Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
854 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
EVALUATION OF PARAMETERS OF
IMAGE SEGMENTATION
ALGORITHMS-JSEG & ANN
Amritpal Kaura
Mrs. Amandeep Kaurb
M.Tech ECEa Head of Departmentb
Global Institute of Engineering and Technology Global Institute of Engineering and Technology
Amritsar, India Amritsar, India
Abstract: In this paper image segmentation using JSEG and Artificial neural networks is proposed. JSEG is
applied to provide characterization as well as classification. ANN is used for pattern recognition. Several
parameters are introduced to calculate and measure the performance of the system. These parameters
include Mean square error, Peak signal to noise ratio (PSNR), Computational Time. All these parameters are
evaluated and compare with both segmentation methods- ANN and JSEG. Moreover, to check the
performance of ANN, several parameters are calculated like Epochs, Time, Gradient, learning rate. By
comparing all these parameters, we can evaluate the process of segmentation for natural scene images. All
simulations are performed in Matlab using NNT (Neural Network Toolbox).
Keywords: JSEG, Artificial neural networks, PSNR, MSE, Computational Time
I.INTRODUCTION
Image segmentation is of use in several applications. It can spot the section of interest in a scene. Image
segmentation algorithms can be categorized into different forms. These forms include region based
segmentation, data clustering and edge based segmentation. In region based segmentation includes
seeded, unseeded and JSEG algorithm. These algorithms enlarge each region pixel by pixel based on their
pixel value or quantized value. In data clustering, the concept is based on whole image and segmentation
is performed depending upon the distance between each data. Edge based segmentation is based on the
points at which the image sharpness increases rapidly. These points are called edges. Lot of image
segmentation algorithm has been existed in literature. Chao Wang et.al [1] proposed a multi scale
segmentation method of oil spills in Synthetic Aperture Radar (SAR) images based on JSEG and spectral
clustering. K.Madhu et.al [2] had introduced Multi-class image semantics segmentation (MCISS) that had
many applications such as image editing and content based image retrieval. Luciano C.Lulio et.al[3] had
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
855 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
introduced image processing techniques in computer vision to a agricultural mobile robots used for
trajectory navigation problems as well as localization matters. S.N.Sulaiman[4] presented a new
clustering algorithm called Adaptive Fuzzy K means (AFKM) clustering for image segmentation which
might be applied on general images or specific images (medical and microscopic images).Yong Gang
Wang [5] presented a new color image segmentation method that combined directional operators and the
JSEG algorithm. J.J Charles et.al [6] investigated an evaluation measure of image segmentation for
classification of organic materials obtained from rocks and drill cuttings. Organic materials obtained from
rocks and drill cuttings involved finding multiple objects in the image. J.CHEN et.al [7] proposed a new
approach for image segmentation that was based on low level features for color and texture. It was aimed
at segmentation of natural scenes, in which the color and texture of each segment did not typically
exhibited uniform statistical characteristics. Good and wide range of overviews of clustering algorithms
can be found in literature [8],[9],[10],[11],[12],[13].
II. ARTIFICIAL NEURAL NETWORK
Artificial neural networks are models inspired by human nervous system. They bear a resemblance to our
brain in their working. The basic processing elements of neural networks are called neurons [14]. They
are parallel systems competent of resolving paradigms that linear systems cannot do. Neural networks
works by guidance and need not to be reprogrammed. Basically, an artificial neural network is a system
that receives an input, processes it and provides an output. There are some parameters related with ANN
which we use in this paper. These parameters are discussed as:-
i) Epochs: - Epochs are number of iterations for training of neural network. Number of epochs varies
from one image to another. Epochs are number of samples through which the process of providing the
network with input and updating the network’s weight is carried out. On the whole many epochs are
required to train the network.
ii)Training time:- During training time, the progress is continuously updated in the training window.
Training time is time taken by network to train. Number of validation checks involved in time of training
and used to cease the training. The number of validation checks represents the number of successive
iterations that the validation performance fails to decrease. If the number reaches 6(default value), the
training will stop.
iii) Performance: - The performance plot shows the value of performance function versus iteration
number. It specifies how successfully the system is trained. The performance function is usually selected
to be the sum of square of network and system errors on the training set.
