IMAGE COMPRESSION 85
5.2.1 Fractal Algorithm 85
5.2.2 Neural Network Back Propagation 86
5.2.3 Neural Network Radial Basis Function
for Image Compression
87
5.3 PERFORMANCE PARAMETERS 88
5.4 RESULTS AND COMPARISON 88
5.5 CONCLUSION 98
6. A HYBRID DISCRETE WAVELET
TRANSFORM WITH NEURAL NETWORK
BACK PROPAGATION APPROACH FOR
EFFICIENT MEDICAL IMAGE COMPRESSION
6.1 INTRODUCTION 99
6.2 ALGORITHMS USED 99
6.2.1 Back Propagation Neural Networks
Algorithm
99
6.2.2 Discrete Wavelet Transform 101
6.3 PERFORMANCE PARAMETERS 102
6.4 RESULTS AND DISCUSSION 102
6.5 CONCLUSION 113
7. A HYBRID APPROACH USING FRACTAL AND
NEURAL NETWORK RADIAL BASIS
X
FUNCTION FOR EFFICIENT COMPRESSION
OF MULTI MODAL MEDICAL IMAGES
7.1 INTRODUCTION 114
7.2 METHODOLOGIES 115
7.2.1 Fractal Algorithm 115
7.2.2 Neural Network Radial Basis for Image
Compression
115
7.3 IMPLEMENTATION OF HYBRID
TECHNIQUES
117
7.3.1 Hybrid image compression 117
7.4 IMAGE QUALITY PARAMETER
EVALUATION
117
7.5 SIMULATION RESULTS AND ANALYSIS 118
7.6 CONCLUSION 128
8. CONCLUSION AND FUTURE WORK
8.1 CONCLUSION 130
8.2 FUTURE WORK 135
REFERENCES 136
LIST OF PUBLICATIONS 146
CURRICULUM VITAE 149
XI
LIST OF TABLES
TABLE
NO.
TITLE PAGE
NO.
1.1 Run Length Encoding 3
4.1
Performance comparison of 24 medical images
which are obtained by using DCT, SPIHT and Fractal
algorithm
74
5.1
Performance comparison of 24 medical images
which are obtained by using Fractal, NNRBF and
NNBP algorithm
89
6.1
Performance comparison of 24 medical images
which are obtained by using DWT, BPNN and hybrid
DWT-BP algorithm
104
7.1
Performance comparison of 24 medical images
which are obtained by using NNRBF, Fractal and
Hybrid FNNRBF.
119
8.1 Compression ratio of 24 medical images which are
obtained by using DCT, DWT, Fractal, NNBP,
NNRBF, Hybrid Fractal and NNRBF and Hybrid
DWT-NNBP algorithm
131
8.2 Peak Signal to Noise Ratio of 24 medical images
which are obtained by using DCT, DWT, Fractal,
NNBP, NNRBF, Hybrid Fractal and NNRBF and
Hybrid DWT-NNBP algorithm
132
8.3 Memory of 24 medical images which are obtained by
using DCT, DWT, Fractal, NNBP, NNRBF, Hybrid
Fractal and NNRBF and Hybrid DWT-NNBP
algorithm
133
XII
8.4 Execution time of 24 medical images which are
obtained by using DCT, DWT, Fractal, NNBP,
NNRBF, Hybrid Fractal and NNRBF and Hybrid
DWT-NNBP algorithm
134
XIII
LIST OF FIGURES FIGURE
NO.
TITLE PAGE
NO.
1.1 Schematic view of an MRI scanner 8
1.2 General block diagram of compression technique 13
1.3 General block diagram of de-compression
technique
13
1.4 The compression process on forward transform 16
2.1 Block diagram of Image Compression 23
3.1 Block diagram of DCT 42
3.2 Conversion of special domain to frequency
domain
43
3.3 Block diagram of DWT 47
3.4 Three Level Decomposition Wavelet Filter 49
3.5 2-D Discrete Wavelet Transform in image
compression
50
3.6 A photo copy machine that makes three reduced
copies of the input image
53
3.7 Formation of cells of parent-offspring conditions 55
3.8 Spatial orientation tree in SPIHT 56
3.9 Block diagram of Neural Network 58
3.10 General Structure of BPNN 59
3.11 General structure of NNRBF Algorithm 65
4.1 Flow diagram of Fractal coding 68
4.2 Basic block diagram of SPIHT method 69
4.3 Formation of cells of SPIHT 69
4.4 Two-dimensional DCT of 8-by-8 blocks in the
image
71
4.5 Compression Ratio expressed in percentage 75
XIV
4.6 Shows the PSNR for three different algorithms
DCT, SPIHT and Fractal
75
4.7 Memory expressed in kilo byte 76
4.8 Execution Time expressed in Seconds 76
4.9 Results obtained for various medical images (a).
Input Image (b). DCT (c). SPIHT and (d). Fractal
algorithms
77
5.1 General structure of Neural Network Back
Propagation Algorithm
87
5.2 General structure of Radial Basis Function Neural
Network
88
5.3 Compression Ratio expressed in percentage 90
5.4 Peak Signal to Noise Ratio expressed in decibels 90
5.5 Memory expressed in kilo byte 91
5.6 Execution Time expressed in Seconds 91
5.7 Results obtained for various medical images (a).
Input Images (b). Fractal, (c). Neural Network
Back Propagation (NNBP) and (d). Radial Basis
Function Neural Network algorithms
92
6.1 General Structure of BPNN 100
6.2 Block diagram of Hybrid DWT- BP algorithm 101
6.3 Comparison chart of proposed work and existing
method
102
6.4 Comparison of Compression Ratio for different
Input images
105
6.5 Comparison of PSNR Values for different Input
Image
105
6.6 Memory expressed in kilo byte 106
6.7 Execution Time expressed in Seconds 106
XV
6.8 Results obtained for various medical images (a).
Input Images (b). DCT, (c). SPIHT and (d).
Fractal algorithms
107
7.1 General structure of NNRBF 116
7.2 Hybrid image compression using FNNRBF
method
117
7.3 Compression Ratio expressed in percentage 120
7.4 PSNR expressed in decibel 120
7.5 Memory expressed in kilo byte 121
7.6 Execution Time expressed in Seconds 121
7.7 Results obtained for various medical images (a).
Input Images (b). Fractal, (c). NNRBF and (d).
Hybrid Fractal & NNRBF algorithms
122
XVI
LIST OF SYMBOLS AND ABBREVIATIONS
Symbols
HRL - Run - Length Entropy
H0, H1 - Entropies of the high contrast
L0, L1 - Normal values of high contrast run length
NxN - Dimensions of the images
I(i,j) - Original image
K(i,j) - Approximated version
MAXI - Maximum possible pixel value of the image
F(u,v) - DCT coefficient in row k1 and column k2 of the DCT
matrix
f(x,y) - intensity of the pixel in row i and column j
3D - Three Dimensional
W - Weights between the hidden layer and the output layer
N - Columns of input image
M - Rows of input image
A - Matrix representing the 2D image pixels
T - Consequence of the complete transformation
f (x,y) - Original image
g (x,y) - Reconstructed image
M, N - Rows and columns of input image
N - Quantity of neurons in the hidden layer
Ci - Inside vector for neuron
i, and ai - Weight of neuron
XVII
Abbreviations
RLE - Run-length Encoding
LZW - Lempel-Ziv–Welch
DFT - Discrete Fourier Transform
DCT - Discrete Cosine Transform
IFS - Iterated Function System
CT - Computed Tomography
MRI - Magnetic Resonance Imaging
PET - Positron Emission Tomography
RF - Radio Frequency
MAP - Maximum a Posteriori
FBP - Filtered Back Projection
ANN-BPN - Artificial Neural Network with Back Propagation Network
SVM - Support Vector Machine
ANN-RBF - Artificial Neural Network with Radial Basis Function
MSE - Mean Square Error
PSNR - Peak Signal to Noise Ratio
CR - Compression Ratio
Bpp - Bits per pixel
SPIHT - Set Partitioning in Hierarchical Trees
NNBP - Neural Network Back Propagation
NNRBF - Neural Network Radial Basic Function
RBFNN - Radial Basis Function Neural Network
DWT-BP - Discrete Wavelet Transform-Back Propagation
NN - Neural Network
FNNRBF - Fractal based Neural Network Radial Basis Function
ISO - International Standard Organization
EZW - Embedded Zero Tree Wavelet
XVIII
BPNN - Back Propagation Neural Networks
RBF - Radial Basis Function
MLP - Multi-Layer Perception
PIFS - Partition Iterated Function System
NNTOOL - Neural Network Tool
ANN - Artificial Neural Network
FIF - Fractal Image Format
1
CHAPTER 1
INTRODUCTION
1.1 INTRODUCTION
Signal processing is a type of Image processing, where the input and
output signals are images. Images can be thought as 2-D signals via a matrix
representation. Earlier, image processing was mostly carried out using analog
devices. Now a day, images are processed in digital domain.
Digital image processing overcomes issues for example the
inflexibility of system to change noise, distortion during processing, and
difficulty of implementation. Image processing is a method that enhances
original images got from camera and sensors in day -to-day life.
A different method has been implemented for image processing
during the past few decades. Image processing systems are becoming popular
because of the availability of powerful personal computers and devices of large
memory, availability of graphics software etc. A wide range of application of
image processing includes the following: i) Remote Sensing ii) Medical Imaging
iii) Forensic Studies iv) Textiles v) Material Science vi) Military etc.
1.2 DATA COMPRESSION
Data compression is a process of compressing the number of data bits
for presenting the data which is in difficult form of sequence, so that storing or
transmitting the data is done in a proficient manner. The data could be an image
or video or an audio, and in the present context. Image compression is a method
of data compression, which encodes unique image with less bits. The goal is to
decrease the size of the storage. While retrieving the original image from the
compressed image, the decompressed image should be similar to the original
image.
2
1.2.1 Image Compression
The image has become the most important information carrier in
people’s life or the biggest media containing information. As the purpose of
storing and transmitting images continues to increase, the field of image
compression also continues to develop. An image contains a large amount of
data, mostly with redundant information that occupies massive storage space and
minimizes transmission bandwidth. An image consists of pixels, which are
highly correlated to one another within a close proximity. The correlated pixels
lead to redundant data.
Two types of data redundancy are observed as follows
• Spatial Redundancy: The intensities of neighboring pixels are
correlated. So, the intensity information of an image contains
unnecessarily repeated (i.e. redundant) data within one frame.
• Spectral Redundancy: Different frequencies of an image contain
redundant data because of the relationship between various color
planes.
1.2.2 Compression Techniques
This compression technique is classified into two techniques namely,
Lossless and Lossy compression algorithms which are explained below.
1.2.2.1 Lossless Compression
The Lossless compression methods mean receiving the data without
loss. The initial data might be retrieved exactly from the data compressed. It is
used in various fields that can't ensure any variation between the original data.
Lossless compressed image has a larger size compared with lossy one. In a
power constrained applications like wireless communication, lossless
compression is not preferred as it consumes more energy and more time for
3
image transfer. In the following sections, lossless compression techniques are
discussed.
a) Run length encoding
b) Huffman encoding
c) LZW coding
d) Area coding
a) Run Length Encoding
It is highly basic compressed strategy utilized for continues data. It is
useful in case of redundant information. The representation of run length code
for gray scale images in the sequence is Vi and Ri, where Vi is referred as pixel
and Ri is referred as quantity of successive pixels with the force Vi as appeared
in Table 1.1. It varies the range of 11 pixels coded utilizes five bytes providing a
compression ratio of 11:5 which represent by one byte.
86 86 86 86 86 91 91 91 91 75 75 {86,5} {91,4} {75,2}
Table 1.1 Run Length Encoding
The images which repeat the intensities along their row and column
can be frequently compressed by representing the runs of indistinguishable
intensities where run - length sets, every run-length set indicates to begin another
force. The quantity of back to back pixels has that intensity. This strategy is
utilized for data compression as a part of bitmap image file format. The RLE
(Run-length Encoding) is especially successful when compacting with the binary
images, subsequently to two conceivable intensities with high contrast.
Moreover, a variable - length code can be connected with the run lengths
themselves. The approximate run-length entropy is
1010
LLHHH RL +
+= (1.1)
4
Where (H0, H1) are the entropies of the high contrast, (L0, L1) are the normal
values of high contrast run length.
b) Huffman Encoding
This technique is utilized for coding symbols based on their
measurable event frequencies probabilities. In this method, the pixels in the
image are considered as symbols. The symbols which occur more often are
assigned as lesser number of bits, while the symbols that happen less frequently
are allotted as relatively more number of bits. It is a prefix code. Nearly every
image coding norms uses lossy practices in the initial stages of compression and
uses Huffman coding as the last step.
c) LZW Coding
LZW (Lempel-Ziv–Welch) is a word reference based coding. It may
be static or dynamic. Static dictionary coding: the word reference is fixed during
the encoding and decoding processes. Dynamic word reference coding: the word
reference is updated on fly. It is extensively used in computer industry and is
implemented as compress command on UNIX.
d) Area Coding
Area coding is an improved type of run length coding, mirroring the
2-Dimensional character of images. This is a critical development over the
alternate lossless techniques. For coding an image, it doesn't make a heavy
impact to interpret it as a successive stream, as it is actually an array of
sequences working up a two dimensional object. The calculation of area coding
is to find the rectangular districts with the same attributes. These areas are coded
in an elucidating form as a component with two focuses and a specific structure.
This sort of coding can be very powerful, yet it bears the issue of a nonlinear
technique, which is hard to execute in equipment. Accordingly, the execution as
far as the compression time is not a constraint.
5
1.2.2.2 Lossy Compression
Lossy compression includes some loss of data. The information that
has been packed utilizing lossy systems for the most part can't be recouped or
reproduced precisely. It causes in superior compression ratios to the detriment of
mutilation in recreation. The benefit of lossy over lossless is high compression
ratio, less process time and low energy requirements in case of power
constrained applications. In the following sections the lossy compression
techniques are explained.
a) Transformation coding
b) Vector quantization
c) Fractal coding
d) Block Truncation Coding
e) Sub band coding
a) Transformation Coding
In this coding scheme, DFT (Discrete Fourier Transform) and DCT
(Discrete Cosine Transform) are utilized to change the pixels in the original
image into recurrence space coefficients (it is called as transform coefficients).
These set of coefficients have a few alluring properties. One is the energy
compaction property that results in the greater part of the energy of the first
information being packed in just a couple of the noteworthy change coefficients.
This indicates the essential of accomplishing the compression. Just those couple
of critical coefficients is chosen and the remaining is disposed. The chosen
coefficients are considered for further quantization and entropy encoding. DCT
coding has been the most well-known way to deal with transform coding.
b) Vector Quantization
The fundamental thought in this system is to build up a dictionary of
fixed size vectors, called code vectors. A vector is typically a piece of pixel
6
qualities. A given image is then partitioned into non-overlapping blocks
(vectors) called image vectors. At that point, every vector is resolved and its file
in the word reference is utilized as the encoding of the original image vector.
c) Fractal Coding
The crucial thought here is to decompose the image into portions by
utilizing standard image handling systems, for example, shading partition, edge
discovery, and range and surface analysis. The library really contains codes
called Iterated Function System (IFS) codes, which are the conservative
arrangements of numbers. Utilizing an orderly method, an arrangement of codes
for a given image is resolved, such that when the IFS codes are connected in an
appropriate manner of image, squares yield a picture that is in nearby estimate of
the first. This idea is very viable for packing images that have great normality
and self-similitude.
d) Block truncation coding
In this method, the image is separated into non -overlapping blocks of
pixels. For every square, limit and remaking qualities are resolved. The limit is
normally the mean of the pixel values in the block. Then a bitmap of the piece is
inferred by replacing all pixels whose qualities are more prominent than or break
even with (not exactly) at the edge of a 1 (0). Then for every segment (gathering
of 1s and 0s) in the bitmap, the remaking quality is resolved. This is the normal
estimations of the comparing pixels in the original block.
e) Sub band coding
In this method, the image is examined to deliver the parts containing
frequencies in very much characterized groups, called sub groups. Hence,
quantization and coding are applied to each of the groups. The advantage of this
plan is that the quantization and coding reasonable for every sub band can be
planned independently. Compression strategies can be applied specifically to the
7
images or to the changed image data (changed space). The transform coding
methods are appropriate for image compression. Here, the image is decayed or
changed into segments that are then coded by individual attributes. The change
ought to have high-vitality compaction property, in order to accomplish high
compression proportions. Cases: Discrete Cosine Transform (DCT), Wavelet
Transform, Multi wavelet Transform etc.
1.3 IMAGING TECHNIQUES
Medical image compression techniques form the basis for common
imaging modalities such as CT (Computed Tomography), MRI (Magnetic
Resonance Imaging) and PET (Positron Emission Tomography), and they are
useful in the fields of medicine, biology, earth science, archaeology, materials
science and nondestructive testing. On the other hand, anatomical and
morphological imaging techniques like X-ray, CT and MRI, which are widely
used in clinical offer high anatomical resolution, but are not capable of imaging
metabolic activity.
1.3.1 Computer Tomography (CT)
A Computer Tomography (CT) utilizes a computer System that takes
data from several x-ray images of structures inside a patient’s body and converts
them into images on a monitor. Tomography is the procedure of generating a 2D
image slice or section through a 3D image. A CT Scanner uses digital geometric
processing to create a 3D image of the inside of an object. CT Scanner radiate a
series of narrow beams through the patient body as it moves through an arch, not
like x-ray machine which sends just one radiation beam. The final picture is
more information’s than an X-Ray Image.
1.3.2 Magnetic Resonance Imaging (MRI)
The basic idea of MRI is to study the response of the magnetized
tissue to Radio Frequency (RF) signals and deduce the underlying properties of
8
the tissue. An MRI system consists mainly of three hardware components. The
main magnet produces a high magnetic field Magnet coil (B0) which is used to
magnetize the tissue. The higher the magnetic field is the higher the SNR, which
can be potentially achieved with the scanner. It is essential to be homogeneous
over the imaging volume in order to avoid the distortions in the acquisition;
additional shim coils are used to guarantee the homogeneity even after the
introduction of the patient in the bore. Other than some open MRI scanners using
permanent magnets, the most clinical scanners use a cylindrical superconducting
magnet consisting of a solenoid of wire (typically niobium-titanium), which
operates within liquid helium at 4 Kelvin, in order to have superconducting
properties and not offer resistance to the current. Therefore, the magnetic field
always stays on even when the scanner is not being operated. Modern clinical
MRI scanners have a main magnet producing the field strength of typically 1.5
or 3 Tesla, although preclinical and research scanners can use 7 Tesla. Earth's
magnetic field strength is about 0.00005 Tesla.
Figure 1.1 Schematic view of an MRI scanner
In Figure 1.1 the main magnet produces a strong homogeneous
magnetic field (a) the gradient coils are responsible for the spatial localization of
the signal (b) RF transmission coil (c) the signal response from the excited spins
within the patient being measured by local surface coils placed on the imaging
volume.
9
1.3.3 Positron Emission Tomography (PET)
The Positron Emission Tomography (PET) is an indicative imaging
instrument that gives images of radioactive substances infused into the perished
to delineate characteristic capacities. The radioactive center transmits a positron
which demolishes the electron to produce two 511 keV photons meandering in
pretty much inverse headings to be distinguished unexpectedly by two
indicators. Numerous photons are intrigued or sprinkled, dropping the quantity
of identified emanation occasions. PET images can be utilized specifically or
after dynamic demonstrating to pull out quantitative estimations of a favored
physiological, biochemical or pharmacological element. Since, such depictions
are typically loud, it is imperative to see how clamor influences the subsequent
quantitative standards. A pre-essential for this kind is that the properties of
clamor: that are known in variety (size) and quality (relationship).
Investigational PET information is obtained in Two Dimensional (2D) and Three
Dimensional (3D) ownership mode and reproduced by scientific Filtered Back
Projection (FBP) and factual Maximum a Posteriori (MAP) approach, with
delicate figuring systems like Artificial Neural Network with Back Propagation
Network (ANN-BPN), Support Vector Machine (SVM) and Artificial Neural
Network with Radial Basis Function (ANN-RBF).
One of the major dissimilarities between PET scans and other imaging
tests like CT scan or MRI is the PET scan reveals the cellular level metabolic
changes occurring in an organ or tissue. This is significant and exclusive because
disease processes very often begin with handy vicissitudes at the cellular level.
A PET scan can very often detect these initial changes. Whereas, a CT or MRI
detects vicissitudes a little later as the disease starts to cause vicissitudes in the
structure of organs or tissues.
Compared to CT images, PET images show a lower anatomical
resolution, which can affect the correct localization of lesions and demarcation
10
of their borders. Moreover, the contrast in CT imaging is based on the density of
the tissue, while contrast in PET imaging results from metabolical activity in the
tissue. However, such combination requires a decent alignment of the PET and
CT image data. In image processing, such alignment is called registration. Image
data is organized in a discrete coordinate system containing numerous elements.
In two-dimensional (2D) images these elements are called pixels. When acquired
with standalone scanners, the coordinate systems of PET and CT image data will
most certainly differ from each other. The reason for the difference can be
manifold but the highest influence derives from the change of patient position
between the scans. It is practically almost impossible for a patient to maintain
the exact place when transferred from one scanner to the other. To cope up with
this requirement, the combination of PET and CT scanners into a single device
has been realized in the early 2000s. In contrast to the PET and CT scanners,
such combined scanners represent an effective approach for the acquisition of
accurately registered images and are able to contribute to the reduction of overall
scan time up to 40%. The reason for this significant improvement is based on the
fact that the scans are acquired subsequently, thus minimizing the possibility of
movement. Moreover, the patient is moved on an automatic bed from one
scanner to the other describing a linear. One-dimensional translation can be
registered very fast and accurate by software algorithms.
1.4 IMAGE COMPRESSION PERFORMANCE METRICS
The performance of a compression technique can be assessed in a
number of ways the amount of compression, the comparative difficulty of the
technique, memory constraint for implementation, time required for the
compression on a machine, and the distortion rate in the reconstructed image.
The following are the performance metrics to evaluate the compression
techniques.
11
1. Image Quality 2. Compression ratio 3. Speed of compression
• Computational complexity • Memory resources.
4. Power consumption
1.4.1 Image quality
There is a need for specifying methods to judge image quality after the
reconstruction process and to measure the amount of distortion due to
compression process, as minimal image distortion means better quality. There
are two types of image quality measures, subjective quality measurement and
objective quality measurements. Subjective quality measurement is established
by asking human observers to judge and report the image or video quality
according to their experience; and these measures would be relative or absolute.
