Upload
vucong
View
214
Download
1
Embed Size (px)
Citation preview
CHAPTER – 1
INTRODUCTIONAge estimation is the determination of a person's age based on biometric features. The
determination of the age of a person from a digital photography is an intriguing problem;
It involves understanding of the human aging process. People cannot freely control aging
variation; the collection of sufficient data for age estimation is extremely laborious.
Estimating of exact age is one of the most difficult problems even for human being.
Therefore, most of the researchers who are working on age estimation are trying to get
the results in certain age ranges. The experimented age ranges are still considered to be
wide and in some cases exceed 10 years while in other cases reach 15 or 20 years.
One of the main problems to reduce the size of the age ranges is how correct and
comprehensive the extracted features from the face are. Some researchers use 20, 22,
35,or 68 features, and the accuracy of the results vary depending on the extracted features
and the used approach for age estimation. There are some open databases used for testing
age estimation systems such as FG-NET [5][6[7] and Morph [8].These datasets contain
photos and ages of the people and there are usually ages from 1 year to 70 years.
Many attempts towards age and sex estimation are tried and most of them give
results for wide ranges of ages, or classify the ages in categories such as child, young,
youth and old. The problem of having an appropriate approach for age estimation for
getting more specific categories of age ranges is still a challenging problem. Thus, we
focus our research on more specific age ranges. Actually this problem is also a human
problem where many people miss estimate the human ages. To achieve our research goal,
we have to find a good database that we can use to test and train our proposed approach,
also we have to construct a proper ANN to model our problem. Developing such kind of
systems might help in many security purposes or in cases of having disabled people
(dump and deaf people). Because our main goal is age estimation and not face
recognition, we care only about images of front image, with face free from classes or
beard.
1
As mentioned before, most of the researchers categorize the ages into four classes,
childhood, young, youth and old. This classification is more or less as our classification
in the main classification stage. In our system we go further by classifying each of the
main categories (each class) into two classes which we call secondary classification
stage. The classification in the secondary stage is not partitioned equally, instead the age
partitions are based on some changes in the facial features. For example the face features
for people who are between 13 to 25 are very close. Also the ages from 36 to 45 are
almost negligible. In some cases we make the range a little bit wide because in many
cases it is really difficult to categorize the age in a smaller range.
In this present scenario, image plays vital role in every aspect of business such as
business images, satellite images, medical images and so on. If we analysis these data,
which can reveal useful information to the human users. But, unfortunately there are
certain difficulties to gather those data in a right way. Due to incomplete data, the
information gathered is not processed further for any conclusion. In another end, Image
retrieval is the fast growing and challenging research area with regard to both still and
moving images.
Color image based segmentation aims at searching image databases for specific
image that are similar to a given query image. It also focuses at developing new
techniques that support effective searching and browsing of large digital image libraries
based on automatically derived imagery features. It is a rapidly expanding research area
situated at the intersection of databases, information retrieval, and computer vision. This
method focuses on Image ‘features’ to enable the query and have been the recent focus of
studies of image databases.
The features further can be classified as low-level and high-level features. Users
can query example images based on these features such as texture, color, shape, region
and others. By similarity comparison the target image from the image repository is
retrieved. Meanwhile, the next important phase today is focused on clustering techniques.
Clustering algorithms can offer superior organization of multidimensional data for
effective retrieval. Clustering algorithms allow a nearest neighbor search to be efficiently
performed. Hence, the images are rapidly gaining more attention among the researchers
2
in the field of data mining, information retrieval and multimedia databases. Spatial
Databases is the one of the concepts which plays a major role in Multimedia System.
Researches can extract semantically meaningful information from image data are
increasingly in demand.
Image are generated at increasing rate by sources such as military reconnaissance
flights; defense and civilian satellites; fingerprinting devices and criminal investigation;
scientific and biomedical imaging; geographic and weather information systems; stock
photo databases for electronic publishing and news Agency; fabric and fashion design; art
galleries and museum management; architectural and engineering design; and WWW
search engines. Most of the existing image management systems are based on the verbal
descriptions to enable their mining. A key-word description of the image content, created
by some user on input, in addition to a pointer to the image data is the base of this
system. Image mining is then based on standard mining. However, verbal descriptions are
almost always inadequate, error prone and time consuming. A more efficient approach is
gathered when image example is given by the user on input to the mining process.
Automatically generate matching is required then for an efficient age and gender
recognit.