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
856 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
iv) Gradient: - The magnitude of gradient are used to terminate the training. If the magnitude of gradient
is less than 1e-5, the training will stop. The gradient will turn out to be very small as the training reaches
the minimum of the training.
v) Learning rate: - The rate at which the network learns is called learning rate. It is a training factor that
controls the range and size of weight and bias changes throughout the learning process. Learning in neural
networks is known as training. Too low learning rate makes performance of the network very slowly and
on the contrary, too high learning rate makes weights diverge which leads to no learning at all. Learning
rate should be neither too low nor too high for proper operation.
III. JSEG
JSEG stands for j value segmentation. It is the very commanding scheme to test the homogeneity of the
image with given color reliability pattern and is reasonably proficient in computational provisions. JSEG
is a bottom up approach and make use of spatial segmentation. JSEG is a conventional approach for
categorization. JSEG is based on the perception of region growing. It is robust system of segmenting
natural images. The JSEG algorithm simplifies color and quality of images. There are certain parameters
that need to be defined:
i) Mean Square Error (MSE):- The error that mainly occurs in training of neural networks is mean
square error (mse). Mse is a complex network performance task. It measures the performance of network
according to mean of squared errors. It is the error metrics used to compare image segmentation quality.
The MSE measures the cumulative squared error between the reconstructed and original image. The
lower the value of MSE, the lower is the error and betters the quality of segmentation. MSE is calculated
as:
MSE= ∑ [ I1(m,n) – I2(m,n)]2 Eq.1
M, N M*N
In Eq.1 M, N is number of rows and columns in input images I1 and I2 respectively.
ii) PSNR: - PSNR stands for peak signal to noise ratio. It is measured in decibels. This ratio is often
measured as quality measurement between the original and reconstructed or compressed image. It
represents the measure of the peak error. The higher PSNR, the better is the quality of the reconstructed or
compressed image. In order to calculate PSNR, MSE is calculated first.
PSNR= 10 log10 (R2)/ MSE Eq.2
R is maximum fluctuation in input image data type. R is1 for double precision data type and 255 for 8 bit
unsigned data type.
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
857 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
iii) Computation Time: - Computational time is the length of time required to perform a computational
process. Its value varies from image to image and also has different values for JSEG and artificial neural
networks. As the amount of data increases, the computation time for Matlab code increases significantly.
IV. RESULTS
Figure 1: Segmentation results for image 1
Figure 2: Segmentation results for image 2
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
858 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Figure 3: Segmentation results for image 3
Figure 4: Segmentation results for image 4
Figure 5: Segmentation results for image 5
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
859 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Figure 6: Segmentation results for image 6
Figure 7: Segmentation results for image 7
Figure 8: Segmentation results for image 8
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
860 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
TABLE 1. PERFORMANCE PARAMETERS FOR ANN
Image Epoch Time Performance Gradient Learning Note Training
1 124 04 0.0920 0.0513 4.241 0.99812
2 124 01 0.0940 0.0538 4.241 0.99819
3 128 01 0.0889 0.0415 5.155 0.99781
4 128 01 0.0905 0.0503 5.155 0.99776
5 118 01 0.0946 0.0498 3.164 0.99643
6 126 01 0.0910 0.0427 4.676 0.9979
7 129 01 0.0944 0.0506 5.412 0.99744
8 121 01 0.0959 0.0440 3.663 0.99822
TABLE 2. COMPARATIVE STUDY OF JSEG AND ANN
Image Computational Time PSNR
MSE
Sr. No ANN JSEG ANN JSEG ANN JSEG
1 16 4 50 45 0.10 0.50
2 3.8 2.1 53 50 0.20 0.60
3 3.7 2.25 52 49 0.30 0.58
4 3.9 1.80 53 50 0.25 0.55
5 3.4 1.70 54 50 0.22 0.50
6 4.1 1.75 53 50 0.22 0.60
7 4.0 2.30 52 49 0.28 0.54
8 3.5 1.90 52 50 0.23 0.52
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
861 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
V. GRAPHS
Figure 9: Computation time with improved JSEG for image 1
Figure 10: Computation time with ANN for image 1
10
0.5
1
1.5
2
2.5
3
3.5
4
Computation Time with Improved JSEG
10
2
4
6
8
10
12
14
16
Computaion time with ANN
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
862 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Figure 11: Comparison of Total Computation time with ANN and JSEG for image 1
Figure 12: Comparison of MSE with ANN and JSEG for image 1
1 20
2
4
6
8
10
12
14
16
Total Computation Time
1 2 30
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
MSE
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
863 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Figure 13: Comparison of PSNR (peak signal to noise ratio) with ANN and JSEG for image 1
VI. CONCLUSION AND FUTURE SCOPE
The intention of image segmentation is to categorize the objects in images. Segmentation is a tough
dilemma and comparison with other methods is even tougher. By comparing performance of JSEG and
ANN, it has been seen that JSEG algorithm shows better performance in terms of computation time and
JSEG algorithm takes less time for computations as compared to ANN. Less computation time leads to
faster operation of the system. But, ANN shows better performance in terms of PSNR and MSE. PSNR is
high in all cases in contrast with JSEG. PSNR should be high for enhanced output and correlated with
quality of images. MSE should be as less as possible in order to have error free operation and it is very
less with ANN in all cases. In crux, ANN shows better results in terms of PSNR and MSE while JSEG
shows better performance in terms of computation time. These two algorithms can be used to segment
natural scene images without any pre-processing and post-processing techniques. Experiments show that
the proposed method shows good results. Future work will be concentrate to eliminate the limitations of
ANN in terms of computation time as it takes long time for computation.