Absolute measures classify image quality not regarding to any other image but
according to some criteria of television allocations study organization. On the
other hand, relative measures compare image against another and choose the best
one. The quantitative measurements are discussed in the following.
1.4.1.1 Distortion
The variation between the original and reconstructed image is called
as distortion. It is denoted using Mean Square Error (MSE).
MSE= ( ) ( )[ ]21-n
0i
1n
0jji,K ji,I
nn1
∑=
∑−
=−
× (1.2)
12
n×n (n-rows, n-columns) represents noise free original image I (input image) and
noisy approximation K (output image)
1.4.1.2 Fidelity or Quality
It defines the similarity between the original and reconstructed
images. It can be measured using Peak Signal to Noise Ratio (PSNR) in dB.
PSNR =
MSE
2IMAX
log 10 (1.3)
Here MAXI represents the maximum possible pixel value of the image. Pixels
are represented as eight bits per samples. Logically, a greater value of PSNR is
good because it indicates that the ratio of Signal to Noise is greater. Here, the
'signal' is the genuine image and the 'noise' is the error due to reconstruction.
1.4.1.3 Compression Ratio (CR)
It is the ratio of the number of bits required to represent the image
prior to the compression and to the number of bits required to represent the
image after compression.
Compression Ratio (CR) =
size file Compressed
size file edUncompress (1.4)
Where, CR can be used to judge how the compression efficiency is. The lower
CR means better compression.
1.4.1.4 Bits per pixel (Bpp)
It is the average number of bits required to represent a single sample.
It is represented in terms of Bits per pixel (Bpp).
13
Bits per pixel (Bpp) =
⇒
N x N
bytes ofNumber x 8pixels ofNumber bits ofNumber (1.5)
Where N-stands for rows, N- stands for columns.
1.4.1.5 Speed of Compression
Compression speed depends on the compression technique that has
been used, as well as, the nature of platform that hosts the compression process.
Compression speed is influenced by computational complexity and size of
memory. Lossy compression is a complex process that increases system
complexity, storage space and needs more computational element clock.
1.5 THE COMPRESSION SYSTEM
The compression system model consists of two parts
Compression
De-compression
Figure 1.2 General block diagram of Compression Technique
Figure 1.3 General block diagram of De-compression Technique
The compressor shown in Figure 1.2 consists of a preprocessing stage
that performs data diminution (reduction) and mapping. The encoding stage
performs quantization and coding, whereas, the de-compression consists of a
decoding stage that performs decoding and inverse mapping followed by a post-
14
processing stage, as shown in Figure 1.3. In compression, prior to encoding
process, preprocessing is accomplished to prepare the image for the encoding
process and consists of many operations that are application specific. Post-
processing can be accomplished to remove some of the potentially unwanted
artifacts brought about by the compression process, after the completion of
compressed file has been decoded.
The compression process can be divided into following stages:
• Image Data reduction: Image data can be reduced by gray level and
spatial quantization, and can undergo any desired image improvement
(for example, noise removal) process.
• Mapping: Involves mapping the original image data into one more
mathematical space, wherever it is easier to compress the data.
• Quantization: Involves taking potentially continuous data from the
mapping stage and putting it in discrete form.
• Coding: Involves mapping the quantized data (discrete) onto a code in
an optimal manner.
The mapping procedure is significant because the image data are
highly linked. If the value of one pixel is known, it is likely that the adjacent
pixel value is identical. On finding a mapping equation that de-correlates the
data, such type of data redundancy can be detached.
• Differential coding: Method of reducing the data redundancy is done
by finding the difference between the adjacent pixels and encoding
those values.
• Principal components transform: This provides a theoretically
optimal decorrelation.
15
As the spectral domain can also be used for image compression, the
first stage may include mapping into the frequency or sequence domain, where
the energy in the image is compressed mainly into lower frequency components.
• Quantization may be essential to convert the data into digital form (bit
data type), depending on the mapping equation used. This is because
many of these mapping methods will result in floating point data,
which require multiple bytes for representation and it is not very
efficient as far as the goal is to reduce the data.
Decompression process can be divided into the following stages:
• Decoding: Takes the compressed file and drives back the original
coding by mapping the codes to the original quantized values.
• Inverse mapping: Involves reversing the original mapping process.
• Post-processing: Involves enhancing the structure of the final image.
De-compression might be done to drive back any preprocessing, for
example, enlarging an image that was shrunk in the data reduction
process. In other cases, the post-processing possibly will be used
simply to enrich the image and to improve any excavation from the
compression process itself. The development of a compression
algorithm is highly application specific. The preprocessing stage of
compression consists of processes such as enhancement, noise
removal or quantization. The goal of preprocessing is to prepare the
image for the encoding process by rejecting any inappropriate
information. For example, many images that are used only for the
viewing purposes can be preprocessed by eliminating the lower bit
planes, without losing any useful information.
16
1.6 OVERVIEW OF METHODOLOGY
The methodology for the compression process which takes the image
of (NxN) size as input is shown below.
Figure 1.4 The compression process on forward transform
The Figure 1.4 represents a compression process flow for an input
image. The compression process pre-analyzes rows and columns and performs
encoding techniques like magnitude set and bit plane coding followed by run
length encoding. The sign data of the coefficients are coded as bit plane with
zero thresholds. This bit plane may be used as it is coded to scale back the Bits
per pixel (Bpp). The coefficients are coded by means of run length coding and
magnitude set coding techniques, which in turn result in low bits.
1.7 RESEARCH MOTIVATION
Enormous quantities of data are involved in the process of storage
and/or transmission of images, videos, sound and text in several applications.
17
The application areas are medical imaging, Tele radiology, Satellite/Space
imaging, Multimedia digital video (entertainment, home use) and digital
photography. However, compression becomes very essential in medical
imaging.
MRI is a noninvasive method for producing 3D tomography images of
the human body. It is frequently used for the detection of tumors, grazes and
other irregularities in soft tissues, such as the brain. Clinically, radiologists
qualitatively analyze the brain surface produced by MRI scanners.
Recently, computer-aided techniques for analyzing and visualizing
MR images have been examined. Many researchers have focused on detecting
and quantifying irregularities in the brain. Automatically recognizing the
pathologies in MR images of the head is a vital step in this process. One more
important step in computer-aided analysis is the data quality assurance. MR
images comprise unwanted intensity variations due to imperfections in MRI
scanners. Reducing these dissimilarities can improve the accuracy of automated
analysis.
A single clinical MRI scan occupies numerous megabytes of disk
space. Effective image compression systems are significant for storing
multitudes of scans. By means of tele-radiology, wherever MRI scans are
transmitted by wire to remote sites for assessment by specialists, MR image
compression plays a massive role in rising transmission speeds.
In the MRI scans of the head, doctors are typically more interested in
the brain as opposed to the region outside the brain. Aimed at this reason,
Anderson has developed a lossy MRI compression scheme that selectively
compresses the region outside the brain at a higher compression ratio than the
brain. Thus, high compression ratio has been achieved while upholding image
quality of the brain area. Obviously, automatic intracranial boundary detection is
a prerequisite for such a scheme.
18
1.8 PROBLEM STATEMENT
1. Medical image data like CT, MRI and PET consume maximum
storage and use maximum bandwidth for transmission that frequently
results in degradation of image quality.
2. Medical image compression considering lossy and lossless types, the
medical images to be compressed efficiently with optimal
compression ratio.
3. Image compression is the foremost component of communication and
storage systems where the uncompressed images need considerable
compression technique, which should be competent of reducing the
crippling disadvantages of data transmission and image storage.
4. The compression is being performed on the whole image without
considering the region of diagnostic importance.
5. Existing compression technique doesn’t guarantee the substantial
noticeable quality of an image with optimum bit rate.
6. Supportability of existing compression techniques on telemedicine
remains undiscovered.
1.9 OBJECTIVES OF THE RESEARCH
The main objectives of this research work are
• To improve the compression ratio for medical image's using DCT,
SPIHT and Fractal algorithm and to analyze their performance.
19
• To present an improved image compression algorithm for medical
images using DWT, BPNN and NNRBF.
• To analyze different modality images using Hybrid technique namely
DWT-BPNN.
• To develop Hybrid approach using Fractal and Neural Network Radial
Basis Function for efficient compression of multi modal medical
images.
1.10 CONTRIBUTIONS OF THE THESIS
In this thesis, medical image compression has been carried out using
various methods. The compression algorithms such as DCT, DWT, SPIHT,
Fractal, LZW, Neural NNBP, NNRBF, Hybrid DWT-NNBP and Hybrid
approach for FNNRBF have been applied to multi modal medical images (PET,
MRI and CT). Quality parameters such as CR, PSNR, Execution time and
Memory usage are considered for performance analysis. It is observed that
FNNRBF method has low CR and high PSNR values. Hybrid Fractal and
NNRBF is found to be more efficient than Fractal and FNNRBF methods.
1.11 ORGANIZATION OF THE THESIS
Chapter 1 : This chapter discusses the basics of PET images and image
compression techniques. The problem statement, objectives, contribution and the
scope of research are presented in this chapter. The organization of thesis is also
presented in this chapter.
Chapter 2: This chapter describes literature review of the existing techniques
such as Huffman, LZW, DCT, DWT, SPIHT and Fractal based compression
techniques, NNBP and NNRBF algorithms for PET, MRI and CT image
compression.
20
Chapter 3: In this chapter, a brief discussion on the methods such as Huffman,
LZW, DCT, DWT, SPIHT, Fractal, NNBP and NNRBF based compression
techniques algorithms are described.
Chapter 4: This chapter discusses the various compression techniques like
DCT, Fractal Compression and SPIHT applied to numerous medical images.
Experimental results show that the outlined DCT approach achieves the better
CR, Bpp and PSNR with less MSE on comparison with SPIHT and Fractal
methodology.
Chapter 5: This chapter describes the different compression methods such as
Fractal, NNBP and NNRBF applied to various medical images such as MR, CT
and PET. Experimental results show that the NNRBF technique achieves a low
CR and higher PSNR, with less MSE on MR, CT and PET images, when
compared to Fractal and NNBP techniques.
Chapter 6: This chapter compares a few promising compression techniques
such as DWT algorithm, NNBP and new hybrid techniques for compression.
DWT improves the quality of compressed image. Back-propagation algorithm
can be extensively used as a learning algorithm in ANN. BPNN comes under
Feed-Forward Neural Network Architecture. Error correction learning rule is
particularly used in Neural Network (NN). This is a very efficient algorithm for
image compression, which works with the architecture of ANN. Then, the
performance analysis of different images is carried out (on application of three
different algorithms). The results clearly explain that hybrid image compression
using hybrid DWT-BP(Discrete Wavelet Transform-Back Propagation) provides
better CR and PSNR.
Chapter 7: This chapter explains the design methodology of Fractal based
Neural Network Radial Basis Function (FNNRBF) for image compression.
Generally, to store the digital images, there is a need of large amount of data and
it consumes more time for transmission and storage. So, the image compression
21
technique of this chapter is used to overcome the storage and transmission costs.
The implementation of this technique shows the effectiveness in terms of
compression of the medical images. Also, a comparative analysis is performed to
state that the proposed system is effective in terms of CR, PSNR, memory space
and execution time.
22
CHAPTER 2
LITERATURE SURVEY
2.1 INTRODUCTION
Compression means making the file size smaller by rearranging the
information in the file. Compressing imagery is different from zipping files.
Imagery compression changes the system and content of the information within
a file. Image compressions may used to rearrange the data and decay it to
achieve desired compression level, depending on the compression ratio. The
sacrifice of the data may or may not be noticeable. The quantity of image
compression can be influenced by the type of imagery. Higher compression ratio
can be achieved in portions of the image that have similar tone, such as water
area that have the same shade. Image compression is one of the brightest
disciplines in image processing. Images acquired need to be stored or transmitted
over long distances. Untreated image occupies more memory and hence need to
be compressed. Due to the demand for high quality video on mobile platforms
there is a need to compress untreated images and reproduce the images without
any degradation. Lossy compression is the reverse of lossless image
compression. It is used to compress images and video files. Lossless
compression is a family of data compression that permits the genuine data to be
perfectly reconstructed from the compressed data. Superfluous information is
removed in compression and added during decompression. Almost every lossless
compression programs do two things in sequence: the first step generates a
statistical framework for the input information and the second step uses this
framework to map input information to bit sequences in such a way that
"probable" (e.g. frequently encountered) information will make shorter output
than "improbable" information.
23
2.1.1 Description of the Image Compression block diagram
An image consists of large data and requires more space in the
memory. If more number of data is required for transmission then it takes much
time to deliver the data to the receiver. Thus by using image compression
techniques the time consumption can be greatly reduced. In this method, the
elimination of redundant data in an image can be possible. The image which is
compressed occupies less memory space and less time to transmit in the form of
information from transmitter to receiver. Compression means to make file size
smaller by reorganizing the data in the file. Data that is duplicated or has no
value is saved in shorter format or eliminated, greatly reducing the file size.
Compressing imagery is different than zipping files. Image compression
reorganizes the data and may degrade it to achieve desired compression level,
depending on the compression ratio. If there is better compression ratio, the
smaller the file size here more data is packed into smaller space, but lower the
quality of the compressed product.
Figure 2.1 Block Diagram of Image Compression
This figure explains the block diagram of Image Compression. First
we need transform the input Image using Forward Transform and again us
quantizing the input Image using Quantization and then us following the Entropy
encoding and finally we are getting the Compressed Image. Now we can store or
transmit the compressed Image. These are the steps followed in Compression
Techniques. Here, the Image Compression Techniques is divided into two types
namely Lossy and Lossless Techniques. In lossy, some information is lost during
compression of Image whereas in lossless compression no information is lost
Input Image
Forward Transform
Quantization
Entropy Encoding
Compressed Image
24
during Image Compression. Discrete Wavelet Transform (DWT) and Fractal
algorithm comes under Lossy Compression. Huffman Coding comes under
Lossless Image Compression.
2.1.2 Image format
In general, JPEG format of CT, MRI, and PET medical images are
used to analysis the quality parameters during image compression.
In JPEG format, the degree of compression can be adjusted, allowing
a selectable trade-off between storage size and image quality. This image format
is best for shooting of snap shot with realistic scenes. JPEG is a regularly used
method of lossy compression for digital images than Bitmap Image File (BMP)
and Tagged Image File Format (TIFF) formats.
2.2 LITERATURE REVIEW
It is of utmost importance to discuss about the basics of multi modal
medical image compression, so that the research community could have a better
idea about the processing of CT, MRI and PET image compression.
Jian-Jiun Ding et al [36] have presented a variable length coding
named as Huffman code which is mostly used to increase coding efficiency. It
widely uses the Huffman source-coding algorithm in order to generate the
uniquely decodable Huffman code with a minimum expected codeword length
when the probability distribution of a data source is known to the encoder.
Pawel Turcza et al [55] have proposed image compression using
Huffman coding which is based on an integer version of a Discrete Cosine
Transform and a low complexity entropy encoder making use of an adaptive
Golomb–Rice algorithm, which can be efficiently used in Huffman tables.
25
Ankita Vaish et al [13] have used PCA and Huffman coding. Set of
principal components (PCs) are used for reconstruction. Ill effects of using
number of PCs for reconstruction are overcome by further quantization. Coding
redundancy is removed by Huffman coding.
Arif Sameh Arif et al [15] have introduced a new framework based on
grouping of images, the correlation of pixels. Combination of Run-length coding
and Huffman coding are used. Significant improvement in compression is
achieved.
Jagadish et al [33] have explained that the objective of image
compression technique is to reduce the amount of data required for representing
sampled digital images. It is concluded that Huffman coding is the most efficient
technique for image compression and decompression.
Xiaofeng Li et al [80] have formulated two different stages in lossless
compression scheme related to the medical image compression. At first, current
pixel is predicted from the least-square-based prediction coefficients. Secondly,
residual image is formed by Huffman coding.
Mohamed Abo-Zahhad et al [7] have proposed image compression
(DPCM, DWT and the Huffman) approach. If the first image is pre-processed by
DPCM, its output will undergo wavelet transformation. Resulting coefficients
are encoded using Huffman coding.
Tajallipour et al [72] have presented an efficient adaptive LZW data
compression algorithm. Encoding considers customized library and also
considers custom valued threshold. The library size as well as threshold
parameters are adjustable.
Ng et al [50] have proposed an effective data re-ordering methods, the
SBI technique, that takes care of pre-processing stage in LZW algorithms. With
26
SBI, the dictionary matches grow dramatically and that lead to improved
performance.
Chiang et al [24] have developed an adaptive lossy LZW algorithm
that employs an adaptive mechanism for threshold.
Patil et al [54] have suggested automated multiclass diagnosis of
Dementia, related to the category of MR images (MRI) like human brain. 1D
histogram derived from 2D MR images of types like brain image is compressed
using DCT. A set of DCT coefficients was considered as features for
classification by ANN. This feature helps in identifying a person in distress
either by Huntington or Mild Alzheimer or Alzheimer disease. Classification
rate of 100% is obtained.
Christophe et al [12] have presented a scheme that involves DCT,
Kohonen map based vector quantization, first-order predictor based differential
coding and entropic coding.
Debin Zhao Wen Gao et al [28] have proposed block-based DCT that
enhances DCT applications to compression, retrieval and pattern recognition of
images. Morphological representation of DCT coefficients (MRDCT) is utilized.
Ci Wang et al [14] have suggested a new DCT-based MPEG-2
Moving Picture Experts Group transparent scrambling algorithm using INTRA
blocks. The computation burden is very low and effects can be easily controlled
by the operator. The algorithm has little influence on output bit rate.
Jie Liang et al [37] have proposed a structure for Linear-Phase
Praunitary Filter Banks (LPPUFB) in time as well as frequency domain post-
processing of the DCT. This structure enables the design of DCT based LPPUFB
with partial-block overlapping and variable-length filters. A DCT-oriented
initialization method is developed for improved convergence
27
Renato et al [59] have introduced an approximate transform of the
matrix Rows which is constructed by utilization of diverse mathematical
structure for design hybrid algorithms.
Merav Huber-Lerner et al [46] have identified Hyper Spectral (HS),
the image sensor types to compute the repentance of each pixel and created a 3-
D representation of the recorded scene. But the HSI takes large storage space
and enormous transmission time. PCA-DCT (means principle component
analysis which is followed by the discrete cosine transform) system mingles the
PCA’s ability to remove the unwanted background from the minute quantity of
components.
Yung-Gi Wu et al [84] have proposed an adaptive sampling algorithm
in which significant coefficients are calculated by the difference between exact
points and projected points. Recording or transmitting the important coefficients
only attain the target of compression. In decoder part, a linear equation is
engaged to reconstruct the coefficients.
Yung-Gi Wu et al [85] have presented a strategy that regards the DCT
in the form of band pass type of filter to decompose a sub block into a number of
equally sized bands. The high similar property among bands is found using
similar property, the bit rate of compression is greatly reduced.
Vasanthi Kumari et al [76] have presented an image compression
system by means of graph cut and also utilization of wavelet transform. Initially,
block partition procedure is carried out, and then dissimilar blocks are selected
by applying graph cut algorithm. The Differential Pulse Code Modulation
(DPCM) is utilized for raising the compressibility. Finally, the transformed
image is given to Huffman-encoder.
Alagendran et al [5] have investigated the several types of medical
image compression techniques.
28
Bhammar et al [19] have reviewed several ways of image
compression. The transmission (with less time consuming) of high quality digital
images necessitates the compression-decompression technique, which is to be
simple, high degree of quality image and completely lossless.
Reny Catherin et al [60] have surveyed the various lossy image
compression techniques. From their survey, the crucial inferences are derived:
DWT, SPIHT and DCT provide higher compression ratio and worthy output
images. Though the performance is disturbed by of noise, Soft computing based
compression technique works well in vigorous environment and offers high CR
and PSNR. NN procedure gives better-quality of reconstructed images as it
rejects blocking effects accompanying with DCT.
Sridevi et al [69] have outlined the contrast of compression methods
for instance JPEG2000 Max-Shift ROI Coding, and other related standards like
JPEG2000 Scaling Based ROI Coding, DCT, Shape Adaptive Wavelet
Transform and Scaling Based ROI, DWT and Sub band block Hierarchical
Partitioning. These concepts are evaluated with the utilization of CR and
compression quality.
Vaclav Simek et al [75] have illustrated the particulars around the
acceleration of 2D wavelet related to the medical image compression using
MATLAB and the Compute Unified Device Architecture (CUDA). Acceleration
of processing flow exploits all the immense parallel computational power
obtainable by the modern NVIDIA GPU. A computing system is programmed
by means of C language with the CUDA. similarly, a number of good-looking
features are exploited for a wide class related to rigorous data parallel
computation tasks.
Praisline Jasmi et al [56] have Observed the similarity of image
compression techniques with different coding like Huffman coding, DWT, the
Fractal coding. Compressed images are obtained by utilizing these three
29
algorithms on different types of input image. Efficiency evaluation is made
between compressed images by using the various parameters among these
compression techniques.
Jaffar Iqbal Barbhuiya et al [32] have made a comparative study upon
image compression using DCT and DWT. DWT algorithm does better than DCT
algorithms in terms of Compression, MSE and PNSR.
Kai-jen Cheng et al [40] have proposed Binary Embedded Zero tree
Wavelet algorithm which is related to the newly define two tree structures. In
BEZW algorithm the bit streams are arranged in order based on their
importance, so that the reconstruction fidelity be subject to the set of recovered
bit planes. In 3-D-BEZW algorithm, which is competitive, the prevailing pass
attains the quantization and significance test.
Priya Pareek et al [58] have recommended RLE as an suitable method
for compressing any form of data irrespective of the information contented.
However alternative RLE schemes can encode data along the length of the
bitmap, and the columns are to encode the bitmap in the form of 2-D tiles and
also to encode pixels in the diagonal form are made in to a zigzag form.
Kuppusamy et al [43] have suggested that Fractal image compression
is a quite useful technique, which can be mainly used in the compression of
medical image as well as color image. Fractal compression comes under the
lossy compression method for digital images. The main idea is to decompose the
image into segments by using standard image processing techniques such as
color separation, edge detection, and spectrum and texture analysis.