The basic idea is to extract characteristic features similar to that of object
recognition schemes. After matching, images are ordered with respect to the query image
according to their similarity measures and displayed for viewing. In this work, we present
a framework f
or considering the influence of this distance function on color to identify the age.
This framework assesses a system’s quality from the viewpoints of users; it provides a
basic set of attributes to characterize the ultimate utility of systems. Then we analyze
examples of mining by color and present some conclusions.
Images present special characteristics due to the richness of the data that an image
can show. Effective evaluation of the results of image by content requires that the user
point of view is used on the performance parameters. Comparison among different
mining by similarity systems is particularly challenging owing to the great variety of
methods implemented to represent likeness and the dependence that the results present of
3
the used image set. Other obstacle is the lag of parameters for comparing experimental
performance. In this project we implement an evaluation framework for comparing the
influence of the distance function on image mining by color and also a way to mine an
image from its name. Experiments with color similarity mining by quantization on color
space and measures of likeness between a sample and the image results have been carried
out. Important aspects of this type of mining are also described.
Images are one of the most important binary multimedia data available in the
system. The images are cluttered in various drives in the system. As far as Document data
are concerned, they are indexed by windows indexing search. Therefore when a search
command is generated, the documents are retrieved quickly. The content based image
retrieval system which is an important alternative to document searching, searches a
given image in set of images. This search is one to one search, therefore is time
consuming. The clustering and indexing algorithm which groups the similar images
together such that any matching can retrieve the entire set of images rather than requiring
searching every image independently.
We propose a methodology based on Neural networks to estimate human ages
using face features. Due to the difficulty of estimating the exact age, we developed our
system to estimate the age to be within certain ranges. In the first stage, the age is
classified into four categories which distinguish the person oldness in terms of age. The
four categories are child, young, youth and old. In the second stage of the process we
classify each age category into two more specific ranges. The uniqueness about our
research project is that most of the previous research work do not consider the fine tuning
of age as we are presenting in our research. Our proposed approach has been developed,
tested and trained using the EasyNN tool.
1.1 COMPARISON OF IMAGE MINING WITH OTHER
TECHNIQUESImage mining normally deals with the extraction of implicit knowledge, image
data relationship, or other patters not explicitly stored from the low-level computer vision
and image processing techniques. i.e. the focus of image mining is the in the extraction
of patterns from a large collection of images, the focus of computer vision and image
4
processing techniques is in understanding or extracting specific features from a single
image.
Figure 1: Image Mining Process
Figure 1 shows the image mining process. The images from an image database are
first preprocessed to improve their quality. These images then undergo various
transformations and feature extraction to generate the important features from the images.
With the generated features, mining can be carried out using data mining techniques to
discover significant patterns. The resulting patterns are evaluated and interpreted to
obtain the final knowledge, which can be applied to applications [1].
The field of image retrieval has been an active research area for several decades
and has been paid more and more attention in recent years as a result of the dramatic and
fast increase in the volume of digital images. The development of Internet not only cause
an explosively growing volume of digital images, but also give people more ways to get
those images.
There were two approaches to content-based image retrieval initially. The first
one is based on attribute representation proposed by database researchers where image
contents are defined as a set of attributes which are extracted manually and are
maintained within the framework of conventional database management systems. Queries
are specified using these attributes. This obviously involves high-level of image
5
Images in the DatabasesPreprocess the Image ContentMining the Collected DataInterpretation & EvaluationCreate Knowledge
abstraction. The second approach which was presented by image interpretation
researchers depends on an integrated feature-extraction object-recognition subsystem to
overcome the limitations of attribute-based retrieval. This system automates the feature-
extraction and object- recognition tasks that occur when an image is inserted into the
database. These automated approaches to object recognition are computationally
expensive, difficult and tend to be domain specific. There are two major categories of
features. One is basic which is concerned with extracting boundaries of the image and the
other one is logical which defines the image at various levels of details. Regardless of
which approach is used, the retrieval in content-based image retrieval is done by color,
texture, sketch, shape, volume, spatial constraints, browsing, objective attributes,
subjective attributes, motion, text and domain concepts[2].
Content-based image retrieval has become a prominent research topic in recent
years. Research interest in this field has escalated because of the proliferation of video
and image data in digital form. The goal in image retrieval is to search through a database
to find images that are perceptually similar to a query image. An ideal image retrieval
engine is one that can completely comprehend a given image, i.e., to identify the various
objects present in the image and their properties. Given the state of the art of research in
the image analysis community, such an ideal retrieval system is far from being reality.