1 2 30
10
20
30
40
50
PSNR
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
864 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
REFERENCES
[1] Chao Wang, Li-Zhong Xu, Xin Wang and Feng-Chen Huang, “A Multi-scale Segmentation Method of
Oil Spills in SAR Images Based on JSEG and Spectral Clustering”, International Journal of Signal Processing, Image Processing and Pattern Recognition Vol.7, No.1 (2014), pp.425-432.
[2] K. Madhu and R.I. Minu, “ Image segmentation using improved JSEG”, IEEE Proceedings of the 2013 International Conference on Pattern Recognition, Informatics and Mobile Engineering, February 21-
22,page 37-42.
[3] Luciano C.Lulio, Mario L. Tronco, Arthur J. V. Porto, “Cognitive merged Statistical Pattern
Recognition Method for Image Processing in Mobile Robot Navigation”, IEEE Proceedings of 2012
Brazilian Robotics Symposium and Latin American Robotics Symposium, pg 279-283.
[4] S. N. Sulaiman and N. A. M. Isa, “Adaptive Fuzzy-K-means Clustering Algorithm for Image
Segmentation ” , IEEE Trans. on Consumer Electronics, vol. 56, no. 4, pp. 2661-2668, 2010.
[5] Yong Gang Wang, J. Yang, and Y.-C. Chang, “Color-texture image segmentation by integrating
directional operators into jseg method”,Pattern Recogn. Lett., vol. 27, no. 16, pp. 1983–1990, 2006.
[6] J.J. Charles, L.I. Kuncheva, B. Wells, I.S. Lim, “ An evaluation measure of image segmentation based on object centers”, in: Proc. of the International Conference on Image Analysis and Recognition,
Portugal, September 18–20, 2006.
[7] Junqing Chen, Thrasyvoulos N. Pappas, Aleksandra Mojsilovic and Bernice E. Rogowitz, “Adaptive
Perceptual Color-Texture Image Segmentation”,IEEE TRANSACTIONS ON IMAGE PROCESSING,
VOL. 14, NO. 10, OCTOBER 2005.
[8] A.K. Jain and R.C. Dubes, “Algorithms for Clustering”,Englewood Cliffs, N.J.: Prentice Hall, 1988.
[9] R.O. Duda and P.E. Hart, Pattern Classification and Scene Analysis. New York: Wiley, 1973.
[10] J. Bezdek, “Pattern Recognition with Fuzzy Objective Function Algorithms”,New York: Plenum,
1981.
[11] J. Hartigan, Clustering Algorithms. New York: Wiley, 1975.
[12] J. Tou and R. Gonzalez, Pattern Recognition Principles. Reading, Mass.: Addison-Wesley, 1974.
[13] E. Ruspini, “A New Approach to Clustering”,Information Control, vol. 15, no. 1, pp. 22-32, July
1969.
[14]. M.Schuster and K. K. Paliwal,“Bidirectional recurrent neural networks”, IEEE Transactions on Signal
Processing,vol 45, 2673–2681 (1997)
International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 8, August 2014
865 ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Authors
Amritpal kaur is pursuing M. Tech in Electronics and Communication Engineering from
Global Institute of Engineering & Technology, Amritsar. She did her B.Tech degree in Electronics and
Communication Engineering from Sai Institute of Engineering & Technology, Amritsar, Punjab, India. She has five
papers published in reputed international journals and one paper in national conference.
Mrs.Amandeep Kaur is Assistant Professor and Head of Department in Electronics
and Communication Engineering in Global Institute of Engineering & Technology, Amritsar, India. She has five
papers published in reputed international journals and one paper in national conference.