Sophin Seeli et al [68] have affirmed that Fractal image compression
is based on fractals of various images. The two main advantages of changing the
images to fractal data are 1) the memory size of the compressed image is much
lower than the memory size of the original image, 2) Fractal image Compression
30
can be used to measure the parameters like Compression ratio (CR), Peak Signal
to Noise Ratio (PSNR), Bits per pixel (Bpp) and also many other parameters.
Taha mohammed Hasan et al [71] have designed an Adaptive Fractal
Image Compression (AFIC) algorithm to reduce the long processing time of the
Fractal Image Compression (FIC). AFIC can be used to speed up the encoding
process and achieve a higher compression ratio, with a slight diminution in the
quality of the reconstructed image. In comparison with some methods, AFIC
spends much less encoding time and offers higher compression ratio of the
quality of the reconstructed images.
Al-Fahoum et al [7] have implemented a combination of Fractal and
Wavelet transform that can be used in image compression. These methods can
be mainly used in X-Ray Angiogram. First, the image is decomposed using
Wavelet transform. The smoothness of the low frequency part of the image
appears as an approximation image with higher self-similarities. Therefore, it is
coded using a Fractal coding technique.
Anil Bhagat et al [11] have found that Fractal image compression can
take advantage of redundancy in scale, but its operating principles are very
different from other transform coders. Images are not stored as a set of quantized
transform coefficients, but instead as fixed points of maps on the plane. Just as
the fern has details at every scale, so the decoded image has no natural size and
it can be decoded at any size.
Jianji Wang et al [35] have introduced FIC scheme created on the
concept of affine comparison among two blocks in FIC remains alike the Total
value of Pearson’s Correlation Co-efficient (APCC). Every block is categorized
by means of an APCC-based block classification process. Secondly domain
blocks are sorted based on APCCs.
31
Jyh-Horng Jeng et al [39] have proposed a Huber Fractal Image
Compression (HFIC), in which the linear type of Huber regression system is
used with the robust data entrenched in the form of encoding technique of
fractal. Once the raw image is modified by means of unwanted noise, FIC
scheme should be insensitive. Due to HFIC the computational cost is increased,
we move to Particle Swarm Optimization (PSO) technique that reduces the time
of search.
Chaudhari et al [23] have suggested a wavelet transform related fast
fractal image coding. Fast Fourier Transform (FFT) approach related to fractal
image coding by flexible quad tree partition is applied here. The resemblances
present among wavelet sub tree are utilized for predicting coefficients of better
scale with the coarser scale by means of affine transformation.
Sridharan Bhavani et al [70] have discussed the concepts related to
Fractal based coding processes like standard Fractal coding, and other
algorithms as quasi lossless Fractal coding in addition to improved quasi lossless
fractal coding. The process of machine learning correlated model is utilized for
plummeting the encoding time.
Omar Arif et al [52] have elaborated the use of Mercer kernel methods
in statistical learning theory which provides strong learning capabilities, as seen
in Kernel Principal Component Analysis (KPCA) and Support Vector Machines
(SVM). The technique takes advantage of the universal approximation
characteristics of generalized radial basis function neural networks to
approximate the empirical kernel map associated to KPCA or SVM.
Long Zhang et al [45] have introduced a discrete continuous
procedure used for the edifice of a RBF prototype. Initially orthogonal least
squares (OLS) build forward stepwise selection. It is followed by Levenberg–
Marquardt (LM) related parameter optimization to speed up convergence,
connection amongst the hidden nodes too output weights.
32
Arunpriya et al [17] have proposed an approach consisting of three
stages from preprocessing, and moving to feature extraction and finally the
classification to process the image. The tea leaf shape images can be recognized
precisely in the stage of preprocessing with the help of fuzzy de-noising by
means of dual tree discrete wavelet transform and the digital morphological
feature concept is used to increase classification accuracy. Further, Radial Basis
Function (RBF) is used for efficient classification.
Tiruvenkadam Santhanam et al [74] have developed a Neural Network
for weather forecasting on various issues gained from weather-related
professionals. He evaluates the concept of Radial Basis Function (RBF) with
other neural network concept Back Propagation (BPN) are used to test the
effectiveness of different forecast techniques. His results on this subject prove
that radial basis function outperforms the back propagation.
Arun Vikas Singh et al [16] have combined the concepts of both
wavelets transform and Radial Basis Function Neural Network laterally with
vector quantization. In this approach the image is disintegrated into a set of sub
bands states diverse coding and quantization practices are utilized.
Alex Alexandridis et al [6] have presented a novel approach designed
for training Radial Basis Function (RBF) networks, which increases the accuracy
and parsimony of the projected system. This procedure is based on a non-
symmetric variant of the Fuzzy Means (FM) process.
Panda et al [53] have perceived a neural network based Image
Compression with Back Propagation algorithm which is widely used in all the
areas to compress the image.
Abdul Khader Jilani Saudagar et al [1] have utilized Medical Image
Compression (MIC), which is a basic but important factor in telemedicine. It is
required to have an algorithm for compression of medical imaging modalities
33
like CT, MRI and ultrasound for providing medical care’s to patients in remote
locations. In telemedicine, the broadcasting of medical images requires a high
data rate so as to obtain a good quality transmission.
Birendra Kumar Patel et al [20] have developed image compression
technique based on back propagation neural network with Levenberg-marquardt
algorithm. The training algorithm and back propagation neural network are used
to improve the performance and to reduce the convergence time and to provide
high compression ratio as well as low distortion.
Anna Durai et al [14] have recommended Back Propagation
Algorithm. The efficiency can be decreased if the input image contain a quantity
of dissimilar gray levels with narrow modification amongst neighborhood pixels,
it takes enormous time to converge the image.
Prema Karthikeyan et al [57] have performed image compression
using Back- propagation algorithm in multi-layer neural network. The network
with three layers, input, hidden and output is used. Both the input and output
layers have the same number of neurons. The input and output are connected to
each network the compression can be done with the value of the neurons at the
hidden layer.
Shiqiang Yan et al [66] have selected Neural Network by means of
chaotic neuron. A Chebyshev chaotic charting is utilized to build the Neural
network.
Vilas et al [79] have preferred an ANN with feed forward back
propagation system designed for image compression. The Bipolar Coding
System in addition to LM algorithm helps to obtain a satisfactory result.
Dipta Pratim Dutta et al [29] have utilized ANN that adapts the
psycho visual features, the concept is mostly dependent on the information
34
confined in images. The algorithms preserve utmost of the appearances of the
data in a lossy manner, further maximize the compression performances.
Vilas Gaidhane et al [78] have used Feed Forward BPNN method
along with PCA technique trained by considering the different number of
hidden neurons.
Anjana Jianyu Lin et al [54] have evaluated the performance of the
Two-band analysis–synthesis filters used mainly meant for compressing raw
images. Both the filter concepts (FIR and IIR filters) have been deliberated.
Hybrid FIR–IIR analysis–synthesis filters is located to make best use of
compression performance.
Chakrapani et al [21] have presented an inimitable iterated function
system (IFS) entailing of the assembly with the affine transformation. FIC hires
a distinct nature of IFS called as PIFS (also called as local IFS). Collage
Theorem is hired for PIFS and gray scale images, which is equal to IFS for
binary images.
Alok kumar singh et al [9] that most popular to scale back the
redundancy and irrelevant within the image, for storing and transferring the
image with efficiency.
Chander mukhi et al [22] have utilized encoders that used the DCT to
perform transform coding. The DCT maps time domain signals to frequency
domain. It compresses the frequency domain spectrum by truncating low
intensity regions. The (Discrete Wavelet Transform) DWT which offers a more
robust solution in essence, may be computed by using a collection of digital
filters at a quicker rate by analyzing the complete signal. The DWT captures a
lot of info than the DCT and produces higher results. The DWT separates the
images with high frequency elements from the remaining elements of the images
35
and resizes the remaining components and rearranges them to make a new
transformed image.
Jiaji Wu et al [34] have proposed lossless compression algorithmic
rule containing two stages to support the weighted motion compensation and the
context based modeling. The algorithmic rule makes use of the weighted motion
compensation for getting the motion vector supported bizarre motion in aurora
images. Afterward, the context based modeling is pooled with the motion vector
and the results obtained provide ascendancy of algorithmic rule.
Yongjian Nian et al [83] have proposed a lossless compression
algorithmic rule for hyper spectral images supported distributed source coding.
The algorithmic rule processes block with same location and size in every band.
The importance varies from block to the other block on the spectral orientation.
The algorithmic rule weighs the energy of every block beneath the target rate
constraints introduced. Additionally, a linear prediction model is employed to
construct the aspect data of every block for Slepian–Wolf coding.
Yeo et al [81] have proposed Feed Forward NN trained with the back
Propagation algorithmic rule to compress grayscale medical images. A system of
three stages of the hidden layer Feed Forward Network (FFN) is utilized
unswervingly. Once trained with sufficient variety of sample images, the
compression method is tested on the target image. Compression is then achieved
with the help of smaller variety of the hidden neuron as compared toward the
extent with the image pixel.
Ajay Kumar Bhagat et al [4] have developed a hybrid methodology
victimization SVD and DWT. This can be a suitable method to update the
decomposition, as well as the premise images. The DWT is employed to divide
the image into sub bands. Because the edges consider LH sub band, HL sub band
and HH sun-band, the impact of fusion is to be minimized.
36
Ferni Ukrit et al [30] have developed compression technique that adds
the super spatial Structure Prediction with motion estimation and motion
compensation to obtain higher compression ratio. This is often enforced by an
easy block matching method Binary Tree Search.
Shruti Puniani et al [67] have mentioned some basic compression
techniques like Huffman, LZW coding, VQ compression.
Sridhar et al [62] have developed a wavelet based transform method
and NN based concept for image compression that makes use of both wavelet
transformations and NN. They also discussed how the coefficients present within
the less frequency bands are compressed by making use of Differential Pulse
Code Modulation (DPCM).
Abirami et al [2] have evaluated the performance of wavelet based
Support Vector Machines with completely different combination of methods of
wavelets and kernel function concepts. SVM regression is applied to wavelet
based coefficients to approximate the obtained coefficients from wavelets.
Higher compression is achieved by removing the redundancy.
Yongfei Zhang et al [82] have utilized quantization as a core
component for wavelet-transform related lossy image compression that
successfully minimizes the visual redundancy.
Saravanan et al [63] have developed a compression technique to
obtain more compression ratio by reducing number of source symbols. The
technique adapted to reduce the quantity of source symbols is by combining
symbols to make a reduced symbol. Therefore, the Huffman codes are generated
for the reduced symbols.
Nikita Bansal et al [51] have developed a scheme for image
compression with the use of both DCT concept and DWT concept provided
hybrid compression model. High energy compaction property is present in DCT
37
and usually utilizes a fewer computational resources and multi resolution
transformation in DWT.
Chunlei Jiang et al [26] have proposed hybrid compression technique
that makes use of both fractal concept and SPIHT (set partitioning in hierarch
tree) concept. It utilizes total landscape characteristics and also the human visual
characteristics. The image is divided in the form of low and high frequency type
sub bands, afterwards the low subband frequency makes use of fractal type
technique.
Ali Al-Fayadh et al [73] have steered a hybrid lossy compression
technique that makes use of both classified vector quantization, singular value
decomposition. The methodology is termed as hybrid classified vector. It
involves a better classifier technique based on gradient within the spatial
domain, and utilizes ac coefficients present in the DCT coefficients. It
evaluates orientation of block while not using any form of threshold that leads to
hi-Fideld medical image compressed. The Singular value decomposition is
accustomed to generate the classified codebooks.
Mohamed El Zorkany et al [48] have developed a DCT compression
technique. This technique combines the compression ratio of Neural Network
(NN) and Vector quantisation (VQ) with the energy-compaction property of
DCT. It must increase the ratio of compression and also preserve the image
quality of reconstruct image, so image is compressed by NN.
Shaou-Gang Miaou et al [65] have proposed a technique that utilizes
both JPEG-LS and the concept of interframe coding with motion vectors to
provide the better compression performance. Since, the interframe correlation
between adjacent images in a sequence of medical image is typically not as high
as that in the case of general video image sequence. the interframe technique is
activated only if the interframe correlation is sufficiently high.
38
Robina Ashraf et al [61] have introduced a technique, that provides
high CR's for images of type radiographic without any loss in quality of
diagnostic. During this process, an image is compressed with loss during first
stage at a high compression ratio and then blunder image again compressed
lossless. Finally ensuing compression isn't solely austerely lossless, other than
additionally it is expected to reach compression ratio to a high ratio, particularly
if lossy type of compression technique is chosen correctly. Neural Network
Vector Quantizer (NNVQ) will be employed as same as the lossy system.
Kaur et al [41] have developed an adaptive image-coding algorithmic
rule for compression of medical ultrasound (US) image within the wavelet based
domain. The histograms of wavelet related coefficients of the sub bands within
the North American nation images are heavy-tailed and might be highly shaped
by making use of the generalized Student’s t-distribution. Exploiting the
statistics, adaptive image coder named as JTQVS-WV is implemented, Rate–
Distortion (R–D) optimized quantizer and R–D optimum thresholding are
predicated.
Bairagi et al [18] have proposed an automated, efficient and low
complexity, lossless, scalable RBC for Digital Imaging and Communications in
Medicine (DICOM) images. RBC is utilized and the regions are segmented in a
number of different types based importance of region and the subjecting
changing bit-rates for optimal performance. Utilization of the integer wavelet
transform and technique of limited bit rate compression is used in minor
important regions which helps to reconstruct the image of desired quality.
Tamilarasi et al [73] have proposed an extension to the WT in separate
two dimensions by making use of non separable and the directional filter banks.
As MI is involved, the diagnosis part (ROI) is very vital. Initially ROIs are
segmented from the intact image by making use of NN related to FL technique.
Contour let transform is then applied to ROI portion. The region of less
39
significance is made use of DWT and finally modified embedded zero tree
wavelet algorithm is applied and it uses six symbols instead of four.
Jonathan Taquet et al [38] have proposed a different hierarchical
process for resolution scalable lossless compression, Near-Lossless (NLS)
compression. It utilizes adaptability from the DPCM schemes, with the concept
of new type of hierarchical oriented predictor which provides scalable
resolution. Then the hierarchy oriented prediction performance is less when
utilized for smooth images. New predictors are introduced. They are
dynamically optimized employing a LS criterion.
Harjeetpal singh et al [31] have presented DWT and DCT
implementation, because these are the lossy techniques. He extended his
research with the Huffman encoding technique. At last he implemented lossless
technique. Its PSNR and the MSE enhance the results when compared to the
previous algorithms.
Monika Narwal et al [49] have proposed SPIHT –DCT algorithm for
compression of an image. SPIHT and DCT both have some limitations. By
combining them the limitations are overcome.
Kesavamurthy Thangavelu et al [42] have developed lossless method
related to volumetric MI compression process and decompression process by
making use of adaptive block concept related encoding technique. Further
algorithm is tested with various collections of CT color image with use of
MATLAB. The digital imaging and communications in medicine images be
compressed with the help of the proposed algorithmic rule and store as DICOM
format image. The contrary step of adaptive block related algorithm is utilized to
reconstruct actual image losslessly with the use of compressed files realted to
DICOM.
40
Vidhya et al [77] have proposed an algorithm that extracts edge
information of MI by making use of fuzzy edge detector. The image is
disintegrated by utilizing the concept of cohen daubechies feauveau wavelet.
The hybrid technique is a combination of JPEG2000 and SPIHT. The
coefficients present in the approximate sub bands are encoded with the help of
tier-1 part of JPEG2000. The coefficients present in thorough sub bands are
encoded by making use of SPIHT. Finally, quality images are obtained from this
process at a lower bit rate compare to other compression technique.
Adnan Khashman [3] have related the Neural Network with the image
of radiograph contents to obtain image compression ratio at optimal image
quality. After the training, neural network provides the perfect Haar wavelet
compression ratio related to x-ray images when they are subjected to the
network.
2.3 CONCLUSION
In essence, this thesis summarizes the selected literature survey that
has been carried out in the area of Discrete Cosine Transform (DCT), Discrete
Wavelet Transform (DWT), Fractal Algorithm, Neural Network Back
Propagation (NNBP), Neural Network Radial Basis Function (NNRBF) and
Hybrid Techniques based approaches for Medical Image Compression.
41
CHAPTER 3
METHODOLOGY
3.1 DISCRETE COSINE TRANSFORM (DCT)
A Discrete Cosine Transform (DCT) communicates a limited
grouping of information focuses as far as an entirety of cosine capacities
wavering at various frequencies. DCTs are useful in various fields such as
science and building, from lossy compression of sound (e.g. MP3) and images
(e.g. JPEG) (where little high-recurrence fragments can be disposed of) to
otherworldly procedures for the numerical game plan of mostly differential
scientific explanations. The utilization of cosine as opposed with sine capacities
is basic for compression, since it turns out (as portrayed beneath) less cosine
capacities that are expected to surmise a common flag. For various comparisons
the cosines express a specific decision of limit conditions.
Specifically, a Discrete Cosine Transform relates to a Fourier
transform that resembles Discrete Fourier Transform (DFT). DCTs are
proportional to DFTs in twofold length and doing with genuine information with
even symmetry, where in a considerable difference the data and/or yield data are
moved. There are eight standard DCT assortments of which four are universal. It
is mainly used for the specific image compression and one of the main tissues
according to the growth of technology is found that in the midst of a colossal
measure of data deferral, with such gigantic data can regularly show challenges.
The fundamental motivation behind image compression is to minimize the size
with no adjustment in the nature of image, so it is valuable to store the substance
in a given measure of circle furthermore it is helpful for transmission.
42
Figure 3.1 Block diagram of DCT
Figure 3.1 shows that the given source of input image undergoes
Discrete Wavelet Transform, where quantization is done with the quantization
table. Thus, the quantized values are given to the entropy encoder. The entropy
values are subjected to pre-processing steps, which gives the compressed image
data.
3.1.1 Image Compression in DCT
JPEG is a standard leading group of trustees which has its own
particular inceptions inside the International Standard Organization (ISO). JPEG
may be adjusted to make little compacted images that are of reasonably low
quality in appearance, yet in the meantime fitting for a few applications. JPEG
gives a compression procedure that is set up for compacting unending tone data
with a pixel importance of 6 to 24 bits with sensible speed and viability.
JPEG is outlined especially to discard information that is not visible to
the human eyes. Slight changes in shading cannot be seen incredibly by the
human eye. Hence JPEGs lossy encoding is capable of storing with the dark
scale part of a image and to be more pointless with the shading.
The Discrete Cosine Transform (DCT) helps in separating the images
into parts (or nebulous vision sub-bunches) with respect to the image visual
quality. The DCT resemble the discrete Fourier change and it changes a sign or
image from the spatial zone to the repeat space. DCT has various inclinations.
43
i. It has been realized in single facilitated circuit
ii. It can pack most data in smallest coefficients
iii. It minimizes the piece like appearance called blocking that
outcome when purposes of restriction between sub-images persuade the chance
to be perceivable.
Figure 3.2 Conversion of special domain to frequency domain
3.1.2 DCT Encoding
The general scientific formula for a 2D (M by M picture) DCT is
described by the accompanying comparison:
+
+
= ∑∑−
=
−
= Mvx
MuxyxfvCuC
MvuF
M
X
M
Y 2)12(cos
2)12(cos),()()(2),(
1
0
1
0
ππ (3.1)
for u=0,…..,M-1 and v=0,….., M-1
Where M=8 and (3.2)
The operation of the DCT is according to the accompanying
• The input signal is given as an image by the matrix M × M;
• The intensity of the pixel is given in row(i) and column (j) matrix
represented by f(i, j)
• The coefficient of DCT in row k1 and column k2 of the DCT grid is F(u,
v)
• The signal vitality lies at low frequencies for some images, which indicate
the upper left corner of the DCT.
44
• Compression is expert ensuing to the lower right values, which address
higher frequencies, and are regularly sufficiently little to be disregarded
with minimal visible distortion.
• The input image of DCT is 8×8. This exhibits every pixel's grey scale
level.
• Every 8 bit pixel has levels from 0 to 255.
3.1.3 Compression Steps in DCT
The following steps are followed in image compression using DCT
• The image selected is considered as blocks of size 512 × 512.
• A high contrast image has pixel values varying from 0 to 511. Yet,
DCT considers pixel values extending in the range from - 128 to 127.
Along with this, each and every block is assigned to work in its range.
• For computing the DCT grid, equation 3.1 is used.
• Every block is connected to the DCT by increasing the adjusted block
with DCT framework on the left and transposes of DCT network to its
right side.
• Each and every block is compressed through quantization technique.
• The quantized framework is entropy encoded.
3.1.4 Quantization Steps
Quantization is obtained through compressing an arrangement of
qualities to singular quantum esteem. For this situation, the discrete images
amount is decreased and the stream ends up being more compressible. A
quantization grid is used as a part of blend with a DCT coefficient cross section
to finish the change. Quantization is the progression where the majority of the
45
compression happens. DCT truly does not pack the images, since it avoids
lossless compression mode.
Quantization makes use of the way that the higher repeat segments are
less essential than the lower repeat parts. It awards various levels for image
compression and quality through choice of particular quantization systems. In
this way, the quality levels going from 1 to 100 can be picked up, where '1' gives
the poorest picture quality and '100' gives the best quality. JPEG board
recommends grid with quality level '50' as standard framework.
Quantization is a master in isolating the changed image structure.
Estimations of the resultant structure are then balanced off. In the resultant
lattice coefficients placed close to the upper left corner have lower values.
3.1.5 Entropy Encoding
After quantization, the high repeat coefficients will be zeros. To know
the number of zeros, a crosswise sweep framework is being utilized for yielding
the long string zeros. Once a bit has been changed over to an extent and
quantized, the JPEG pressure figuring then takes the result and changes over it
into a one dimensional direct appearance or vector of 64 qualities, playing out an
across yield by selecting the parts in the numerical deals appeared by the
numbers in the cross section underneath:
0 1 2 3 4 5 6 7
0: 0 1 5 6 14 15 27 28
1: 2 4 7 13 16 26 29 42
2: 3 8 12 17 25 30 41 43
3: 9 11 18 24 31 40 44 53
4: 10 19 23 32 39 45 52 5
5: 20 22 33 38 46 51 55 60
46
This places the components of the coefficient block in a sensible
request of expanding frequency. Since, the higher frequencies will probably be
zero after quantization. This tends to group zero qualities in the high end of the
vector.