Moreover retrieval based on human annotation is to no avail, because of the size of the
video and image databases and the varying interpretations that different humans can
attach to an image.
In a practical scenario, like the Internet, the number of images can be of
the order of millions and is ever growing. Even if the time required to compare two
images is very short, the cumulative time needed to compare the query image with all the
database images is rather long and is probably longer than the time an average user wants
to wait. We solve this problem by grouping or clustering the images according to their
similarity beforehand, so that at the time of the query, it is not necessary to perform an
exhaustive comparison with all the images in the database. The clustering is performed
based on visual features extracted automatically from the images. Performance evaluation
has been a challenging issue in the field of content-based retrieval, primarily because of
6
the difficulty associated with calculating quantitative measures to evaluate the quality of
retrieval. The precision and recall measures have been frequently used by many
researchers to evaluate the performance of retrieval algorithms. In this paper we
introduce a quantitative method to evaluate the retrieval accuracy of clustering
algorithms. Our goal is not to subjectively evaluate the quality of retrieval, but to
quantitatively compare the performance of Retrieve with and without clustering.
Content Based Image Retrieval is the retrieval of images based on visual features
such as color and texture. Reasons for its development are that in many large image
databases, traditional methods of image indexing have proven to be insufficient,
laborious, and extremely time consuming. These old methods of image indexing, ranging
from storing an image in the database and associating it with a keyword or number, to
associating it with a categorized description, have become obsolete. This is not in CBIR.
In CBIR, each image that is stored in the database has its features extracted and compared
to the features of the query image. It involves two steps:
Feature Extraction: The first step in the process is extracting image features to a
distinguishable extent.
Matching: The second step involves matching these features to yield a result that
is visually similar.
The importance of an effective technique in searching and retrieving images from
the huge collection cannot be overemphasized. One approach for indexing and retrieving
image data is using manual text annotations.
Advances in image acquisition and storage technology have led to tremendous
growth in significantly large and detailed image databases. These images, if analyzed,
can reveal useful information to the human users. Image mining deals with the extraction
of implicit knowledge, image data relationship, or other patterns not explicitly stored in
the images. Image mining is more than just an extension of data mining to image domain.
It is an interdisciplinary endeavor that draws upon expertise in computer vision, image
processing, image retrieval, data mining, machine learning, database, and artificial
intelligence. Despite the development of many applications and algorithms in the
individual research fields cited above, research in image is still in its infancy. In this
7
paper, we will examine the research issues in image mining, current developments in
image mining, particularly, image mining frameworks, and state-of-the-art techniques
and systems. The annotations can then be used to search images indirectly. But there are
several problems with this approach. First, it is very difficult to describe the contents of
an image using only a few keywords. Second, the manual annotation process is very
subjective, ambiguous, and incomplete.
Those problems have created great demands for automatic and effective
techniques for content-based image retrieval (CBIR) systems. Most CBIR systems use
low-level image features such as color, texture, shape, edge, etc., for image indexing and
retrieval. It’s because the low-level features can be computed automatically. Content
Based Image Retrieval (CBIR) has emerged during the last several years as a powerful
tool to efficiently retrieve images visually similar to a query image. The main idea is to
represent each image as a feature vector and to measure the similarity between images
with distance between their corresponding feature vectors according to some metric.
Finding the correct features to represent images with, as well as the similarity metric that
groups’ visually similar images together, are important steps in the construction of any
CBIR system.
The efficiency of different clustering approaches for selecting a set of exemplar
images, to present in the context of a semantic concept. We evaluate these approaches
with different images, and comparing the image based on clustering. Affinity Propagation
is effective in selecting exemplars that match the top search images but at high
computational cost.
8
CHAPTER – 2
LITERATURE REVIEWImage mining is rapidly gaining more attention among the researchers in the field of data
mining, information retrieval and multimedia databases. In this section, briefly the Image
Mining work in the literature is reviewed.
2.1 IMAGE MININGThe need for image mining in view of the rapidly growing amounts of image data. We
have pointed out the unique characteristics of image databases that bring with it a whole
new set of challenging and interesting research issues to be resolved. In addition, we have
also examined two frameworks for image mining: function-driven and information-
driven image mining frameworks. They also discussed techniques that are frequently
used in the early works in image mining, namely, object recognition, image retrieval,
image indexing, image classification and clustering, association rule mining and neural
network .