3.2 DISCRETE WAVELET TRANSFORM (DWT)
The Discrete wavelet transform is processed independently for various
fragments of the time-space signal at various frequencies. Multi-determination
analysis: analyzes down the signal at various frequencies giving diverse
resolutions. It is useful for the sign having high frequency parts for brief lengths
and low frequency segments for long duration. E.g. images and video frames.
The wavelet transformation is made out of an arrangement of low-pass
and high-pass channels. The following channel arrangements can be connected
similarly as a discrete FIR channel in DSP, utilizing the MACP order, aside from
as different progressive FIR channels. The low pass channel performs an
averaging/obscuring operation, and is expressed as:
H=1/√2(1,1) (3.3)
The differencing operation of high pass channel is communicated as follows:
G=1/√2(-1,1) (3.4)
On any adjacent pixel pair, the complete wavelet transform can be represented in
a matrix format.
(3.5)
TN NT W AW=
First half: Applying 1D Transformation to Rows of Image
Second half: Applying 1D Transformation to Columns of Image
47
Where, A is the matrix representing the 2D image pixels, wavelet transformation
of the image.
1 1 0 0 0 0 0 02 2
1 10 0 0 0 0 02 2
1 10 0 0 0 0 02 2
1 10 0 0 0 0 02 2
1 1 0 0 0 0 0 02 2
1 10 0 0 0 0 02 2
1 10 0 0 0 0 02 2
1 10 0 0 0 0 02 2
NHWG
= = − − − − (3.6)
The consequence of the complete transformation, T, is made out of 4
new sub-images, which compare to the obscured images, and the vertical,
diagonal, and horizontal contrasts between the original picture and the blurred
image. The blurred representation of the image evacuates the subtle elements
(high frequency components).
Compressed Output
Figure 3.3 Block diagram of DWT
In the above Figure 3.3 the image is given as an input to get
compressed and it undergoes some of the preprocessing steps and it is followed
by some coding algorithm of wavelet transform. For further compression, the
arithmetic compression is used then finally reconstruction is done to get the
reconstructed image.
Input Image
Pre-Processing Wavelet
Transform Coding
Algorithm
48
3.2.1 Advantages of DWT
No need to divide the data coding into non-covering 2-D
pieces, it has higher compression proportions ratios from
blocking antiques.
Allows great restriction both in time and spatial frequency
domain.
Transformation of the entire image
Introduces innate scaling
Better recognizable proof of which information is pertinent to
human perception higher compression ratio
3.2.2 Wavelets used in Image Compression
Wavelets are signs, which are close-by in time and scale and generally
have a irregular shape. A wavelet is a waveform of sufficiently obliged term that
has a typical estimation of zero. The expression "wavelet" starts from the way
that they incorporate to zero.
There are two methods of compression. They are lossy and lossless.
Here, DWT is one of the algorithms in lossless method. DWT is considered as
one of the important methods for image compression, where there is no loss of
information during the compression of image. Wavelets have more advantages
over the compressing signals.
DWT can be applied to the process of image compression by using the
threshold value. Applying DWT can help us to get different levels of bands.
After deciding the threshold value, these values will neglect the certain wavelet
coefficients. In wavelet change, the deterioration of a specific image comprises
two sections, one is the lower recurrence or approximation of an image (scaling
49
capacity) and another is the higher recurrence or point by point part of a image
(wavelet capacity).
3.2.3 Aspects of Wavelets
DWT assumes an essential part to pack the given image without the
loss of any data in that specific image. DWT comes under the lossless sort of
image compression. Wavelets have more points of interest over packing signals.
The wavelet change is considered as the most beneficial and helpful
computational instruments for a variety of sign and image preparing
applications. Wavelet changes are basically utilized for images to decrease the
undesirable commotion and obscuring. Wavelet change has developed as the
most intense device for both information and picture pressure. Wavelet change
performs multi determination image analysis. DWT has effectively been utilized
as a part of numerous image preparing applications including noise reduction,
edge identification and compression.
When we apply high frequency (utilize high pass channel) on an
image, there are high varieties in the dark level between the two contiguous
pixels. So, edges are present in the image. When we apply low recurrence
(utilize low pass channel) on a image, there are smooth varieties between the
nearby pixels. So, edges are not produced. All data of image stays as same as
genuine picture data (it shows as estimate image). Figure 3.4 explains the three
level decomposition wavelet filters.
LL3 LH3 LH2
LH1
HL3 HH3
HL2 HH2 HL1 HH1
Figure 3.4 Three Level Decomposition Wavelet Filter
Utilizing DWT, images are decomposed into four sections:
approximate image, horizontal points of interest, vertical details and diagonal
50
details. When we apply high recurrence on an image, there are high varieties in
the grey level between the two nearby pixels. When we apply low frequency on
an image, there are smooth varieties between the adjacent pixels.
All data of picture stay as same as genuine image data (it shows as estimate
image).
Figure 3.5 2-D Discrete Wavelet Transform in image compression
• Wavelet transform is fundamentally the same to the customary Fourier
change; however, it depends on the little waves, called wavelet, which is
made out of time differing and restricted term waves. We use 2-D discrete
wavelet transform in image compression.
• The data sign will be shifted into low pass and high go parts through
examination channels.
• The human recognition framework has distinctive affectability to various
recurrence bands
– The human eyes are less touchy to high recurrence band
shading parts
3.3 FRACTAL ALGORITHM
Fractal image compression is a lossy compression strategy for
advanced images, in the view of fractals. This compression technique is the most
51
appropriate for surfaces as well as for normal images, depending on the way that
segments of an image frequently take after different segments of the same
image. Fractal calculations change over these parts into scientific information
called "Fractal codes", which are utilized to reproduce the encoded image.
3.3.1 Presentation about Fractal Algorithm
One of the lossy image compression techniques right now accessible
is the strategy for fractal image compression, created by Michael Barnsley and
his partners in 1987. The strategy is a restrictive innovation of Iterated Systems,
Inc., a firm helped to establish by Barnsley. Image compression techniques can
likewise be delegated either symmetrical or asymmetrical. Fractal image
compression, then again, is an illustration of asymmetrical strategies.
Asymmetric strategies take additional time/exertion compacting an image than
decompressing it. The thought is to do a large portion of the work amid the
compression.
Given a unique image (in digital, bit-mapped group), say B (here we
accept B is nonempty, generally there is not something to be compressed), with a
determination of M×N pixels, the image record comprises of a header took
succeeded by M×N cells of force information, one for every pixel. Given the
determination, the spatial directions of every pixel are suggested. The extent of
the cell connected with every pixel changes, contingent upon the sort of the
image as depicted beneath. This procedure is autonomous of the determination
of the original image.
The yield realistic will resemble the first at any determination, since
the compressor has found an IFS whose attractor reproduces the first one (i.e. an
arrangement of comparisons portraying the original image). Obviously, the
procedure takes a ton of work, particularly amid the quest for the suitable extent
districts. Be that as it may, once the compression is done, the FIF (Fractal Image
Format) record can be decompressed rapidly. In this way, the Fractal image
52
compression is uneven. The down to earth executions of a fractal compressor
offer diverse levels of compression.
3.3.2 Features of Fractal Algorithm
With the Fractal compression, encoding is significantly
computationally exorbitant because of the search used to find the self-similitude.
Decoding however is altogether fast. While this asymmetry has so far made it
unrealistic for steady applications, when video is reported for dispersion from
circle stockpiling or record downloads, fractal compression ends up being more
focused.
At the essential compression ratios, up to around 50:1, Fractal
pressure compression gives equivalent results to DCT-based counts, for instance,
JPEG. At high compression extents, the Fractal compression may offer prevalent
quality. For satellite imagery, extents of more than 170:1 have been proficient
with commendable results. Fractal video compression extents of 25:1-244:1 have
been defined in sensible compression times (2.4 to 66 sec/layout). The
compression adequacy increases with higher image multifaceted nature and
shading significance, appearing differently in relation to the fundamental
grayscale images.
3.3.3 Fractal Image Compression
Imagine a remarkable kind of photocopying machine that reduces the
image to be copied fundamentally and rehashes it three times on the duplicate
(see Figure. 3.6). We can watch that all the duplicates appear to join to the same
resultant image. Since the replicating machine lessens the information image,
any basic image sets on the copying machine will be decreased to a point as we
run the machine; more than once, it is just the position and the introduction of
the duplicates figures out what the resultant image resembles.
53
Figure 3.6 A photo copy machine that makes three reduced copies of the
input image
3.4 SET PARTITIONING IN HIERARCHICAL TREES (SPIHT)
SPIHT means Set Partitioning in Hierarchical Trees. This is proposed
by Pearlman in 1996. It is an image compressing method of DWT and it belongs
to lossless compression technique. SPIHT is a method of coding and decoding
the wavelet transform of an image. The spatial-orientation tree (or) the three-
Level Haar function of Wavelet transformation structure is used to describe how
an image gets split and compressed.
3.4.1 Haar Wavelet
In arithmetic, the Haar wavelet is a plan of arranging rescaled "square-
molded" capacities which together frame a wavelet family or premise. Wavelet
analysis resembles Fourier analysis, which allows a target limit over the time
duration and it needs to be addressed with respect to an orthonormal limit
premise. The Haar gathering is right now seen as the key allocated to wavelet
premise and comprehensively used as a teaching outline.
The Haar grouping was proposed in 1909 by Alfred Haar. Haar used
these abilities to give an outline of an orthonormal framework for the space of
square-integral limits on the unit between the time [0 and 1]. The examination of
54
wavelets, and even the expression "wavelet", did not come until much later. As
an exceptional case of the daubechies wavelet, the Haar wavelet is generally
called D2.
The Haar wavelet is moreover the least troublesome possible wavelet.
The particular obstruction of the Haar wavelet is that it is not consistent, and
thus not differentiable. This property can regardless be inclination for the
examination of signs with sudden moves, for instance, checking of the
equipment disillusionment in machines.
The Haar wavelet's function can be described as
(3.7)
Its scaling function can be described as
(3.8)
3.4.2 Formation of Cells
The smallest conceivable square matrix produced from the wavelet
disintegrated image, has the same level of wavelet decay structure as the original
image.
55
Figure 3.7 Formation of cells of parent-offspring conditions
The Haar wavelet transform is simple, and the better compression can be
achieved by other wavelet filters. It seems that the different wavelet filters produce
different results depending on the image type, but it is currently not clear what filter is
the best for any given image type. Regardless of the particular filter used, the image is
decomposed into sub bands, such that the lower sub bands correspond to the higher
image frequencies (they are the high pass levels) and the higher sub bands correspond
to the lower image frequencies (low pass levels), where most of the image energy is
concentrated (Figure 3.8). This is why we can expect the detailed coefficients to get
smaller as we move from high to low levels. Also, there are spatial similarities among
the sub bands. An image part, such as an edge, occupies the same spatial position in
each sub band. These features of the wavelet decomposition are exploited by the SPIHT
(Set Partitioning In Hierarchical Trees) method
3.4.3 Zero Tree Encoding
In zero tree based image compression arrangement, for instance EZW
and SPIHT, the point is to use the truthful properties of the trees with a particular
finished objective to adequately code the ranges of the tremendous coefficients.
56
Since, most of the coefficients will be zero or close to zero, the spatial zones of
the critical coefficients make up an expansive bit of the aggregate size of a run
of the mill compacted image. A coefficient (in addition a tree) is seen as
enormous if its degree (or sizes of a center and each one of its relatives because
of a tree) is over a particular edge. By starting with a breaking point, which is
close to the best coefficient degrees and iteratively lessening the edge, it is
possible to make a pressed representation of a image which consistently
incorporates better detail. In light of the structure of the trees, it is likely that if a
coefficient in a particular repeat band is insignificant, then every one of its
relatives (the spatially related higher repeat band coefficients) will moreover be
immaterial.
3.4.4 SPIHT Algorithm
• O(i,j): set of coordinates of all posterity of hub (i,j); children alone.
• D (i, j): set of coordinates of all relatives of hub (i, j); children,
grandchildren, incredible amazing, and so on.
• H (i,j): set of all tree roots (hubs in the most noteworthy pyramid level);
folks
• L (i, j): D (i, j) – O (i, j) (all descendents aside from the posterity);
grandchildren, incredible amazing.
Figure 3.8 Spatial Orientation Tree in SPIHT
57
3.5 INTRODUCTION TO NEURAL NETWORKS
For the most part, an image is isolated into a number of non-covering
pixel blocks, and sustained as examples for system preparing. In examination
with the vector quantization, the required and encoding/interpreting time is
substantially less. In any case, extremely restricted measure of compression is
accomplished, as it is misused just the connection between pixel inside of each
of the preparation designs.
The higher pressure proportion is accomplished in creating
progressive NN that costs vigorously because of the physical structure of the
NN. To make the picture pressure reasonable, it is compulsory to lessen the
gigantic size off the most picture information that in the long run diminishes the
physical structure of the NN. To decrease the size impressively, a few image
preparing steps like edge identification and thresholding are made and discussed.
The principle worry of the second period of the work is to adaptively decide the
structure of the NN that encodes the picture utilizing back engendering preparing
strategy.
Another strategy has been embraced while introducing the weight
amidst the middle layer and shrouded layer neurons in place of randomizing the
basic weight. Here, the spatial directions of the pixel of the image piece are
changed from two to one dimensional esteem and standardized with in '0' and '1'.
This methodology exhibits the quick rate of merging of the preparation
calculation and has been tried for various pictures. In this research, the
exploration of a managed learning calculation for fake neural systems i.e. the
error back spread learning calculation for a layered food forward system has
been executed for picture pressure and the investigation of the recreation
consequences of back propagation calculation is finished. There are two types of
learning algorithms in Neural Networks. They are supervised learning algorithm
and unsupervised learning algorithm.
58
Figure 3.9 Block diagram of Neural Network
Figure 3.9 explains that the input image is given to the training kit. Then, this
training kit will compare this input image with its database. After doing this
training process, by using supervised or unsupervised learning algorithms
compressed image is produced.
3.5.1 Back Propagation Neural Networks (BPNN)
Most of the image compression methods depend on the back
propagation feed forward neural network which is able to find possible solutions
to the problem and for application in many fields where high computational rates
are required. Initially the image is decomposed into numerous pixels using
image compression. These pixels are then encoded and given as the input
training pattern to the network which is to be transmitted and then reconstructed
at the receiver side. In the back propagation process, the entire network consists
of input layer, output layer and one or more hidden layers.
The spatial co-ordinates of the pixel value are encoded and converted
from two to one dimensional values and finally compressed when the inputs are
multiplied with their corresponding weight to get the total sum of the input. This
result of weighted sum undergoes sigmoidal function to yield output pattern.
This is the first phase or forward phase. Once the output is gained then the error
is calculated after which the process is propagated reversely by finding the
changes occur between the output and hidden layer, hidden layer and input layer.
Input Image Training Kit Database
Compressed Image
59
.
Figure 3.10 General Structure of BPNN
Initially the image is decomposed into numerous pixels using image
compression. These pixels are then encoded and given as the input training
pattern to the network which is to be transmitted and then reconstructed at the
receiver side. In the back propagation process, the entire network consists of
input layer, output layer and one or more hidden layers. When the input is given,
it gets multiplied with their corresponding weight to get the total sum of the
input. This result of weighted sum undergoes sigmoidal function to yield output
pattern. This is the first phase or forward phase. Once the output is gained then
the error is calculated after which the process is propagated reversely by finding
the changes occur between the output and hidden layer, hidden layer and input
layer.
Andrew et al [10] have proposed techniques explored two basic neural
systems (i.e., the BPNN-L and BPNN-R models) for online summed up network
reversal. In addition, two discrete-time Hopfield-sort neural systems (i.e., the
HNN-L and HNN-R models) are introduced for online arrangement of the
summed up converse. Seyun Kimand et al [64] have proposed this technique
unmistakably and portrayed CR, PSNR for cameraman image utilizing this
BPNN calculation. In his research, the packed image happens just at 1100, 1300,
1900 epoches etc.
60
In any case, our commitment is that we have changed the execution of
CR, PSNR from existing paper Leyuan Fang et al [44]. In this research, we get
the compacted image at less time i.e. inside of the normal of 22 epoches for
information restorative images. Change of the execution contrasted with existing
papers can be found.
3.5.2 Image Compression using Back Propagation
The usage of back engendering neural system calculation on image
compression framework with great execution has been illustrated. The back
engendering neural system has been prepared and tried for the examination of
various images. It has been watched that the merging time for the preparation of
back proliferation neural system is quicker. Diverse characteristics of
compression, for example, compression proportion, crest sign to clamor
proportion, bits per pixel are ascertained. It has been watched that it is essential
to change the compression proportion from .99 to .9556 in the event of
cameraman image.
It has likewise been watched that there is a remarkable change in crest
sign to clamor proportion from 19.3181 to 20.722. The versatile qualities of the
proposed approach give seclusion in organizing the engineering of the system,
which accelerates the handling as well as less vulnerable to disappointment and
simple for amendment. The procedure of introducing weights displays quick rate
of joining and utilizing the prepared weight sets, great nature of recovered
images is accessible at the desirable end.
A standout amongst the most famous NN calculations is back
engendering calculation, asserted that BP calculation could be separated into
four fundamental steps. Subsequent to picking the weights of the system
haphazardly, the back engendering calculation is utilized to Figure the vital
rectifications. The calculation can be deteriorated in the accompanying four
stages:
61
• Feed-forward calculation
• Back spread to the Output layer
• Back spread to the Hidden layer
• Weight redesigns
The calculation is halted when the estimation of the blunder capacity has
turned out to be adequately little. This is harsh and essential equation for BP
calculation.
3.5.3 Use of Image Compression in Back Propagation
On the other hand, here are a few circumstances where a BPNN may
be a smart thought:
• A vast measure of input/output information is accessible, yet you're not
certain how to relate it to the output.
• The issue seems to have overpowering unpredictability; however there is
plainly an answer.
• It is anything but difficult to make various case of the right conduct.
• The answer for the issue may change after some time, inside the limits of
the given data and yield parameters (i.e., today 2+2=4, however later on
we may find that 2+2=3.8).
• Outputs can be "fluffy" or non-numeric.
Training Algorithm
Step 1: Normalize the inputs and outputs regarding their greatest
qualities. It is demonstrated that the neural systems work better if
inputs and outputs lie somewhere around 0 and 1.
Step 2: The image is parted into non-covering sub-images. Say for
instance 512X512 piece image will be parted into 4 x 4 or 8 x 8 or
16 x 16 pixels.
62
Step 3: The standardized pixel estimation of the sub-image will be
the contribution to the nodes. The three-layered Radial Basis
Function Network will prepare every sub-image.
Step 4: The quantity of neurons in the hidden layer will be
intended for the fancied compression. The quantity of neurons in
the output layer will be the same as that in the input layer.
Step 5: The output of the information layer is assessed utilizing the
exchange capacity for a radial basis neuron.
Step 6: The input layer and output layer are completely associated
with the hidden layer. The weights of neurotransmitters associating
info neurons and concealed neurons and the weight of
neurotransmitters interfacing shrouded neurons and the weight of
neurotransmitters interfacing concealed neurons and yield neurons
are introduced.
Step 7: The contribution to the hidden layer is figured by
multiplying the comparing weights of neural connections. The
hidden layer units assess the yield utilizing the exchange capacity
for a spiral premise neuron.
Step 8: The contribution to the output layer is processed by
increasing the relating weights of neural connections. The output
layer neuron assesses the yield utilizing direct capacity.
Step 9: The neural system is tried for various images. At that point
the yield downsizes to the first gray scale range.
Step 10: Calculate image quality parameters.
63
3.5.4 Neural Network Radial Basis Function (NNRBF)
A Radial Basis Function system is an ANN that utilizes RBF as an
initiation capacity. The output of the system is a direct combination of radial
basis elements of the inputs and neuron parameters. Spiral premise capacity
systems have numerous utilizations, including capacity guess, time arrangement
expectation, grouping and framework control.
RBF system can be utilized to locate a set weight for a bend fitting
issue. The weights are in the higher dimensional space than the original data.
Learning is proportional to finding a surface in high dimensional space that gives
the best fit to training data. Hidden layers give an arrangement of capacities that
constitute a discretionary premise for input designs.
3.5.4.1 Radial basis function operation
ANN every neuron in a MLP (Multilayer Perception) holds the
weighted sum of its input values. That is, every input value is multiplied by a
coefficient and all the outcomes are summed up. A single MLP neuron is a plain
linear classifier, but difficult non-linear classifiers can be constructed by
introducing these neurons into a network. RBFN method is more spontaneous
than the MLP. To classify a fresh input, every neuron estimates the Euclidean
distance between the model and the input. Figure 3.13 shows the general
structure of NNRBF Algorithm. An input vector x is employed as input to all
radial basis functions with different properties. Each RBF neuron compares the
input vector to its model, and outputs a value in range [0, 1] which measures the
similarity. If the input is identical to the model, then RBF neuron’s output will
be 1. As the distance between the model and input increases, the output falls off
exponentially towards 0. RBF neuron’s output resembles a bell curve. The
output of the network is composed of a set of nodes. Every output node
calculates a score for the linked class. The score is calculated by taking a
weighted total of the activation values from each RBF neuron. By weighted total
64
we mean that an output node links a weight value with every RBF neuron, and
multiplies the neuron’s activation by this weight before adding it to the total
output. The data can be shown as a vector of bona fide numbers. The yield of the
framework is then a scalar limit of the information vector, φ
Where, N is the amount of neurons in the hidden layer, Ci is within
vector for neuron i, and ai is the weight of neuron "i" in the linear output neuron.
Limits that depend just on the division from a center vector is radially symmetric
about that vector. Therefore it is named as Radial Basis Function. In the central
frame, all inputs are connected with each disguised neuron. The standard is
regularly taken as the Euclidean partition and the extended reason limit is
conventionally taken as Gaussian.
The Gaussian basis functions are local to the center vector in the sense that
i.e. changing parameters of one neuron has just a little impact for
information values that are far from the focal point of that neuron.
Given certain mild conditions on the shape of the activation function,
RBF networks are universal approximations on a compact subset of Rn. This
means that an RBF network with enough hidden neurons can approximate any
continuous function with arbitrary precision.
The parameters ai, Ci, and βi are resolved in a way that improves the fit
amongst φ and the data.
65
Figure 3.11 General Structure of NNRBF Algorithm
An info vector "x" is utilized as a contribution to all outspread premise
works, each with various parameters. The output of the system is a direct
combination of the output from spiral premise capacities
The exact interpolation of a set of N data points in a multi-dimensional
space requires every one of the D dimensional input vectors xp i= {x p: i = 1,...