The main objective of the image mining is to remove the data loss and extracting
the meaningful information to the human expected needs. The images are preprocessed
with various techniques and the texture calculation is highly focused. Here, images are
clustered based on RGB Components, Texture values and Fuzzy C mean algorithm.
Entropy is used to compare the images with some threshold constraints. This application
can be used in future to classify the medical images in order to diagnose the right disease
verified earlier.
Image mining is the advanced field of Data mining technique and it has a great
challenge to solve the problems of various systems. The main objective of the image
mining is to remove the data loss and extracting the meaningful information to the human
expected needs. Here, we have furnished some of the techniques, which can be applied
for image retrieval system in future.
In this system, the color based and texture based image retrieval yields high
accuracy. That is, it retrieves the most matching images from the collection of the
images, with respect to the query image.
9
In future, it is to implement a voice recognition system to give our desired image
as a keyword to find the related images from the database.
2.2 K- MEANS ALGORITHM The K-means algorithm gradually finds k clusters with segmentation for each cluster. Our
algorithm does not require the users to select and try various parameters combinations in
order to get the desired output.
The benefit of the algorithm is to find similarity structures (segments) within
clusters. All of these look nice from theoretical point of view. However from practical
point of view, there is still some room for improvement for running time of the clustering
algorithm. This could perhaps be accomplished by using some appropriate data structure.
2.3 TEXT BASED APPROACHEarly Techniques were not generally based on visual features but on the textual
annotation of images. In other words, images were first annotated with text and then
search using text based approach from traditional database management system. Text
based image retrieval system uses traditional database techniques to manage images.
Through text description, images can be organized by topical or semantic hierarchies to
facilate easy navigation and browsing base on standard Boolean queries. However since
automatically generating descriptive text for wide spectrum of images is not feasible,
most text based image retrieval system requires manual annotation of images. Obviously,
annotating images manually is a cumbersome as an expensive task for large image
databases, and is often subjective, context sensitive and incomplete. As a result it is
difficult for traditional text based method to support variety of task dependant queries.
It is said that one image is worth a thousand words. Visual information accounts
for about 90% of the total information content that a person acquires from the
environment through his sensory systems. This reflects the fact that human being relies
heavily on his highly developed visual system compared with other sensory pathways.
The external optical signal is perceived by eyes, and then converted into neural signal;
the corresponding neural subsystem specialized for visual system is specially organized
to detect subtle image features and perform high-level processing, which is further
processed to generate object entities and concepts. The anatomy of the visual system
10
explains from the structure aspect why visual information is so important to human
cognition. The cognitive functions that such a system must support include the capability
to distinguish among objects, their positions in space, motion, sizes, shapes, and surface
texture. Some of these primitives can be used as descriptors of image content in machine
vision research.
There are two formats, in which visual information can be recorded and presented
– static image, and motion picture, or video. Image is the major focus of research interest
in digital image processing and image understanding. Although a relatively recent
development, computerized digital image processing has attracted much attention and
shed lights to a broad range of existing and potential applications. This is directly caused
by rapid accumulation of image data, a consequence of exponential increases of digital
storage capacity and computer processing power. There are several major types of digital
images depending on the elemental constituents that convey the image content. Images
can take the form of:
1. Printed text and manuscript. Some examples of the kind are micro-films of old
text documents, photograph of handwriting.
2. Line sketch, including diagrams, simple line graphs
3. Halftones. Images are represented by a grid of dots of variable sizes and
shapes.
4. Continuous tone. Photographic images that use smooth and subtle tones.
5. Mixture of above.
Feature extraction plays an important role in content-based image retrieval to
support for efficient and fast retrieval of similar images from image databases.
Significant features must first be extracted from image data. Retrieving images by their
content, as opposed to external features, has become an important operation.
A fundamental ingredient for content based image retrieval is the technique used
for comparing images. There are two general methods for image Comparison: intensity
based (color and texture) and geometry based (shape).
One of the most important features that make possible the recognition of images
by humans is color. Color is a property that depends on the reflection of light to the eye
11
and the processing of that information in the brain. We use color everyday to tell the
difference between objects, places, and the time of day. Usually colors are defined in
three dimensional color spaces. These could either be RGB (Red, Green, and Blue), HSV
(Hue, Saturation, and Value) or HSB (Hue, Saturation, and Brightness). The last two are
dependent on the human perception of hue, saturation, and brightness. Most image
formats such as JPEG, BMP, GIF, use the RGB color space to store information. The
RGB color space is defined as a unit cube with red, green, and blue axes. Thus, a vector
with three co-ordinates represents the color in this space. When all three coordinates are
set to zero the color perceived is black. When all three coordinates are set to the color
perceived is white. The other color spaces operate in a similar fashion but with a different
perception.