D} to be mapped onto the corresponding target output t p. The objective is to
discover a capacity f (x) such that
f (xp ) = t p p = 1,...,N (3.12)
Each RBF neuron stores a “prototype” vector, which is just one of the
vectors from the training set. Each RBF neuron compares the input vector to its
prototype and outputs a value between 0 and 1, which is a measure of similarity.
If the input is equal to the training, then the output of that RBF neuron
will be ‘1’ and the distance between the input and test data grows, the response
falls off exponentially towards 0. The shape of the RBF neuron’s response is a
bell curve, as illustrated in the network architecture diagram, Figure. 3.13.
66
The neuron’s response value is also called as “activation” value. The
prototype vector is also often called as the neuron’s “center”, since it is the value
at the center of the bell curve.
3.5.4.2 Output nodes
The output of the network consists of nodes equivalent to the number
of output. Each output node computes a sort of score for the associated category.
Typically, a classification decision is made by assigning the input to the category
with the highest score.
The score is computed by taking a weighted sum of the activation
values from every RBF neuron. By weighted sum, it is meant that an output node
associates a weight value with each of the RBF neurons and multiplies the
neuron’s activation by this weight before adding it to the total response. Because
each output node is computing the score for a different category, every output
node has its own set of weights. The output node will typically give a positive
weight to the RBF neurons that belong to its category and a negative weight to
the others.
3.5.4.3 Training of RBF neural networks
As specified some time recently, preparing of a RBF neural system
can be acquired with the choice of the ideal values for the accompanying
parameters:
1) (w) is the weights in between the hidden layer and the output layer.
2) (β) is defined as the parameters of the neuron in the output layer
3) (c) be the centre vector in the hidden layer
4) (α)be defined as the parameters of the hidden layer base function.
67
CHAPTER 4
COMPRESSION TECHNIQUES FOR MEDICAL IMAGES
USING FRACTAL, SPIHT AND DCT ALGORITHMS
4.1 INTRODUCTION
Recent radiology techniques offer crucial medical information for
radiologists to diagnose diseases and find out appropriate treatments. Such
medical information must be obtained through medical imaging processes.
Since, the medical images are in digital format, more cost-effective compression
technologies are required to reduce the mass volume of the digital image data
produced in the hospitals. Medical image compression is a challenging task as
the high frequency components contain details relevant for medical diagnosis. In
medical image compression, diagnosis is efficient only when compression
techniques conserve all the significant and important image information.
The idea of image compression technique is to minimize the
redundancy of the image data in order to store or transmit data in a competent
form. This results in the diminution of file size and allows more images to be
accumulated in a given amount of disk or memory space. Typically,
compression scheme can be categorized into two major categories: lossless and
lossy compressions. The lossy image compression is not very commonly used in
medical practice and diagnosis because even with a minor data loss, it is possible
that the physicians and radiologists fail to spot the critical information that could
be a crucial element for the diagnosis of a patient. In a lossless compression,
compressed data can be used to reconstruct an exact replica of the original
image; no information is lost due to the compression process. These necessities
are not satisfied with old techniques of compression like Fourier Transform,
Hadamard and Cosine Transform etc. due to high mean square error occurring
between original and compressed images. Fractal compression is a lossy
68
compression method for digital images, based on Fractals. The method is the
best suited for textures and natural images, relying on the fact that parts of an
image often resemble other parts of the same image. Fractal algorithms convert
these parts into mathematical data called Fractal codes, which are used to
recreate the encoded image. Out of which, Set Partitioning in Hierarchical Trees
(SPIHT) is the powerful wavelet-based image compression method. Thousands
of people, researchers and practitioners have tested and used SPIHT.
4.2 FRACTAL, SPIHT AND DCT METHODS
Fractal, SPIHT and DCT by using this algorithm, we can get better
PSNR values and the error is much reduced in this process.
4.2.1 Fractal
Fractal image coding depends on the basis of Partition Iterated
Function System (PIFS), in which a unique information image is divided into an
arrangement of non-covering sub-blocks, called range obstruct spread over the
entire image. Figure 4.1 illustrates the steps involved in Fractal image
compression.
Figure 4.1 Flow diagram of Fractal coding
69
4.2.2 Set Partitioning in Hierarchical Trees
The Set Partitioning in Hierarchical Trees (SPIHT) methodology is
not an easy extension of earliest ways for compression, and represents a
vigorous advance within the field. The SPIHT is a skilled image compression
routine utilizing the wavelet transform, where image coding exhausting the
wavelet change has intrigued unnecessary thought. The SPIHT has been
exceptionally effective. The tactic which deserves special attention and it
provides numerous advantages over the traditional methods. Figure 4.2
illustrates the basic operation involved in SPIHT method. Figure 4.3 illustrates
the formation of cells in SPIHT method.
Figure 4.2 Basic block diagram of SPIHT method
Figure 4.3 Formation of cells of SPIHT
Training Algorithm
Step 1: The medical image is being partitioned into small, non-overlapping, square blocks, typically called “parent blocks”.
70
Step 2: Each parent block is divided into 4 blocks named as “child blocks.”
Step 3: Now compare each child block against a subset of all possible overlapping blocks of parent block size.
Step 4: Reduce the size of the parent block to allow the comparison to work.
Step 5: Determine which larger block has the lowest difference, according to some measurement.
Step 6: Now to match the intensity levels between large block and the child block we will be calculating the grayscale transform. Where an affine transform is used (w*x = a*x + b) to match grayscale levels.
Step 7: Upper left corner child block is very similar to upper right parent block.
Step 8: Compute affine transform.
Step 9: Store location of parent block (or transform block), affine transform components, and related child block into a file.
Step 10: Repeat for each child block.
4.2.3 Discrete Cosine Transform
It registers 2-D DCT of 8-by-8 blocks in an input image, It is widely
used as a part of the JPEG image compression calculation. At first, the info
image is separated into 8-by-8(see Figure.4.4) or 16-by-16 blocks, and the 2-D
DCT is processed for every block. The DCT coefficients are then quantized,
coded and transmitted. The JPEG collector (or JPEG file reader) interprets the
quantized DCT coefficients, processes the converse two-dimensional DCT of
every block and after that assembles the blocks again into a solitary image. For
normal images, a lot of considerable DCT coefficients have values near zero.
These coefficients can be disposed of without truly influencing the nature of the
reproduced image.
71
v Figure 4.4 Two-dimensional DCT of 8-by-8 blocks in the image
Step 1: We need to convert an image into class double by reading an entire
image in the given workspace.
Step 2: The image is being computed by the 2-D of 8 by 8 blocks in the image
by mxm DCT matrix
Step 3: Discard all, but 10 of the 64 DCT coefficients in each block.
Step 4: The image is being reconstructed by using 2D inverse DCT of each
block.
Step 5: The reconstructed image is being easily recognizable that almost 85% of
DCT coefficients are discarded, while displaying the original image and
the reconstructed image side by side.
4.3 IMAGE QUALITY PARAMETER EVALUATION
The simulations are done utilizing MATLAB-Simulink and confirmed
utilizing scientific conditions. PET, CT and MRI images are favoured for an
intricate investigation. In Fractal, DCT and SPIHT calculation, it is important to
predefine a few parameters for contrasting their outcomes and past compression
algorithm. The various image parameters have been listed out from the
u
72
examination. These quality measures are being used for the assessments of
imaging frameworks. It demonstrates the productivity of the calculation and
shows the outcome.
4.3.1 Performance Parameter
The quality of the compressed image can be measured by many
parameters. The most commonly used parameters are Compression Ratio (CR),
Mean Square Error (MSE), Peak Signal to Noise Ratio Error (PSNR), Bits per
pixel (Bpp), Elapsed Time and Memory.
A. Compression Ratio (CR)
It is defined as the ratio of the size of the original image to the size of the compressed image.
(4.1)
where the n1 and n2 are defined as the input and output of the given input image.
B. Mean Square Error (MSE)
MSE is utilized to appraise the nature of compacted image. The lesser
the estimation of MSE is the higher the nature of packed image. It can be
communicated as MSE.
(4.2)
Where f (x,y) is the original image and g (x,y) is the reconstructed image and m,
n are the rows and columns of input image.
C. Peak Signal to Noise Ratio (PSNR)
( ) ( )MSENgfPSNR maxlog20),( = (4.3)
It is characterized as the measure of the data image to the MSE. If
PSNR is high, then the quality of reconstructed image is also increased.
73
D. Bits per pixel (Bpp)
The number of bits used to encode each pixel value is termed as Bpp.
Thus for the purpose of compression, Bpp should be minimum to reduce the
storage on the memory.
E. Elapsed Time
The compressed time gives the value of elapsed time in seconds from
the process of input image to the compressed image.
4.4 RESULTS AND COMPARISON
From Table 4.1, it is observed that, all the parameters obtained in the
evaluation test are in the acceptable limit. Hence it shows that after the
compression, the quality of the image is not degraded.
The simulation results for MR, PET and CT images are shown below
and the comparison tables are also included. The images with the highest CR and
PSNR are included below. The following figures (Figure 4.5 to 4.8) show the
comparison chart of CR, PSNR, Memory used and Execution time of various
medical images for methods like Fractal, Neural Network Back Propagation and
Radial Basis Function Neural Network.
The analysis of various techniques depicts that SPIHT provides better
CR values among all the techniques (DCT, Fractal). But DCT provides higher
PSNR values with less execution of time. And it also occupies less memory
space as compared to other techniques (SPIHT, Fractal). Therefore, we can
conclude that DCT compression method is best suited to MR, PET and CT
medical images.
74
Table 4.1 Shows performance comparison of 24 medical images which are obtained by using DCT, SPIHT
and Fractal algorithm
CR PSNR MEMORY EXECUTION TIME Images DCT SPIHT Fractal DCT SPIHT Fractal DCT SPIHT Fractal DCT SPIHT Fractal CT Image 1 1.4088 0.4410 6.6507 88.8087 38.1902 38.2519 21.40 66.40 26.80 0.7572 42.9628 11.7500 CT Image 2 1.3442 0.3542 6.6622 88.3888 37.1780 38.7416 23.20 60.70 28.40 0.7645 45.9385 11.7969 CT Image 3 1.4679 0.5264 13.2030 85.8091 36.8513 35.8632 31.60 75.50 43.20 0.7522 100.4340 16.3281 CT Image 4 1.3443 0.5526 5.0707 92.4192 36.7236 38.9335 21.10 60.40 25.00 0.8272 35.4521 10.8594 MR Image 5 1.4402 0.9332 6.5164 90.5865 36.3537 43.6002 18.00 66.20 21.90 0.9514 60.0308 12.0000 MR Image 6 1.6127 1.3050 10.9643 91.8776 36.3546 41.7960 18.90 65.90 26.40 0.8773 46.2459 14.8125 MR Image 7 1.3586 0.7977 4.4960 92.2370 35.8857 39.0490 22.60 61.10 23.00 0.7620 32.1609 10.5625 MR Image 8 1.5927 0.8920 7.5873 97.7195 35.8076 42.3063 19.20 65.60 25.10 1.5385 69.1809 12.3906 MR Image 9 1.5900 0.5976 5.1806 101.4771 35.3169 43.6769 15.70 52.60 20.50 0.7825 20.8594 10.9688 MR Image 10 1.2792 0.2437 6.9771 90.5101 35.9551 43.3527 20.50 66.40 23.70 1.1055 39.8822 12.1094 MR Image 11 1.3006 0.1853 6.5397 90.6810 35.8193 43.0880 19.60 73.00 23.00 0.8540 35.8102 11.9219 MR Image 12 1.4355 0.4519 3.4217 99.6740 35.4337 40.7357 17.50 61.50 22.30 0.7832 39.5033 9.8438 MR Image 13 1.4994 0.4558 1.8435 114.2706 36.6741 43.4892 12.40 46.10 17.80 0.7919 39.6286 9.0625 MR Image 14 1.3451 0.1391 2.3289 97.6726 36.5000 41.3458 14.40 36.20 15.90 1.4770 48.2454 9.0938 MR Image 15 1.6541 0.6722 3.0478 98.4225 35.9084 39.6035 17.20 56.90 20.50 0.7612 48.3930 9.4844 MR Image 16 1.5732 1.0766 3.4444 95.1357 35.8668 39.9027 18.20 53.90 22.70 0.8803 45.3308 10.0625 MR Image 17 1.4197 0.1334 4.8070 96.8298 35.7687 39.0449 23.40 62.40 29.90 1.0694 45.0365 10.7031 MR Image 18 1.6541 0.6722 3.0478 98.4225 35.9084 3.0478 17.20 56.90 20.50 0.9104 38.9254 9.5000 MR Image 19 1.7575 0.1213 3.4270 90.0821 36.3353 3.4270 26.80 40.70 28.20 1.0884 51.7431 10.1094 MR Image 20 1.0884 0.4835 4.1237 91.1779 36.2339 34.1952 27.40 51.90 32.70 1.2724 45.5717 10.5156 MR Image 21 1.2792 0.2437 6.2566 90.5101 35.9551 43.3019 20.50 66.40 20.60 0.7581 46.0090 11.5313 PET Image 22 1.5891 0.5745 5.5154 97.2405 36.1077 43.3578 14.80 59.20 19.20 0.8768 42.3300 11.2656 PET Image 23 1.4908 0.4473 1.4721 112.8808 35.1495 43.9157 9.97 41.10 14.30 0.7791 28.1462 8.5625 PET Image 24 1.5791 0.3454 1.1946 112.1955 36.6452 39.3448 13.20 34.50 18.80 1.0188 41.1812 8.8125
75
Figure 4.5 Compression Ratio expressed in percentage
Figure 4.5 shows the compression ratio of three different algorithms
namely DCT, SPIHT and Fractal. It is clearly evident that SPIHT provides better
CR values.
Figure 4.6 Shows the PSNR for three different algorithms DCT, SPIHT and Fractal.
76
Figure 4.7 Memory expressed in kilo byte
Figure 4.7 shows the memory usage of three different algorithms
namely, DCT, SPIHT and Fractal. It is noticed that DCT uses lesser memory for
image compression.
Figure 4.8 Execution Time expressed in Seconds
Figure 4.8 Shows the Execution time of three different algorithms
namely, DCT, SPIHT and Fractal. Here, DCT produces compression results
within minimal time duration.
77
Figure 4.9 (A) Results obtained for various medical images
(a).Input Image (b).DCT (c).SPIHT and (d).Fractal algorithms
78
Figure 4.9 (B) Results obtained for various medical images
(a).Input Image (b).DCT (c).SPIHT and (d).Fractal algorithms
79
Figure 4.9 (C) Results obtained for various medical images
(a).Input Image (b).DCT (c).SPIHT and (d).Fractal algorithms
80
Figure 4.9 (D) Results obtained for various medical images
(a).Input Image (b).DCT (c).SPIHT and (d).Fractal algorithms
81
Figure 4.9 (E) Results obtained for various medical images
(a).Input Image (b).DCT (c).SPIHT and (d).Fractal algorithms
82
Figure 4.9 (F) Results obtained for various medical images
(a).Input Image (b).DCT (c).SPIHT and (d).Fractal algorithm
83
The image quality parameters can be represented in graphical form to
study the character of each parameter with respect to other parameters for better
understanding. This graphical representation of test images with respect to CR,
PSNR, Execution time and memory is shown in the Figures 4.5 to 4.8. The
graphical representation of test input images with respect to MRI, CT and PET
and Elapsed time in seconds is shown in the Figures 4.9A to 4.9F.
4.5 CONCLUSION
In this chapter, to enhance the performance of three completely
different approaches which are compared for medical images like Fractal,
Discrete Cosine Transform and Set Partitioning in Hierarchical Trees, these
approaches are tested against completely different medical images like human
MRI image, PET and CT images, using specific image quality parameters like
Compression Ratio, Bits per pixel, Peak Signal Noise Ratio and Mean Square
Error. The results clearly show that SPIHT methodology has higher Compression
Ratio (CR) and PSNR value with less BPP and MSE for PET and MR brain
images. Future work lies in developing a neural network framework, so as to
realize the higher compression results.
84
CHAPTER 5
EFFICIENT IMAGE COMPRESSION TECHNIQUES FOR
MULTIMODAL MEDICAL IMAGES USING RADIAL BASIS
FUNCTION NEURAL NETWORK APPROACH
5.1 INTRODUCTION
Recent radiology techniques offer vital medical information for
radiologists to diagnose diseases and find out appropriate treatments through
image processing techniques. Image compression is one of the image processing
techniques, which is used to reduce the number of bits that are required to
represent an image. Since, the medical images are in digital format, more cost-
effective compression techniques are required to reduce the mass volume of
digital image data produced in the hospitals. Medical image compression is a
challenging task as the high frequency components may contain important
information for medical diagnosis. In medical image compression applications,
diagnosis is efficient only when compression techniques preserve all the
significant and important image information. The idea of image compression
technique is to minimize the redundancy of the image data, in order to store or
transmit the data in a competent form. This results in size reduction and allows
more images to be accumulated in a given amount of disk or memory space.
Typically, compression scheme can be categorized into two major categories:
lossless and lossy compressions. The lossy image compression is not commonly
used in medical diagnosis, because it fails to interpret the critical information
for radiologists to diagnose the patient. In a lossless compression, no useful
information is lost due to the compression process. Huffman comes under lossless
and Fractal comes under lossy image compression. An image with size 512×512
is given as an input to all these above methods.
85
5.2 ALGORITHMS FOR IMAGE COMPRESSION
Different compression methods such as Fractal, Neural Network Back
Propagation (NNBP) and Radial Basis Function Neural Network (RBFNN) are
applied to various medical images such as MR and CT images. Experimental
results show that the NNRBF technique achieves a higher Compression Ratio
(CR), Bits per pixel (Bpp) and Peak Signal to Noise Ratio (PSNR), with less
Mean Square Error (MSE) on CT and MR images when compared to Fractal and
Neural Network Back Propagation techniques.
5.2.1 Fractal Algorithm
Fractal encoding is to a great extent used to change over bitmap images
to fractal codes. This encoding procedure is to a great degree computationally
escalated. Millions of cycles is being required to discover the fractal designs in a
image. Contingent on the determination and substance of the info bitmap
information and yield quality, compression time, and record size parameters
chose, packing a solitary image can take anyplace from a few moments to a
couple of hours on even a quick PC. All the decoding process needed to do is to
interpret the Fractal codes and translate them into a bitmap image. Two huge
advantages are instantly acknowledged by changing over routine bitmap images
to Fractal information. The first is the ability to scale any Fractal image up or
down in size without the presentation of image curios or an adversity in
inconspicuous component that happens in bitmap images. The strategy of "Fractal
zooming", which is free of the determination of the main bitmap image and the
zooming is compelled just by the measure of available memory in the PC. The
second point of interest is the way that the measure of the physical data is used to
store the fractal codes, which is much smaller than the extent of the primary
bitmap data. Really, it is not exceptional for the fractal images to be more than
86
100 times humbler than their bitmap sources. This part of Fractal development,
called Fractal compression has advanced the best eagerness inside the PC imaging
industry. The process of matching Fractals does not involve looking for exact
matches, but instead looking for the "best fit", it matches based on the
compression parameters (encoding time, image quality, and size of output). But
the encoding process can be controlled to the point where the image is "visually
lossless." That is, the occurrence of the data loss is undetectable. Fractal
compression contrasts from other lossy compression techniques, for example,
JPEG, in various ways. JPEG fulfills the compression by discarding the image
data that are not required for the human eye to see the image. The resulting data
are then further compacted by using a lossless procedure for compression. To
finish more significant compression extents, more image data must be discarded,
achieving a low quality image with a pixelized (blocky) appearance. Fractal
images are not considered as an aid of pixels, nor the encoding weight to the
visual characteristics of the human eye. Or maybe, bitmap data is discarded when
it is required to make a best-fit Fractal outline. More noticeable compression
extents are expert using the more vital computationally genuine changes that may
degrade the image, yet the bending appears significantly more normal in view of
the Fractal portions.
5.2.2 Neural Network Back Propagation (NNBP)
The back-propagation learning algorithm is one of the most
significant improvements in neural networks. This learning algorithm can be
mainly applied to feed-forward networks that can be consists of dispensation
elements with uninterrupted differentiable activation functions. In BPNN,
training input-output pair is given as an input for training, this algorithm make
available a procedure for varying the weights in a BPNN to categorize the given
87
input patterns appropriately. The very vital concept for this weight update
algorithm is gradient-descent method. The back-propagation algorithm is
completely different from other networks in respect to the process by which the
weights are calculated during the learning period of the network. There will be
three different layers, one input layer, one output layer and one hidden layer, are
assigned. Both of input layer in BPNN and output layer are fully connected to
hidden layer. Compression is obtained by designing the value of the number of
neurons in the both input layer and output layers neuron less than the hidden
layer neuron. Figure 5.1 shows General structure of Neural Network Back
Propagation Algorithm.
Figure 5.1 General structure of Neural Network Back Propagation
Algorithm.
5.2.3 Neural Network Radial Basis Function for Image Compression
Radial basis function neural networks (RBFNN) are feed-forward
networks trained using a supervised training algorithm. They are normally put
together with a single hidden layer of units whose output function is selected
from a class of functions called basis functions. The structure of an RBF
networks in its most basic form involves three entirely different layers as
shown in Fig.4. The input layer is made up of source nodes (sensory units)
whose number is equal to the dimension N of the input vector. The second layer
88
is the hidden layer which is composed of nonlinear units that are connected
directly to all of the nodes in the input layer. Each hidden unit takes its input
from all the nodes at the components at the input layer. The hidden unit
contains a basis function, which has the parameters center and width. Figure 5.2
shows the General structure of Radial Basis Function Neural Network.
Figure 5.2 General structure of Radial Basis Function Neural Network.
5.3 PERFORMANCE PARAMETERS
There are several parameters that can be used to compare the
various image compression techniques. The efficiency of the compression
algorithm is measured in terms of performance measuring parameters such as
Compression Ratio (CR), Peak Signal Noise Ratio (PSNR), Bits per pixel (Bpp),
Mean Square Error (MSE) and Testing and Training Time.
5.4 RESULTS AND COMPARISON
The simulation results for various medical images and the comparison
tables are also included. The images with the highest CR and PSNR are included
below. The following figures (Figure 5.3 to 5.6) show the original and
reconstructed image of MR and CT images by using Fractal, NNBP and NNRBF
methods.
89
Table 5.1 Shows performance comparison of 24 medical images which are obtained by using Fractal, NNRBF
and NNBP algorithm
CR PSNR MEMORY EXECUTION TIME
Images Fractal NNRBF NNBP Fractal NNRBF NNBP Fractal
NNRBF NNBP Fractal NNRBF NNBP
CT Image 1 6.6507 1.0537 1.0495 38.2519 39.1906 69.6090 26.80 28.60 28.70 11.7500 17.4231 850.4214 CT Image 2 6.6622 0.8386 1.0356 38.7416 29.9663 42.3719 28.40 37.20 30.10 11.7969 16.4425 278.7227 CT Image 3 13.2030 1.0632 1.0330 35.8632 30.6855 62.5396 43.20 43.60 44.90 16.3281 18.6835 542.2186 CT Image 4 5.0707 1.1561 1.0328 38.9335 19.5331 40.2157 25.00 24.50 27.40 10.8594 18.5866 935.6339 MR Image 5 6.5164 0.9678 1.0564 43.6002 32.7573 62.0237 21.90 26.70 24.50 12.0000 18.5779 387.0713 MR Image 6 10.9643 1.1854 1.0423 41.7960 22.9515 46.7436 26.40 25.70 29.20 14.8125 18.2230 667.3835 MR Image 7 4.4960 1.2092 1.0398 39.0490 22.3658 44.7684 23.00 25.40 29.50 10.5625 18.2627 658.0410 MR Image 8 7.5873 1.117 1.0442 42.3063 29.9071 48.5713 25.10 27.40 29.30 12.3906 18.7018 1164.2615 MR Image 9 5.1806 1.1416 1.0208 43.6769 24.6009 45.7868 20.50 21.90 24.50 10.9688 18.3615 836.9937 MR Image 10 6.9771 1.0116 1.0591 43.3527 44.3570 73.0390 23.70 25.90 24.80 12.1094 18.6340 751.2015 MR Image 11 6.5397 0.9904 1.0605 43.0880 44.2258 68.0856 23.00 25.80 24.10 11.9219 18.7667 294.9351 MR Image 12 3.4217 0.956 1.0536 40.7357 29.3755 54.1458 22.30 26.30 23.90 9.8438 17.8159 415.8369 MR Image 13 1.8435 1.0728 1.0837 43.4892 50.9330 66.8312 17.80 17.30 17.10 9.0625 17.9883 497.0650 MR Image 14 2.3289 1.1096 1.0728 41.3458 35.7803 47.6965 15.90 71.50 18.10 9.0938 17.2854 313.2422 MR Image 15 3.0478 1.0365 1.0514 39.6035 46.2050 72.3725 20.50 27.40 27.00 9.4844 17.7939 285.6100 MR Image 16 3.4444 1.0499 1.0513 39.9027 32.9302 57.2598 22.70 27.30 27.30 10.0625 18.7660 610.1797 MR Image 17 4.8070 0.9978 1.0426 39.0449 37.5945 84.0402 29.90 33.30 31.90 10.7031 26.7360 843.5427 MR Image 18 3.0478 1.0365 1.0514 3.0478 1.0365 72.3725 20.50 27.40 27.00 9.5000 27.1599 285.4747 MR Image 19 3.4270 1.1379 0.9938 3.4270 1.1379 36.2982 28.20 41.40 47.10 10.1094 17.9907 654.8289 MR Image 20 4.1237 1.003 1.0398 34.1952 38.2271 69.1297 32.70 40.60 39.20 10.5156 18.1835 504.8591 MR Image 21 6.2566 1.0636 1.0591 43.3019 32.7369 73.0390 20.60 22.90 24.80 11.5313 18.5056 744.2455 PET Image 22 5.5154 1.0352 1.0530 43.3578 33.4930 50.2760 19.20 22.80 22.40 11.2656 18.8691 652.3676 PET Image 23 1.4721 0.7923 0.7662 43.9157 24.4163 42.7221 14.30 18.70 19.40 8.5625 16.5292 452.7775 PET Image 24 1.1946 1.0696 1.0709 39.3448 52.1345 84.0081 18.80 19.60 19.50 8.8125 16.7909 378.8078
90
Fractal, NNBP and NNRBF, of which NNRBF produces better compression ratio.
Figure 5.3 Compression Ratio expressed in percentage
Figure 5.4 Shows the Peak Signal to Noise Ratio of the algorithms Fractal,
NNBP and NNRBF. It is clearly identifiable that NNRBF is capable of producing high
PSNR values.
Figure 5.4 Peak Signal to Noise Ratio expressed in decibels
91
Figure 5.5 Memory expressed in kilo byte
Figure 5.5 shows the memory usage of three different algorithms namely, Fractal, NNRBF and NNBP. It is noticed that Fractal uses lesser memory for image compression.
Figure 5.6 Execution Time expressed in Seconds
92
Figure 5.6 Shows the Execution time of three different algorithms namely, Fractal, NNRBF and NNBP. Here, Fractal produces compression results within minimal time duration.
Figure 5.7(A) Results obtained for various medical images
(a).Input Images (b).Fractal (c).Neural Network Back Propagation (NNBP)
and (d).Radial Basis Function Neural Network algorithms
93
Figure 5.7(B) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). Neural Network Back Propagation (NNBP) and (d). Radial Basis Function Neural Network algorithms.
94
Figure 5.7 (C) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). Neural Network Back Propagation(NNBP) and (d). Radial Basis Function Neural Network algorithms.
95
Figure 5.7 (D) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). Neural Network Back Propagation(NNBP) and (d). Radial Basis Function Neural Network algorithms.
96
Figure 5.7(E) Results obtained for various medical images (a). Input Images (b). Fractal, (c). Neural Network Back Propagation(NNBP)
and (d). Radial Basis Function Neural Network algorithms.
97
Figure 5.7(F) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). Neural Network Back Propagation(NNBP) and (d). Radial Basis Function Neural Network algorithms.
The above figures (Figure 5.7A to 5.7F) show the original and
reconstructed image of MR and CT images by using Fractal, Neural Network
Back Propagation methods and Radial Basis Function Neural Network.
98
5.5 CONCLUSION
In this chapter, four different approaches such as Huffman, Fractal,
Neural Network Back Propagation and Radial Basis Function Neural Network
algorithms are compared for medical image compression. These approaches are
tested with different medical images such as MR and CT images. In order to
identify a better compression method, specific image quality parameters like
compression ratio, bits per pixel, peak signal to noise ratio, mean square error and
execution time have been calculated. The results clearly show that the Radial
Basis Function Neural Network method has low Compression Ratio (CR) and
high PSNR value with less BPP and MSE for MR and CT images. Thus, Radial
Basis Function is found to be efficient when compared to Huffman, Fractal and
Neural Network Back Propagation algorithms. Future work is to develop a hybrid
approach by combining two or more algorithms, in order to achieve better
compression results. Further, it is essential to prepare a hybrid approach by
mixing two or more soft computing techniques to achieve better compression
results.
99
CHAPTER 6
A HYBRID DISCRETE WAVELET TRANSFORM WITH
NEURAL NETWORK BACK PROPAGATION APPROACH
FOR EFFICIENT MEDICAL IMAGE COMPRESSION
6.1 INTRODUCTION
An image as a rule comprises of colossal measure of information and
requires an expansive number of space in the memory. Compacted image
possesses less number of spaces in memory and it requires less time for
transmission. The principle capacity of image compression is to improve the
information that may debase the image, contingent upon the compression ratio
(CR). CR is one of the best parameters to get great nature of a compressed image.
By evaluating this CR, the quality of an image may be predicted. Usually, the
input image will be in the form of analog images. Thus these analog images can
be sampled and quantized to get the digital images. By using these digital images,
the compression algorithm can be made available. For example, in order to
transfer an image of size 512×512, it will take around a few minutes to reach the
receiver. Thus by utilizing the compression algorithm, many of the medical
images can be sent simultaneously in less time duration. The importance of this
image compression is to decrease the cost for storage space and communication.
6.2 ALGORITHMS USED
6.2.1 Back Propagation Neural Networks Algorithm
Neural Network is nowadays an important emerging tool that can be
very applicable to image processing techniques. There will be many training pairs
in BPNN, the most important and useful training pair is input-output pair. The
BPNN algorithm will be able to give the procedures for varying the weights after
100
giving the input; this input to output pair will categorize the given input patterns.
Gradient-descent method is one of the apt methods for weight updation in BPNN.
In BPNN, the weight can be computed amid the learning time of the system, in
this way it’s different from all other algorithms. Back propagation algorithm
involves three different types of layers namely input, output, and hidden layer.
The value of neurons in BPNN can be evaluated by both input and output layer.
Figure 6.1 General Structure of BPNN
The first step of image compression in BPNN is to decompose the input
images in pixels; this can be done by the algorithm of BPNN. These pixels, which
are encoded in previous step can be given as an input to the network. Now, this
image is transmitted and recovered in the receiver side. There are three important
layers in BPNN, which are named as input layer, hidden layer, output layer,
where this hidden layer should be more than one. The next process is to encode
spatial coordinates of the pixel. Entropy encoding, which is a way of lossless
compression will convert the image from two to one dimensional value and then
the image is compressed. After getting the compressed image, the error can be
calculated in all the three layers.
101
6.2.2 Discrete Wavelet Transform
There may be different types of image processing techniques in which
this DWT can be of successive high pass and low pass filter, where the images
can be divided into pixels. Decomposition in wavelet transform consist of two
parts, i) approximation of an image (scaling function) and ii) detailed part of an
image (wavelet function).
Nowadays, in the developing world, many researches are done on
wavelet representation and transforms. The area which is having no noise can be
represented as plain areas in an image. These areas will have very high degree of
redundancy. This chapter discusses the hybrid combination of DWT and NNBP.
In order to get better compressed image without degrading the quality of image,
there should be low CR and PSNR. Thus in this chapter, we hybrid these two
algorithms viz., DWT and BP and this gives better CR and PSNR.
Figure 6.2 Block diagram of Hybrid DWT-BP Algorithm
Here the input images of size 512×512 are given to get compressed.
First, these images can be given to DWT algorithm which undergoes image
compression process and the image which gets out from DWT algorithm be the
input given to BPNN for further compression. The Image obtains as the output
from the BPNN be the compressed image, which has better CR and PSNR. We
can predict that our proposed method will be having efficient CR and PSNR.
Thus the comparison chart for both existing method and proposed method is
given in Figure 6.3.
Input Images
Hybrid DWT-BP DWT Algorithm BPNN Compressed
Images
102
Figure 6.3 Comparison chart of proposed work and existing method
6.3 PERFORMANCE PARAMETERS
Many parameters are used to measure the quality of any compressed
image. Commonly used parameters are Mean Square error (MSE), Peak signal to
Noise Ratio (PSNR), Compression Ratio (CR) and Bits per pixel (BPP).
6.4 RESULTS AND DISCUSSION
From Table 6.1, it is observed that, all the parameters obtained in the
evaluation test are in the acceptable limit. Hence it shows that after the
compression, the quality of the image is not degraded. The simulation results for
MR, PET and CT images are shown below and the comparison tables are also
included. The images with the highest CR and PSNR are included below. The
following figures (Figure 6.4 to 6.7) show the comparison chart of CR, PSNR,
Memory used and Execution time of various medical images for methods like
DWT, NNBP and Hybrid DWT with NNBP.
103
The analysis of various techniques depicts that Hybrid DWT with
NNBP provides better CR values and it occupies less memory space as compared
to other techniques (DWT, NNBP). DWT provides higher PSNR values with less
execution of time. Therefore, we can conclude that Hybrid DWT with NNBP
compression method is best suited to MR, PET and CT medical images in terms
of CR and memory space.
104
Table 6.1 Shows performance comparison of 24 medical images which are obtained by using DWT, BPNN
and hybrid DWT-BP algorithm.
CR PSNR MEMORY EXECUTION TIME
Images DWT NNBP Hybrid DWT-BP
DWT NNBP Hybrid DWT-BP DWT NNBP Hybrid
DWT-BP DWT NNBP Hybrid DWT-BP
CT Image 1 1.0002 1.0495 1.0005 71.7728 69.6090 61.3201 28.70 28.70 24.1 1.0867 850.4214 848.0286 CT Image 2 1.0011 1.0356 1.0121 71.3010 42.3719 39.1269 29.80 30.10 17.2 1.0818 278.7227 279.6484 CT Image 3 1.0015 1.0330 1.0016 55.7033 62.5396 47.2912 44.80 44.90 30.3 2.3126 542.2186 546.848 CT Image 4 1.0020 1.0328 0.568 67.3440 40.2157 29.4004 26.90 27.40 26.9 1.6483 935.6339 938.7183 MR Image 5 1.0005 1.0564 0.9254 67.7753 62.0237 38.5935 24.50 24.50 51.4 1.1104 387.0713 354.3677 MR Image 6 1.0048 1.0423 0.8844 63.8786 46.7436 31.9976 28.90 29.20 70.5 1.1089 667.3835 668.1684 MR Image 7 1.0229 1.0398 0.0969 62.7552 44.7684 29.4417 28.60 29.50 49.6 1.2349 658.0410 656.5766 MR Image 8 1.0265 1.0442 0.5601 60.1968 48.5713 23.6815 28.40 29.30 44.9 1.0894 1164.2615 874.3936 MR Image 9 1.0144 1.0208 0.0999 64.2066 45.7868 25.814 23.30 24.50 55.5 1.0646 836.9937 798.0145 MR Image 10 1.0001 1.0591 1.005 67.8580 73.0390 47.1406 24.80 24.80 11.6 1.1229 751.2015 786.3682 MR Image 11 0.9999 1.0605 0.5603 68.4217 68.0856 43.8376 24.10 24.10 10.3 1.1197 294.9351 287.1838 MR Image 12 1.0073 1.0536 0.8914 67.9670 54.1458 38.7958 23.70 23.90 22.1 1.1865 415.8369 425.006 MR Image 13 1.0013 1.0837 0.9743 72.8183 66.8312 38.0442 17.10 17.10 17.8 1.0898 497.0650 504.9017 MR Image 14 1.0006 1.0728 0.0518 72.5798 47.6965 24.6016 18.00 18.10 3.32 1.0197 313.2422 321.9819 MR Image 15 1.0001 1.0514 1.0052 64.5986 72.3725 43.2033 27.00 27.00 27 1.0705 285.6100 293.2718 MR Image 16 1.0023 1.0513 1.0115 65.2455 57.2598 40.3127 27.20 27.30 45.2 1.0782 610.1797 633.1677 MR Image 17 1.0001 1.0426 0.9986 61.9274 84.0402 64.8598 31.90 31.90 5.78 1.1614 843.5427 858.127 MR Image 18 1.0001 1.0514 1.0052 64.5986 72.3725 43.2033 27.00 27.00 27 1.0800 285.4747 293.2718 MR Image 19 1.0152 0.9938 0.0471 54.2999 36.2982 23.7107 45.00 47.10 70.5 1.0605 654.8289 663.0196 MR Image 20 1.0015 1.0398 1.002 60.5071 69.1297 49.3025 39.10 39.20 19.2 1.2132 504.8591 551.1406 MR Image 21 1.0001 1.0591 1.005 67.8580 73.0390 47.1406 24.80 24.80 11.6 1.0645 744.2455 758.223 PET Image 22 1.0069 1.0530 0.4557 66.0979 50.2760 27.168 21.90 22.40 23.3 1.1157 652.3676 653.9199 PET Image 23 1.0006 0.7662 0.2325 78.9087 42.7221 28.3994 13.40 19.40 15 1.0681 452.7775 463.315 PET Image 24 1.0048 1.0709 1.0007 70.6387 84.0081 48.427 19.40 19.50 9.54 1.1340 378.8078 388.1468
105
Figure 6.4 Comparison of Compression Ratio for different Input images
Figure 6.4 shows the comparison of CR for different input images. CR for
DWT algorithm is lower.
Figure 6.5 Comparison of PSNR Values for different Input Image
Figure 6.5 shows that PSNR for DWT algorithm is very low compared to
other two techniques. Therefore, it can be concluded that hybrid DWT-BP
algorithm gives better PSNR values.
106
Figure 6.6 Memory expressed in kilo byte
Figure 6.6 shows the memory usage of three different algorithms
namely, DWT, NNBP and Hybrid DWT-BP. It is noticed that DWT uses lesser
memory for image compression.
Figure 6.7 Execution Time expressed in Seconds
Figure 6.7 Shows the Execution time of three different algorithms
namely, DWT, NNBP and Hybrid DWT-BP. Here, DWT produces compression
results within minimal time duration.
107
Figure 6.8 (A) Results obtained for various medical images
(a). Input Images (b). DCT, (c). SPIHT and (d). Fractal algorithms
108
Figure6.8 (B) Results obtained for various medical images
(a). Input Images (b). DCT, (c). SPIHT and (d). Fractal algorithms
109
Figure 6.8 (C) Results obtained for various medical images
(a). Input Images (b). DCT, (c). SPIHT and (d). Fractal algorithms
110
Figure 6.8 (D) Results obtained for various medical images
(a). Input Images (b). DCT, (c). SPIHT and (d). Fractal algorithms
111
Figure 6.8 (E) Results obtained for various medical images
(a). Input Images (b). DCT, (c). SPIHT and (d). Fractal algorithms
112
Figure 6.8(F) Results obtained for various medical images
(a). Input Images (b). DCT, (c). SPIHT and (d). Fractal algorithms
113
Figures 6.8A to 6.8F show the compressed images obtained using
DWT, BPNN and Hybrid DWT-BP along with respective original images (by
simulation). All the input images are in the size of 512×512.
6.5 CONCLUSION
Image compression based on DWT, NNBP and Hybrid DWT-NNBP is
discussed. The input image of size 512×512 is given, where the compressed
image is obtained by these above algorithms. Various parameters are calculated to
know the quality of the compressed image. By viewing the comparison charts
which are given in the Figures 6.4 to 6.7, it can be concluded that (for both CR
and PSNR) among the three algorithms, Hybrid DWT-NNBP gives efficient
results.
114
CHAPTER 7
A HYBRID APPROACH USING FRACTAL AND NEURAL
NETWORK RADIAL BASIS FUNCTION FOR EFFICIENT
COMPRESSION OF MULTI MODAL MEDICAL IMAGES
7.1 INTRODUCTION
Compression relates to the process of reducing the file size by
rearranging the information in the file. Compressing images is different from
zipping files. Image compression changes the system and the content of the
information within a file. Loss of the data may or may not be noticeable. The
quantity of image compression can be influenced by the type of images. The
greater compression ratio can be accomplished in the portions of the image.
Where compression is an important technique which is essential to store and
transmit an image over the long distance. An untreated image acquires more
memory to be compressed. the techniques of de compressed image to be
compressed because of the requirement of high quality image in building a video
for this loss less compression is a compression methods that permits to
reconstructed the compressed data from the original data in common it generally
two things at first it generates the statistical framework for the input information
second it matches the input information to bit sequences.
Fractal image compression is based on the fractals of various images.
The merits of converting the images to fractal data are 1) Reduced memory space
requirement of the compressed image. 2) Quantification of parameters like
Compression Ratio (CR), Peak Signal to Noise Ratio (PSNR), Bits per pixel
(Bpp) and others. The number of RBFs used to encode a sub image is too lower
than the number of data points that result in reduction of data size. RBF networks
115
configure a neural network architecture that is extensively used for modeling and
controlling nonlinear systems.
In the existing fractal algorithm, CR is good but Mean Square Error
(MSE) is high and PSNR value is low. But the proposed Radial Basis Function
Neural Network (RBFNN) takes small convergence time during the training
period and produces good PSNR values. New hybrid approach combining the
working principles of Fractal and RBFNN is implemented here. Comparisons of
existing algorithms are also exhibited.
7.2 METHODOLOGIES
7.2.1 Fractal Algorithm
The term Fractal is being first used by Benoit Mandelbrot in the year
1975. It uses the original image and makes three exact copies. Fractal encoding is
generally used to encode the bit map images by using mathematical techniques. A
set of numerical information expresses the Fractal properties of image.
7.2.2 Neural Network Radial Basis for Image Compression
The idea of Radial Basis Function (RBF) Networks derives from the
theory of function approximation. We have already seen how Multi-Layer Perceptron
(MLP) networks with a hidden layer of sigmoidal units can learn to approximate
functions. RBF Networks take a slightly different approach. It has a single hidden
layer. The basic neuron model as well as the function of the hidden layer is different
from that of the output layer. The hidden layer is nonlinear but the output layer is a
linear activated function of the hidden unit which computes the euclidean distance
between the input vector and the center of that unit.
116
The main features are, It has two-layer feed-forward networks and hidden nodes are
implemented a set of radial basis functions. The output nodes are implemented using
linear summation functions as in MLP. The network training is divided into two stages:
first the weights from the input to hidden layer are determined and then the weights from
the hidden to output layer. The networks are very good at interpolation.
Radial basis networks can be used to approximate functions. To add
neurons to the hidden layer of a radial basis network until it meets the specified mean
squared error goal. The larger spread is, the smoother the function approximation. Too
large a spread means a lot of neurons are required to fit a fast-changing function. Too
small a spread means many neurons are required to fit a smooth function, and the
network might not generalize well. Call new radial basis function with different spreads
to find the best value for a given problem.
Figure 7.1 General structure of NNRBF
117
7.3 IMPLEMENTATION OF HYBRID TECHNIQUES
7.3.1 Hybrid image compression
Figure 7.2 Hybrid image compression using FNNRBF method
Figure 7.2 Shows the proposed Hybrid image compression using
FNNRBF. The NN-RBF algorithm is used to improve the transformation process,
which increases the edge threshold. Simultaneously, the fractal coding and NN-
RBF algorithm are combined to obtain hybrid FNNRBF coding, in order to get
better quality in image compression.
7.4 IMAGE QUALITY PARAMETER EVALUATION
Generally, for evaluating the parameters of image quality we use
MATLAB-simulating and it is verified using the mathematical equations. Two
dimensional multi modal medical images are preferred for an elaborate analysis.
Different image quality parameters are being computed from this analysis. The
image quality measures are the figures of authenticity used for the evaluation of
imaging systems. It exhibits the capability of the computation and demonstrates
the outcome. The nature of the compacted image can be measured by numerous
118
parameters. The most commonly used parameters are Compression Ratio (CR),
Mean Square Error (MSE), Peak Signal to Noise Ratio Error (PSNR), Bits per
pixel (Bpp), Memory and Execution Time. In general, CR value is low and the
PSNR value is high. From these two parameters the better will be the greater
compressor.