An image can be an array of pixel values stored in uncompressed bitmap digital
image format. In this format, each value represents the color intensity at discrete points or
pixels. A well-known example of this is Microsoft's BMP format. Although BMP format
allows for pixel packing and run-length encoding to achieve certain level of compression,
its uncompressed version is more popular.
Popular Internet standard image formats see more extensive image
transformation and compression, such as GIF and JPEG. The GIF standard defines a
color degeneration process, which maps the colors in an image into no more than 256
new colors.
Specially defined signature file stores specifically extract images features in
numeric or Boolean format, indicating the presence (or non-presence) and strength of the
features. This is a very compact representation of image that is targeted for fast retrieval
instead of display, archival, etc. This allows for easy indexing and fast search for
matching features of the query. An example of this can be found in image coding method
using vector quantization (VQ), in which image blocks are coded according to a carefully
chosen codebook. If the image blocks are similar to each other or the images in a set bear
significant similarity, higher compression ratio can usually be achieved than the general-
purpose compression algorithms such as GIF, JPEG.
12
Textual annotation can also be thought of as an instantiation of mental image,
and sometimes, the descriptors can be coded by a predefined convention, or a thesaurus.
The fact that two visually different images can convey the same concept and different
concepts may present in images that share many similar optimal properties brings about a
gap between image retrieval by content, and retrieval by concept.
Image mining requires that images be retrieved according to some requirement
specifications. The requirement specifications can be classified into three levels of
increasing complexity (a) Level 1 comprises image retrieval by primitive features such as
color, texture, shape or the spatial location of image elements. Examples of such queries
are “Retrieve the images with long thin red objects in the top right-hand corner” and
“Retrieve the images containing blue stars arranged in a ring” (b) Level 2 comprises
image retrieval by derived or logical features like objects of a given type or individual
objects or persons. Examples include “Retrieve images of round table” and “Retrieve
images of Jimmy” (c) Level 3 comprises image retrieval by abstract attributes, involving
a significant amount of high-level reasoning about the meaning or purpose of the objects
or scenes depicted.
For example, we can have queries such as “Retrieve the images of football
match” and “Retrieve the images depicting happiness”. Here are three fundamental bases
in content-based image retrieval, namely, visual information extraction, image indexing
and retrieval system application. Many techniques have been developed in this direction,
and many image retrieval systems, both research and commercial, have been built. In the
area of commercial systems, IBM’s QBIC system is probably the best known of all image
content retrieval systems. It offers retrieval by any combination of color, texture or shape,
as well as text keyword [8].
Image mining is a promising field for research. Image mining research is still in
its infancy and many issues remain solved. Specifically, we believe that for image mining
research to progress to a new height, the following issues need to be looked at:
(a) Propose new representation schemes for visual patterns that are able to encode
sufficient contextual information to allow for meaningful extraction of useful visual
characteristics;
13
(b) Devise efficient content-based image indexing and retrieval techniques to
facilitate fast and effective access in large image repository;
(c) Design semantically powerful query languages for image databases;
(d) Explore new discovery techniques that take into account the unique
characteristics of image data;
(e) Incorporate new visualization techniques for the visualization of image
patterns.
Image clustering is usually performed in the early stages of the mining process.
Feature attributes that have received most attention for clustering are color, texture and
shape. Once the images have been clustered, a domain expert is needed to examine the
images of each cluster to label the abstract concepts denoted by the cluster.
Image classification and image clustering are the supervised and unsupervised
classification of images into groups respectively. In supervised classification, one is
provided with a collection of labeled (pre-classified) images, and the problem is to label
newly encountered, unlabeled images. Typically, the given labeled (training) images are
used to do the machine learning of the class description which in turn is used to label a
new image.
In image clustering, the problem is to group a given collection of unlabeled
images into meaningful clusters according to the image content without a priori
knowledge. The fundamental objective for carrying out image classification or clustering
in image mining is to acquire content information the users are interested in from the
image group label associated with the image.