7.5 SIMULATION RESULTS AND ANALYSIS
The image quality parameters can be represented in graphical form to
study the character of each parameter with respect to other parameters for better
understanding. The following figures (Figure 7.4 to 7.7) show the comparison
chart of CR, PSNR, Memory used and Execution time of various medical images
for methods like Fractal, NNRBF and Hybrid FNNRBF.
CR and PSNR are better with Hybrid FNNRBF. It is identifiable that
the compressed image size is much less in Hybrid FNNRBF. It is clearly stated
that the execution of time is greatly reduced by using Hybrid FNNRBF. Fractal
provides higher PSNR values with less execution time. Therefore, we can
conclude that Hybrid FNNRBF compression method is the best suited to MR,
PET and CT medical images in terms of CR and memory space. Table 7.1 Shows
CR obtained using NNRBF, Fractal and Hybrid FNNRBF.
119
Table 7.1 Performance comparison of 24 medical images which are obtained by using NNRBF, Fractal and Hybrid FNNRBF
CR PSNR MEMORY EXECUTION TIME
Images Fractal NNRBF Hybrid
Fractal & NNRBF
Fractal NNRBF Hybrid
Fractal & NNRBF
Fractal NNRBF
Hybrid Fractal & NNRBF
Fractal NNRBF Hybrid
Fractal & NNRBF
CT Image 1 6.6507 1.0537 1.0424 38.2519 39.1906 34.1945 26.80 28.60 25.70 11.7500 17.4231 19.6851 CT Image 2 6.6622 0.8386 0.6958 38.7416 29.9663 29.4611 28.40 37.20 28.40 11.7969 16.4425 18.5039 CT Image 3 13.2030 1.0632 1.0482 35.8632 30.6855 31.0718 43.20 43.60 41.20 16.3281 18.6835 23.0971 CT Image 4 5.0707 1.1561 1.0581 38.9335 19.5331 20.6044 25.00 24.50 23.60 10.8594 18.5866 21.1245 MR Image 5 6.5164 0.9678 0.9345 43.6002 32.7573 33.1091 21.90 26.70 23.40 12.0000 18.5779 23.6943 MR Image 6 10.9643 1.1854 1.1463 41.7960 22.9515 23.1293 26.40 25.70 23.10 14.8125 18.2230 21.6958 MR Image 7 4.4960 1.2092 0.3788 39.0490 22.3658 16.4623 23.00 25.40 23.00 10.5625 18.2627 18.4104 MR Image 8 7.5873 1.117 1.1054 42.3063 29.9071 30.0208 25.10 27.40 22.70 12.3906 18.7018 20.8762 MR Image 9 5.1806 1.1416 0.9305 43.6769 24.6009 24.6205 20.50 21.90 22.00 10.9688 18.3615 18.9826 MR Image 10 6.9771 1.0116 1.0085 43.3527 44.3570 46.0139 23.70 25.90 23.50 12.1094 18.6340 20.9603 MR Image 11 6.5397 0.9904 0.9703 43.0880 44.2258 44.5761 23.00 25.80 23.00 11.9219 18.7667 21.6570 MR Image 12 3.4217 0.956 0.9271 40.7357 29.3755 31.4664 22.30 26.30 22.30 9.8438 17.8159 19.7808 MR Image 13 1.8435 1.0728 0.8469 43.4892 50.9330 44.3137 17.80 17.30 21.00 9.0625 17.9883 20.0986 MR Image 14 2.3289 1.1096 0.7947 41.3458 35.7803 27.3508 15.90 71.50 15.90 9.0938 17.2854 18.6382 MR Image 15 3.0478 1.0365 0.9788 39.6035 46.2050 42.7904 20.50 27.40 20.50 9.4844 17.7939 18.6295 MR Image 16 3.4444 1.0499 0.8910 39.9027 32.9302 28.8155 22.70 27.30 22.70 10.0625 18.7660 20.1662 MR Image 17 4.8070 0.9978 0.9794 39.0449 37.5945 37.4232 29.90 33.30 29.90 10.7031 26.7360 20.4192 MR Image 18 3.0478 1.0365 0.9788 3.0478 1.0365 0.9788 20.50 27.40 20.50 9.5000 27.1599 18.9999 MR Image 19 3.4270 1.1379 0.8874 3.4270 1.1379 0.8874 28.20 41.40 28.20 10.1094 17.9907 19.0742 MR Image 20 4.1237 1.003 0.9852 34.1952 38.2271 38.9271 32.70 40.60 33.20 10.5156 18.1835 20.2288 MR Image 21 6.2566 1.0636 0.6189 43.3019 32.7369 26.1410 20.60 22.90 20.60 11.5313 18.5056 18.7640 PET Image 22 5.5154 1.0352 1.0080 43.3578 33.4930 32.3096 19.20 22.80 19.10 11.2656 18.8691 18.8583 PET Image 23 1.4721 0.7923 0.3612 43.9157 24.4163 21.6748 14.30 18.70 39.60 8.5625 16.5292 19.5478 PET Image 24 1.1946 1.0696 1.0687 39.3448 52.1345 52.1528 18.80 19.60 17.50 8.8125 16.7909 20.1511
120
Figure 7.3 Compression Ratio expressed in percentage
Figure 7.3 shows the compression ratio of three different algorithms namely
Neural Network Radial Basis Function, Fractal and Hybrid Fractal & NNRBF. It is
clearly evident that Hybrid Fractal & NNRBF provides better CR values.
Figure 7.4 PSNR expressed in decibel
Figure 7.4 Shows the PSNR for three different algorithms NNRBF, Fractal and
Hybrid Fractal & NNRBF. Here Hybrid Fractal & NNRBF provides better PSNR values.
121
Figure 7.5 Memory expressed in kilo byte
Figure 7.5 shows the memory usage of three different algorithms namely,
NNRBF, Fractal and Hybrid Fractal & NNRBF. It is noticed that Hybrid Fractal &
NNRBF uses lesser memory for image compression.
Figure 7.6 Execution Time expressed in Seconds
Figure 7.6 Shows the Execution time of three different algorithms namely,
NNRBF, Fractal and Hybrid Fractal & NNRBF. Here, Hybrid Fractal & NNRBF
produces compression results within minimal time duration.
122
Figure 7.7(A) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). NNRBF and (d). Hybrid Fractal & NNRBF algorithms
123
Figure 7.7(B) Results obtained for various medical images (a). Input Images (b). Fractal, (c). NNRBF and (d). Hybrid Fractal & NNRBF
algorithms
124
Figure 7.7(C) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). NNRBF and (d). Hybrid Fractal & NNRBF algorithms
125
Figure 7.7(D) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). NNRBF and (d). Hybrid Fractal & NNRBF algorithms
126
Figure 7.7(E) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). NNRBF and (d). Hybrid Fractal & NNRBF algorithms
127
Figure 7.7(F) Results obtained for various medical images
(a). Input Images (b). Fractal, (c). NNRBF and (d). Hybrid Fractal & NNRBF algorithms
128
Figures 7.7 (A-F) (a) Input images, (b) output obtained from Fractal, (c) output
obtained from NNRBF and (d) output obtained from Hybrid Fractal & NNRBF
compression algorithms.
7.6 CONCLUSION
The standard prior techniques have been studied in the past to overcome the
issues of compressing medical images. After reviewing the prior studies, it can be said
that performance of medical image compression is highly dependent on the compression
ratio as well as perceptible quality of a compressed image. Compressed image with better
perceptual quality will retain the information of the clinical importance to higher degree
and will aid the diagnostic to have better result. Although a plenty of research work has
been carried out in the past for exploring an efficient solution for performing medical
image compression. In the existing Fractal algorithm, Compression Ratio is efficient but
Mean Square error is high, wherein PSNR value becomes low. Neural Network Radial
Basis Function takes very less convergence time during training period. Implementation
of a new radial basis function network based on neural network scheme and the
comparison with existing algorithms is carried out. The proposed Hybrid Fractal with
NNRBF based image compression method undergoes better compression ratio and PSNR
values. The proposed algorithm requires minimum time duration and reduced memory
space for performing image compression. Image compression is provided by Fractal,
Radial Basis Function Neural Network and Hybrid Fractal & NNRBF algorithms. Hybrid
FNNRBF provides compressed images with lower CR in minimized time duration on
comparison with Fractal and NNRBF algorithms. The soft computing techniques
proposed through this research have been considered in image compression and have
produced better image quality on comparison with analytical and statistical algorithms.
Three different approaches such as Fractal, Radial Basis Function Neural Network and
Hybrid Fractal & NNRBF are applied to medical image compression and compared.
129
Here, MR and CT images are compared using quality parameters such as CR and PSNR,
Execution time and Memory usage. It is observed that FNNRBF method has low CR and
high PSNR values. Hybrid Fractal & NNRBF is found to be more efficient than Fractal
and NNRBF methods.
130
CHAPTER 8
CONCLUSION AND FUTURE WORK
8.1 CONCLUSION
In brief, the following conclusions are made:
• Image compression process is carried out using Fractal, Discrete Cosine
Transform, Discrete Wavelet Transform, Set Partitioning in Hierarchical Trees,
Neural Network Back Propagation, Neural Network Radial Basis Function,
Hybrid Discrete Wavelet Transform with Neural Network Back Propagation
and Hybrid Fractal with Neural Network Radial Basis Function methods.
• The performances of conventional methods are found to be dissatisfied for
MRI, CT and PET image compression.
• The performance of Neural Network Back Propagation and Neural Network
Radial Basis Function for compression process is improved by means of
introducing hybrid technology.
• Neural Network Radial Basis Function algorithm is effective in obtaining
better compression results of various CT, MRI and PET images.
• Hybrid Fractal with Neural Network Radial Basis Function based image
compression is found to be better when compared to conventional compression
methods.
131
Table 8.1 Shows Compression ratio of 24 medical images which are obtained by using
DCT,DWT,Fractal, NNBP, NNRBF, Hybrid Fractal and NNRBF and Hybrid DWT-NNBP
algorithm •
Images DCT DWT Fractal NNBP NNRBF Hybrid
Fractal & NNRBF
Hybrid DWT-NNBP
CT Image 1 1.4088 1.0002 6.6507 1.0495 1.0537 1.0424 1.0005 CT Image 2 1.3442 1.0011 6.6622 1.0356 0.8386 0.6958 1.0121 CT Image 3 1.4679 1.0015 13.2030 1.0330 1.0632 1.0482 1.0016 CT Image 4 1.3443 1.0020 5.0707 1.0328 1.1561 1.0581 0.568 MR Image 5 1.4402 1.0005 6.5164 1.0564 0.9678 0.9345 0.9254 MR Image 6 1.6127 1.0048 10.9643 1.0423 1.1854 1.1463 0.8844 MR Image 7 1.3586 1.0229 4.4960 1.0398 1.2092 0.3788 0.0969 MR Image 8 1.5927 1.0265 7.5873 1.0442 1.117 1.1054 0.5601 MR Image 9 1.5900 1.0144 5.1806 1.0208 1.1416 0.9305 0.0999 MR Image 10 1.2792 1.0001 6.9771 1.0591 1.0116 1.0085 1.005 MR Image 11 1.3006 0.9999 6.5397 1.0605 0.9904 0.9703 0.5603 MR Image 12 1.4355 1.0073 3.4217 1.0536 0.956 0.9271 0.8914 MR Image 13 1.4994 1.0013 1.8435 1.0837 1.0728 0.8469 0.9743 MR Image 14 1.3451 1.0006 2.3289 1.0728 1.1096 0.7947 0.0518 MR Image 15 1.6541 1.0001 3.0478 1.0514 1.0365 0.9788 1.0052 MR Image 16 1.5732 1.0023 3.4444 1.0513 1.0499 0.8910 1.0115 MR Image 17 1.4197 1.0001 4.8070 1.0426 0.9978 0.9794 0.9986 MR Image 18 1.6541 1.0001 3.0478 1.0514 1.0365 0.9788 1.0052 MR Image 19 1.7575 1.0152 3.4270 0.9938 1.1379 0.8874 0.0471 MR Image 20 1.0884 1.0015 4.1237 1.0398 1.003 0.9852 1.002 MR Image 21 1.2792 1.0001 6.2566 1.0591 1.0636 0.6189 1.005 PET Image 22 1.5891 1.0069 5.5154 1.0530 1.0352 1.0080 0.4557 PET Image 23 1.4908 1.0006 1.4721 0.7662 0.7923 0.3612 0.2325 PET Image 24 1.5791 1.0048 1.1946 1.0709 1.0696 1.0687 1.0007
132
Table 8.2 Shows Peak Signal to Noise Ratio of 24 medical images which are obtained
by using DCT, DWT, Fractal, NNBP, NNRBF, Hybrid Fractal and NNRBF and
Hybrid DWT-NNBP algorithm •
Images DCT DWT Fractal NNBP NNRBF Hybrid Fractal
-NNRBF Hybrid DWT-NNBP
CT Image 1 88.8087 71.7728 38.2519 69.6090 39.1906 34.1945 61.3201 CT Image 2 88.3888 71.3010 38.7416 42.3719 29.9663 29.4611 39.1269 CT Image 3 85.8091 55.7033 35.8632 62.5396 30.6855 31.0718 47.2912 CT Image 4 92.4192 67.3440 38.9335 40.2157 19.5331 20.6044 29.4004 MR Image 5 90.5865 67.7753 43.6002 62.0237 32.7573 33.1091 38.5935 MR Image 6 91.8776 63.8786 41.7960 46.7436 22.9515 23.1293 31.9976 MR Image 7 92.2370 62.7552 39.0490 44.7684 22.3658 16.4623 29.4417 MR Image 8 97.7195 60.1968 42.3063 48.5713 29.9071 30.0208 23.6815 MR Image 9 101.4771 64.2066 43.6769 45.7868 24.6009 24.6205 25.814 MR Image 10 90.5101 67.8580 43.3527 73.0390 44.3570 46.0139 47.1406 MR Image 11 90.6810 68.4217 43.0880 68.0856 44.2258 44.5761 43.8376 MR Image 12 99.6740 67.9670 40.7357 54.1458 29.3755 31.4664 38.7958 MR Image 13 114.2706 72.8183 43.4892 66.8312 50.9330 44.3137 38.0442 MR Image 14 97.6726 72.5798 41.3458 47.6965 35.7803 27.3508 24.6016 MR Image 15 98.4225 64.5986 39.6035 72.3725 46.2050 42.7904 43.2033 MR Image 16 95.1357 65.2455 39.9027 57.2598 32.9302 28.8155 40.3127 MR Image 17 96.8298 61.9274 39.0449 84.0402 37.5945 37.4232 64.8598 MR Image 18 98.4225 64.5986 3.0478 72.3725 1.0365 0.9788 43.2033 MR Image 19 90.0821 54.2999 3.4270 36.2982 1.1379 0.8874 23.7107 MR Image 20 91.1779 60.5071 34.1952 69.1297 38.2271 38.9271 49.3025 MR Image 21 90.5101 67.8580 43.3019 73.0390 32.7369 26.1410 47.1406 PET Image 22 97.2405 66.0979 43.3578 50.2760 33.4930 32.3096 27.168 PET Image 23 112.8808 78.9087 43.9157 42.7221 24.4163 21.6748 28.3994 PET Image 24 112.1955 70.6387 39.3448 84.0081 52.1345 52.1528 48.427
133
Table 8.3 Shows Memory of 24 medical images which are obtained by using DCT,
DWT, Fractal, NNBP, NNRBF, Hybrid Fractal and NNRBF and Hybrid DWT-
NNBP algorithm •
Images DCT DWT Fractal NNBP NNRBF Hybrid
Fractal & NNRBF
Hybrid DWT-NNBP
CT Image 1 21.40 28.70 26.80 28.70 28.60 25.70 24.1 CT Image 2 23.20 29.80 28.40 30.10 37.20 28.40 17.2 CT Image 3 31.60 44.80 43.20 44.90 43.60 41.20 30.3 CT Image 4 21.10 26.90 25.00 27.40 24.50 23.60 26.9 MR Image 5 18.00 24.50 21.90 24.50 26.70 23.40 51.4 MR Image 6 18.90 28.90 26.40 29.20 25.70 23.10 70.5 MR Image 7 22.60 28.60 23.00 29.50 25.40 23.00 49.6 MR Image 8 19.20 28.40 25.10 29.30 27.40 22.70 44.9 MR Image 9 15.70 23.30 20.50 24.50 21.90 22.00 55.5 MR Image 10 20.50 24.80 23.70 24.80 25.90 23.50 11.6 MR Image 11 19.60 24.10 23.00 24.10 25.80 23.00 10.3 MR Image 12 17.50 23.70 22.30 23.90 26.30 22.30 22.1 MR Image 13 12.40 17.10 17.80 17.10 17.30 21.00 17.8 MR Image 14 14.40 18.00 15.90 18.10 71.50 15.90 3.32 MR Image 15 17.20 27.00 20.50 27.00 27.40 20.50 27 MR Image 16 18.20 27.20 22.70 27.30 27.30 22.70 45.2 MR Image 17 23.40 31.90 29.90 31.90 33.30 29.90 5.78 MR Image 18 17.20 27.00 20.50 27.00 27.40 20.50 27 MR Image 19 26.80 45.00 28.20 47.10 41.40 28.20 70.5 MR Image 20 27.40 39.10 32.70 39.20 40.60 33.20 19.2 MR Image 21 20.50 24.80 20.60 24.80 22.90 20.60 11.6 PET Image 22 14.80 21.90 19.20 22.40 22.80 19.10 23.3 PET Image 23 9.97 13.40 14.30 19.40 18.70 39.60 15 PET Image 24 13.20 19.40 18.80 19.50 19.60 17.50 9.54
134
Table 8.4 Shows Execution time of 24 medical images which are obtained by using
DCT, DWT, Fractal, NNBP, NNRBF, Hybrid Fractal and NNRBF and Hybrid DWT-
NNBP algorithm
Images DCT DWT Fractal NNBP NNRBF Hybrid Fractal & NNRBF Hybrid DWT-NNBP
CT Image 1 0.7572 1.0867 11.7500 850.4214 17.4231 19.6851 848.0286 CT Image 2 0.7645 1.0818 11.7969 278.7227 16.4425 18.5039 279.6484 CT Image 3 0.7522 2.3126 16.3281 542.2186 18.6835 23.0971 546.848 CT Image 4 0.8272 1.6483 10.8594 935.6339 18.5866 21.1245 938.7183 MR Image 5 0.9514 1.1104 12.0000 387.0713 18.5779 23.6943 354.3677 MR Image 6 0.8773 1.1089 14.8125 667.3835 18.2230 21.6958 668.1684 MR Image 7 0.7620 1.2349 10.5625 658.0410 18.2627 18.4104 656.5766 MR Image 8 1.5385 1.0894 12.3906 1164.2615 18.7018 20.8762 874.3936 MR Image 9 0.7825 1.0646 10.9688 836.9937 18.3615 18.9826 798.0145 MR Image 10 1.1055 1.1229 12.1094 751.2015 18.6340 20.9603 786.3682 MR Image 11 0.8540 1.1197 11.9219 294.9351 18.7667 21.6570 287.1838 MR Image 12 0.7832 1.1865 9.8438 415.8369 17.8159 19.7808 425.006 MR Image 13 0.7919 1.0898 9.0625 497.0650 17.9883 20.0986 504.9017 MR Image 14 1.4770 1.0197 9.0938 313.2422 17.2854 18.6382 321.9819 MR Image 15 0.7612 1.0705 9.4844 285.6100 17.7939 18.6295 293.2718 MR Image 16 0.8803 1.0782 10.0625 610.1797 18.7660 20.1662 633.1677 MR Image 17 1.0694 1.1614 10.7031 843.5427 26.7360 20.4192 858.127 MR Image 18 0.9104 1.0800 9.5000 285.4747 27.1599 18.9999 293.2718 MR Image 19 1.0884 1.0605 10.1094 654.8289 17.9907 19.0742 663.0196 MR Image 20 1.2724 1.2132 10.5156 504.8591 18.1835 20.2288 551.1406 MR Image 21 0.7581 1.0645 11.5313 744.2455 18.5056 18.7640 758.223 PET Image 22 0.8768 1.1157 11.2656 652.3676 18.8691 18.8583 653.9199 PET Image 23 0.7791 1.0681 8.5625 452.7775 16.5292 19.5478 463.315 PET Image 24 1.0188 1.1340 8.8125 378.8078 16.7909 20.1511 388.1468
Medical Image Compression is provided by Fractal, DCT, DWT,
SPIHT, NNBP, NNRBF, Hybrid DWT-NNBP and Hybrid Fractal-NNRBF
algorithms. Among the above said algorithms, the comparison of results show
that Hybrid Fractal-NNRBF produce compression images with better CR, PSNR
and bandwidth required to save the image. Soft computing techniques for image
135
compression have produced better image quality when compared to analytical and
the statistical algorithms. This proves that Hybrid Fractal-NNRBF algorithm is
commendable and this research concludes that high quality compression of CT,
MRI and PET image is offered through Hybrid Fractal with Neural Network
Radial Basis Function method.
8.2 FUTURE WORK
• The Hybrid-Neural Network based method can be combined with other
evolutionary methods and Level set methods to give better compression
results.
• The extension of this application can be suggested for other images for
compression and in the analysis of real time diagnosis.
• The methodologies used in this research can also be extended for the
compression of images pertaining to Oncology.
136
REFERENCES
[1] Abdul Khader Jilani Saudagar and Omar A. Shathry, Neural Network Based Image
Compression Approach to Improve the Quality of Biomedical Image for
Telemedicine, British Journal of Applied Science & Technology, 4(3), (2014), 510-
524.
[2] Abirami.J, Siva sankari.S and Narashiman.K, Image compression based on
WaveletSupport Vector Machine Kernels. International Journal of Engineering and
Technology (IJET), 5(2), (2013), 1584-1588.
[3] Adnan Khashman and Kamil Dilmililer, Medical Radiographs Compression using
Neural Networks and Haar Wavelet, IEEE Transaction, (2009), 1448-1453.
[4] Ajay Kumar Bhagat and Er. Dipti Bansal, Image Fusion Using Hybrid Method with
Singular Value Decomposition and Wavelet Transfor, International Journal of
Emerging Technology and Advanced Engineering, 4, (2014), 827-830.
[5] Alagendran B, Manimurugan S, A Survey on Various Medical Image Compression
Techniques. International Journal of Soft Computing and Engineering (IJSCE), 2(1),
(2012), 425-428.
[6] Alex Alexandridis, Eva Chondrodima, and Haralambos Sarimveis, Radial Basis
Function Network Training Using a Nonsymmetric Partition of the Input Space and
Particle Swarm Optimization. IEEE Transactions on Neural Networks and Learning
Systems, 24(2), (2013), 219-230.
[7] Al-Fahoum A and Harb B, A Combined Fractal and Wavelet Angiography Image
Compression Approach, The Open Medical Imaging Journal, 7, (2013), 9-18.
[8] Ali Al-Fayadh, Abir Jaafar Hussain, Paulo Lisboa, Dhiya AI- Jumeily and M.AI-
Jumaily, A Hybrid Image Compression Method and Its Application to Medical
Images, Second International Conference on Developments in eSystems Engineering,
(2009),107-112.
137
[9] Alok kumar singh, G.S. Tripathi, A Comparative Study of DCT, DWT & Hybrid
(DCT-DWT) Transform, GJESR Review Paper, 1(4), (2014), 2349-2355.
[10] Andrew Martchenko and Guang Deng, Bayesian Predictor Combination for Lossless
Image Compression, IEEE Transactions on Image Processing, 22(12), (2013), 5263-
5270.
[11] Anil Bhagat, Balasaheb Deokate, Improve Image Quality At Low Bitrate Using
Wavelet Based Fractal Image Coder, International Journal of Advanced Research in
Electrical, Electronics and Instrumentation Engineering , 2(8), (2013), 3693-3702.
[12] Anjana Jianyu Lin and Mark J. T. Smith, Two-Band Hybrid FIR–IIR Filters for
Image Compression, IEEE Transactions on Image Processing, 20(11), (2011),
3063-3072.
[13] Ankita Vaish and Manoj Kuma, A new Image Compression Technique using
Principal Component Analysis and Huffman Coding. International Conference on
Parallel, Distributed and Grid Computing, (2014), 301-305.
ANN and DWT, International Journal of Computer Applications, 95(11), (2014),
35-38.
[14] Anna Durai, and E. Anna Saro, Image Compression with Back-Propagation Neural
Network using Cumulative Distribution Function, World Academy of Science,
Engineering and Technology, 2, (2008), 802-806.
[15] Arif Sameh Arif, Sarina Mansor, Hezrul abdul karim and Rajasvaran Logeswaran,
Lossless Compression of Fluoroscopy Medical Images using Correlation and the
Combination of Run-length and Huffman Coding, IEEE EMBS International
Conference on Biomedical Engineering and Sciences, (2012), 759-762.
[16] Arun Vikas Singh and Srikanta Murthy K, Neuro-Wavelet based Efficient Image
Compression using Vector Quantization, International Journal of Computer
Applications, 49(3), (2012), 33-40.
[17] Arunpriya C. and Antony Selvadoss Thanamani, An Effective Tea Leaf Recognition
Algorithm for Plant Classification Using Radial Basis Function Machine,
138
International Journal of Modern Engineering Research (IJMER), 4(3), (2014),35-
44.
[18] Bairagi and A.M. Sapkal, Automated region-based hybrid compression for
[19] Bhammar and Prof. K.A. Mehta, Survey of Various Image Compression Techniques,
International Journal of Darshan Institute on Engineering Research & Emerging
Technologies, 1(1), (2012), 85-90.
[20] Birendra Kumar Patel, Prof. Suyash Agrawal, Image Compression Techniques Using
Artificial Neural Network, International Journal of Advanced Research in Computer
Engineering & Technology (IJARCET), 2(10), (2013), 2725-2729.
[21] Chakrapani and K. Soundera Rajan, Implementation of Fractal Image Compression
Employing Hybrid Genetic-Neural Approach. International Journal Of
Computational Cognition, 7(3), (2009), 33-39.
[22] Chander mukhi, Pallavi Nayyar, Mandeep Singh Saini, Improved Image
Compression using Hybrid Transform. International Journal of Science and
Research (IJSR), 2(1), (2013), 316-319.
[23] Chaudhari and S. B. Dhok, Wavelet Transformed based Fast Fractal Image
Compression, International Conference on Circuits, Systems, Communication and
Information Technology Applications (CSCITA), (2014), 65-69.
[24] Chiang and L.M.Po. Adaptive lossy LZW algorithm for palettised image
compression. Electronics Letters, 33, (1997),10.
[25] Christophe Amerijckx, Michel Verleysen, Philippe Thissen, and Jean-Didier Legat,
Image Compression by Self-Organized Kohonen Map. IEEE Transactions on Neural
Networks, 9(3), (1998), 503-507.
[26] Chunlei Jiang and Shuxin Yin, A Hybrid Image Compression Algorithm Based on
Human Visual System, International Conference on Computer Application and
System Modeling (ICCASM 2010), 9, (2010), 170-173.
[27] Ci Wang, Hong-Bin Yu, and Meng Zheng, A DCT-based MPEG-2 Transparent
Scrambling Algorithm, IEEE Transactions on Consumer Electronics, 49(4), (2003),
1208-1213.
139
[28] Debin Zhao, Wen Gao, and Y. K. Chan, Morphological Representation of DCT
Coefficients for Image Compression. IEEE Transactions On Circuits And Systems
For Video Technology, 12(9), (2002), 819-823.
digital imaging and communications in medicine magnetic resonance imaging
images for telemedicine applications, IET Sci. Meas. Technol, 6(4), (2012), 247-
253.
[29] Dipta Pratim Dutta, Samrat Deb Choudhury, Md. Anwar Hussain, and Swanirbhar
Majumder, Digital Image Compression using Neural Networks, International
Conference on Advances in Computing, Control, and Telecommunication
Technologies, (2009), 116-120.
[30] Ferni Ukrit and G.R.Suresh, Hybrid Algorithm for Medical Image Sequences using
Super-Spatial Structure Prediction with LZ8. International Journal of Computer
Applications, 86(11), (2014), 10-15.
[31] Harjeetpal singh and Sakhi Sharma, Hybrid Image Compression Using DWT, DCT
& Huffman Encoding Techniques, International Journal of Emerging Technology
and Advanced Engineering, 2(10), (2012), 300-306.
[32] Jaffar Iqbal Barbhuiya, Tahera Akhtar Laskar and K. Hemachandran, An Approach
for Color Image Compression of JPEG and PNG Images using DCT And DWT.
Sixth International Conference on Computational Intelligence and Communication
Networks, (2014), 129-133.
[33] Jagadish H. Pujar and Lohit M. Kadlaskar, A New Lossless Method Of Image
Compression And Decompression Using Huffman Coding Techniques. Journal of
Theoretical and Applied Information Technology, (2010), 18-23.
[34] Jiaji Wu, Chong Liang, Jianxiang Han, Zejun Hu, Dehong Huang, Hongqiao Hu,
Yong Fang, Licheng Jiao, Two-stage lossless compression algorithm for aurora
image using weighted motion compensation and context-based model, Elsevier
Journal, 290, (2013), 19-27.
140
[35] Jianji Wang, and Nanning Zheng, A Novel Fractal Image Compression Scheme with
Block Classification and Sorting Based on Pearson’s Correlation Coefficient. IEEE
Transactions on Image Processing, 22(9), (2013), 3690-3702.
[36] Jian-Jiun Ding, Hsin-Hui Chen, and Wei-Yi Wei, Adaptive Golomb Code for Joint
Geometrically Distributed Data and Its Application, Image Coding. IEEE
Transactions On Circuits And Systems For Video Technology, 23(4), (2013), 661-
670.
[37] Jie Liang, Trac D. Tran and Ricardo L. de Queiroz, DCT-Based General Structure
for Linear-Phase Paraunitary Filterbanks. IEEE Transactions On Signal Processing,
51(6), (2003), 1572-1580.
[38] Jonathan Taquet and Claude Labit, Hierarchical Oriented Predictions for Resolution
Scalable Lossless and Near-Lossless Compression of CT and MRI Biomedical
Images, IEEE Transactions on Image Processing, 21(5), (2012), 2641-2652.
[39] Jyh-Horng Jeng, Chun-Chieh Tseng, and Jer-Guang Hsieh, Study on Huber Fractal
Image Compression, IEEE Transactions on Image Processing, 18(5), (2009), 995-
1003.
[40] Kai-jen Cheng, and Jeffrey Dill, Lossless to Lossy Dual-Tree BEZW Compression
for Hyperspectral Images, IEEE Transactions on Geoscience and Remote Sensing,
52(9), (2014), 5765-5770.
[41] Kaur, R.C. Chauhan and S.C. Saxena, Adaptive compression of medical ultrasound
images, IEE Proc.-Vis. Image Signal Process., 153(2), (2006), 185-190.
[42] Kesavamurthy Thangavelu, Thiyagarajan Krishnan, Lossless Color Medical Image
Compression using Adaptive Block-Based Encoding for Human Computed
Tomographic Images, Wiley Periodicals, Inc, 23, (2013), 227-234.
[43] Kuppusamy, R.Ilackiya, Fractal Image Compression & Algorithmic Techniques,
International Journal of Computer & Organization Trends, 3(4), (2013), 141-145.
[44] Leyuan Fang, Shutao Li, Xudong Kang, Joseph A. Izatt, and Sina Farsiu, 3-D
Adaptive Sparsity Based Image Compression With Applications to Optical
141
Coherence Tomography. IEEE Transactions on Medical Imaging, 34(6), (2015),
1306-1320.
[45] Long Zhang, KangLi, HaiboHe and George W. Irwin, A New Discrete-Continuous
Algorithm for Radial Basis Function Networks Construction. IEEE Transactions on
Neural Networks and Learning Systems, 24(11), (2013), 1785-1798.
[46] Merav Huber-Lerner, Ofer Hadar, Stanley R. Rotman, and Revital Huber-Shalem,
Compression of Hyperspectral Images Containing a Subpixel Target. IEEE Journal
Of Selected Topics In Applied Earth Observations And Remote Sensing, 7(6), (2014),
2246-2255.
[47] Mohamed Abo-Zahhad, Reda Ragab Gharieb, Sabah M. Ahmed, Mahmoud Khaled
Abd-Ellah, Huffman Image Compression Incorporating DPCM and DWT, Journal of
Signal and Information Processing, 6, (2015), 123-135.
[48] Mohamed El Zorkany, A Hybrid Image Compression Technique Using Neural
Network and Vector Quantization With DCT, Springer International Publishing
Switzerland, 5, (2014), 233-244.
[49] Monika Narwal and Er.Tarun Jeet Singh Chugh, Image Compression By Hybrid
Transformation Technique, International Journal of Advanced Research in
Computer Science and Software Engineering, 3(7), (2013), 474-479.
[50] Ng and L.M. Cheng, Data re-ordering technique for lossless LZW continuous-tone
image compression. Electronics Letters, 35, (1999), 20.
[51] Nikita Bansal and Sanjay Kumar Dubey, Image Compression Using Hybrid
Transform Technique, Journal of Global Research in Computer Science, 4(1),
(2013) 13-17.
[52] Omar Arif and Patricio Antonio Vela, Kernel Map Compression using Generalized
Radial Basis Functions, IEEE 12th International Conference on Computer Vision
(ICCV), (2009), 1119-1124.
[53] Panda, M.S.R.S Prasad, MNM Prasad, Ch. SKVR Naidu, Image Compression Using
Back Propagation Neural Network, International Journal Of Engineering Science &
Advanced Technology, 2(1), (2012), 74-78.
142
[54] Patil and A.R.Yardi, ANN based Dementia Diagnosis using DCT for Brain MR
Image compression. International conference on Communication and Signal
Processing, (2013), 451-454.
[55] Pawel Turcza and Mariusz Duplaga, Hardware-Efficient Low-Power Image
Processing System for Wireless Capsule Endoscopy, IEEE Journal Of Biomedical
And Health Informatics, 17(6), (2013),1046-1056.
[56] Praisline Jasmi, Mr.B.Perumal and Dr.M.Pallikonda Rajasekaran, Comparison of
Image Compression Techniques using Huffman Coding, DWT and Fractal
Algorithm, International Conference on Computer Communication and Informatics
(ICCCI), (2015).
[57] Prema Karthikeyan, Narayanan Sreekumar, A Study on Image Compression with
Neural Networks Using Modified Levenberg Maruardt Method, Global Journal of
Computer Science and Technology, 11(3), (2011).
[58] Priya Pareek, Manish Shrivastava, An Image Compression Using Multilayer Wavelet
Transform with 2DTCWT: A Review, International Journal of Computer
Applications, 102(1), (2014), 13-17.
[59] Renato J. Cintra, and Fábio M. Bayer, A DCT Approximation for Image
Compression. IEEE Signal Processing Letters, 18(10), (2011), 579-582.
[60] Reny Catherin L, Thirupurasunthari P, Sherley Arcksily Sylvia A, Sravani Kumari
G, Joany R.M and N.M. Nandhitha, A Survey on Hybrid Image Compression
Techniques for Video Transmission. International Journal of Electronics and
Communication Engineering, 6(3), (2013), 217-224.
[61] Robina Ashraf and Muhammad Akbar, Absolutely Lossless Compression of Medical
Images, Engineering in Medicine and Biology 27th Annual Conference, (2005),
4006-4009.
[62] Sridhar M.I.S.T.E, V.Venugopal, S. Ramesh, Srinivas nd Sk. Mansoob, Wavelets
and Neural Networks based Hybrid Image Compression Scheme, International
Journal of Emerging Trends & Technology in Computer Science (IJETTCS), 2,
(2013), 195-200.
143
[63] Saravanan and R. Ponalagusamy, Lossless Grey-scale Image Compression using
Source Symbols Reduction and Huffman Coding, International Journal of Image
Processing (IJIP), 3, (2010), 246-251.
[64] Seyun Kimand, Nam Ik Cho, Hierarchical Prediction and Context Adaptive Coding
for Lossless Color Image Compression, IEEE Transactions On Image Processing,
23(1), (2014), 445-449.
[65] Shaou-Gang Miaou, Fu-Sheng Ke, and Shu-Ching Chen, A Lossless Compression
Method for Medical Image Sequences Using JPEG-LS and Interframe Coding, IEEE
Transactions on Information Technology in Biomedicine, 13(5), (2009), 818-821.
[66] Shiqiang Yan and Zhong Xiao, Application of BP Neural Network with Chebyshev
Mapping in Image Compression, International Conference on Instrumentation,
Measurement, Computer, Communication and Control, (2013), 398-402.
[67] Shruti Puniani, Er. Nishi Madaan, Various Image Compression Techniques: A
Review, International Journal of Advanced Research in Computer Engineering &
Technology (IJARCET), 3(4), (2014), 1164-1170.
[68] Sophin Seeli , Dr.M.K.Jeyakumar, A Study on Fractal Image Compression using
Soft Computing Techniques, IJCSI International Journal of Computer Science
Issues, 9(2), (2012), 420-430.
[69] Sridevi, Dr.V.R.Vijayakuymar and Ms.R.Anuja, A Survey on Various Compression
Methods for Medical Images. I.J. Intelligent Systems and Applications, 3, (2012),
13-19.
[70] Sridharan Bhavani, Kepanna Gowder Thanushkodi, Comparison of fractal coding
methods for medical image compression, IET Image Process, 7(7), (2013), 686-693.
[71] Taha mohammed Hasan and Xingqian Wu, An Adaptive Fractal Image
Compression., IJCSI International Journal of Computer Science Issues, 10(1),
(2013), 98-108.
[72] Tajallipour and Khan Wahid, Efficient Implementation of Adaptive LZW Algorithm
for Medical Image Compression. International Conference on Computer and
Information Technology, (2011), 987.
144
[73] Tamilarasi and Dr.V.Palanisamy, Contourlet Based Medical Image Compression
Using Improved EZW, International Conference on Advances in Recent
Technologies in Communication and Computing, (2009), 800-804.
[74] Tiruvenkadam Santhanam and A.C. Subhajini, An Efficient Weather Forecasting
System using Radial Basis Function Neural Network, Journal of Computer Science,
7(7), (2011), 962-966.
[75] Václav Šimek and Ram Rakesh ASN, GPU Acceleration of 2D-DWT Image
Compression in MATLAB with CUDA. Second UKSIM European Symposium on
Computer Modeling and Simulation, (2008), 274-277.
[76] Vasanthi Kumari P and K Thanushkodi, Image Compression Using Wavelet
Transform and Graph Cut Algorithm. Journal of Theoretical and Applied
Information Technology, 53(3), (2013), 437-445.
[77] Vidhya and S. Shenbagadevi, Medical Image Compression Using Hybrid Coder
With Fuzzy Edge Detection, ICTACT Journal on Image and Video Processing, 1(3),
(2011), 138-142.
[78] Vilas Gaidhane, Vijander Singh and Mahendra Kumar, Image Compression using
PCA and Improved Technique with MLP Neural Network, International Conference
on Advances in Recent Technologies in Communication and Computing, (2010),
106-110.
[79] Vilas H. Gaidhane, Vijander Singh, Yogesh V. Hote and Mahendra Kumar, New
Approaches for Image Compression Using Neural Network, Journal of Intelligent
Learning Systems and Applications, 3, (2011), 220-229.
[80] Xiaofeng Li and Yi Shen, A Medical Image Compression Scheme Based on Low
Order Linear Predictor and Most-likely Magnitude Huffman Code. International
Conference on Mechatronics and Automation, (2006),1796-1800.
[81] Yeo, David F. W. Yap, T.H. Oh, D.P. Andito, S. L. Kok, Y. H. Ho, M. K. Suaidi,
Grayscale Medical Image Compression using Feedforward Neural Networks,
International Conference on Computer Applications and Industrial Electronics,
(2011), 633-638.
145
[82] Yongfei Zhang, Haiheng Cao, Hongxu Jiang and Bo Li, Visual Distortion Sensitivity
Modeling for Spatially Adaptive Quantization in Remote Sensing Image
Compression, IEEE Geoscience And Remote Sensing Letters, 11(4), (2014), 723-
727.
[83] Yongjian Nian, MiH and Jianwei Wan, Distributed near lossless compression
algorithm for hyperspectral images, Elsevier Journal, 40, (2014), 1006-1014.
[84] Yung-Gi Wu and Shen-Chuan Tai, Medical Image Compression by Discrete Cosine
Transform Spectral Similarity Strategy, IEEE Transactions on Information
Technology in Biomedicine, 5(3), (2001), 236-243.
[85] Yung-Gi Wu, Medical Image Compression by Sampling DCT Coefficients, IEEE
Transactions on Information Technology in Biomedicine, 6(1), (2002), 86-94.
146
LIST OF PUBLICATIONS
International Journal - Published:
1. Perumal, B, M.Pallikonda Rajasekaran ”Efficient Image Compression
Techniques for Compressing Multimodal Medical images using Neural
Network Radial Basis Function Approach” publication in International
Journal of Imaging Systems and Technology Volume 25, Issue 2, pages
115–122, June 2015, Article first published online: 19 MAY
2015, DOI: 10.1002/ima.22127 (Impact Factor 1.301)
2. Praisline Jasmi R, Perumal B, Pallikonda Rajasekaran M “Comparison of
Medical Image Compression using DWT Algorithm and Neural Network
Techniques” AENSI Journal Advances in Natural and Applied Sciences
Vol. 8, No. 19, pp:1-9, November 2014.
3. Perumal, B, M.Pallikonda Rajasekaran “Compression Techniques for
Medical Images Using SPIHT ”Applied Mechanics and Materials
(Volume 626), pp-87-94, 2014, ISSN: 1662-7482 (Scopus Index)
International Journal Communicated:
1. Perumal, B, M.Pallikonda Rajasekaran "A Hybrid Approach Using
Fractal and Neural Network Radial Basis Function for Efficient
Compression of Multi Modal Medical Images" in International Journal of
Imaging Systems and Technology has been communicated
147
International Conference Published
1. Mr.B.Perumal, Dr.M.Pallikonda Rajasekaran “A Hybrid Discrete
Wavelet Transform with Neural Network Back Propagation Approach for
Efficient medical Image Compression “International Conference on
Emerging Trends in Engineering, Technology and Science (ICETETS-
2016) at Kings College of Engineering on 24-26th February 2016
INDEXED IEEE EXPLORE
2. Perumal B, Praisline Jasmi R, Pallikonda Rajasekaran M “Comparison of
Image Compression Techniques using Huffman Coding, DWT and Fractal
Algorithm” attended 2015 International conference on Computer
Communication and Informatics connducted at Sri Shakthi College of
Engineering on 08th – 10th January 2015. INDEXED IEEE EXPLORE
3. Chithra, K., B. Perumal, M. Pallikonda Rajasekaran, and T. Arun Prasath.
"A quantitative assesment of image compression parameters and its
algorithm." In Communication Technologies (GCCT), 2015 Global
Conference on, pp. 294-296. IEEE, 2015. INDEXED IEEE EXPLORE
4. Praisline Jasmi R, Perumal B, Pallikonda Rajasekaran M “Comparison of
Medical Image Compression using DWT Algorithm and Neural Network
Techniques” attended 2015 International conference on Electrical,
Electronics, Instrumentation and Computer communication (E2IC2) 2014 at
Karpagam College of Engineering on 12th – 13th December 2014.
5. Perumal B, Pallikonda Rajasekaran M, and Duraiyarasan S, “Efficient
Image Compression Techniques for PET and MR brain images ”, IEEE
Sponsored Fourth International Conference on Recent Trends in
Information Technology, MIT, Chennai, April 10-12, 2014.
148
6. Perumal B, Pallikonda Rajasekaran M, “Compression Techniques for
Medical images Using SPIHT”, International Conference on Energy
Efficient Technologies for Sustainability (ICEETS’14), St. Xaviers
Catholic College of Engineering, Nagercoil, April 7-9, 2014
149
CURRICULUM VITAE
Mr. B. Perumal was born at Bodinayakanur, India in 1980. He graduated in
Electronics and Communication Engineering from Madurai Kamaraj University
and post graduated in Digital Communication and Network Engineering in 2006
from Anna University, Chennai, India. He is doing Ph.D (Medical Image
Compression) in Kalasalingam University, Krishnankoil. Now he is working as
Assistant Professor in the Department of Electronics and Communication
Engineering, Kalasalingam University. He has published 3 International Journals
published 6 papers in International Level Conferences and 15 National level
Conferences. His research interests include Mobile Computing, Wireless Sensor
Networks, Cloud computing and Bio-medical Instrumentation and Medical Image
Compression.