Age Estimation Approach fall with two mainstreams.According to the first stream
the problem is treated as a standard classification problem, solved using standard
classifiers where age estimation is performed by assigning a set of facial features to an
age group. Within this context facial features used may be associated with the general
appearance of a face or may be associated to age-related features (e.g. wrinkles). As an
alternative age estimation approaches that rely on the modelling of the aging process
have been developed. In this section typical approaches described in the literature are
briefly presented. The aim of this review is not to present an exhaustive literature review
14
of the topic but rather to highlight the evolution of the topic. A more detailed presentation
of the related literature is presented by Ramanathan et al. (Ramanathan 2009) and Fu et
al. (Fu 2010).
One of the first attempts to develop facial age estimation algorithms was reported
by Kwon and Lobo (Kwon 1999), [12]. Kwon and Lobo use two main types of features:
Geometrical ratios calculated based on the distance and the size of certain facial
characteristics and an estimation of the amount of wrinkles detected by deformable
contours (snakes) in facial areas where wrinkles are usually encountered. Based on these
features Kwon and Lobo (Kwon 1999) classify faces into babies, adults and seniors based
on a computational theory for visual age classification from facial images. First, primary
features of the face, namely the eyes, nose, mouth, chin, and virtual top of the head, are
found. The research in age-estimation started in 1990s and up to now, many approaches
have been proposed. They typically consist of two main steps: image representation and
age prediction. For the image representation, the most common models are
Anthropometric model [8], Active Appearance Model (AAM) [5], aging pattern subspace
[6], aging manifolds [1], and patch-based model [7]. The final step for age estimation is
either the multiclass classification problem or the regression problem. In 1999, Kwon [8]
measured the changes of face in shapes, e.g. six geometric ratios of key features, to
classify faces into appropriate age groups. Drawing inspiration from this work,
Ramanathan [9], Dehshibi [10], later used the geometric ratios of facial features and
added information of texture, e.g. wrinkles, in their approaches. Although these
approaches achieved low Mean Absolute Errors (MAEs), they can only deal with young
ages when the shapes of faces vary largely. Moreover, because of the sensitivity to head
pose in the steps of computing geometric ratios in 2D face images, only frontal faces can
be used. Adopting the Active Appearance Models (AAMs) [11] approach, Lanitis et al.
[12], Khoa Luu et al. [5] used AAM features, which combine both shape and texture
information in their ageestimation studies. In 2009, using AAM features extracted from
image with 161 landmarks, Ricanek et al. [11] developed a multiethnic age-estimation
system that can deal with the race problem. Recently, based on the arguments that age
information is often encoded by local information, such as wrinkles around the eye
15
corners, other approaches are to divide face images into many sub-regions, extract
features from these regions, and then combine them together. Yan et al. [7] proposed to
use Spatially Flexible Patch (SFP) and Gaussian Mixture Model (GMM). B. Ni et al. [2]
developed a technique to extend the human aging image dataset by mining the web
resource and then used SFP for representing face images. Suo et al. [11] designed a
multiresolution hierarchical graphical face model for age estimation. LBP features and
Gabor features are also exploited in the work of Günay [5], and Gao [16]. Guo et al. [10],
in 2009, investigated the biologically inspired feature (BIF) derived from a feedforward
model of the primate visual object recognition pathway – HMAX model. The advantages
are that small translations.
The proposed approach to identify the age range algorithm that is free from
previous disadvantages. The proposed method classifies face images into one of four
well-ordered age groups range, which contains four key steps, pre-processing, facial
feature extraction, wrinkle analysis, age range identification.
16
CHAPTER – 3
PROBLEM IDENTIFICATIONMost of the time it is not possible for us to know the age and sex of the person. Most of
the methods are fail to implement this concept. Advances in image acquisition and
storage technology have led to tremendous growth in very large and detailed image
databases.
The system is mainly using supervised neural networks with back propagation
algorithm. The image is entered to the system, features are extracted, the image is
classified in one of the four main age classes, then a more specific age range class is
specified. We firstly classify age into four main age categories and each age category is
classified into two age ranges. We obtained our data to train and test our system from two
databases, FG-NET [7] and MORPH [9].The images in FG-NET are ready and their
features are already extracted. MORPH database have only images with some other
related information, but without extracted features. For this purpose we ought to extract
the features from the images obtained from MORPH database using am-markup tool.
Finally we train the system with Easy- NN tool based on the two datasets FG-NET and
MORPH.
In the color based image retrieval the RGB Color model is used. Color images
normally are in three dimensional. RGB color components are taken from each and every
image. Then, the mean values of Red, Green, and Blue components of target images are
calculated and stored in the database. Based on the RGB component mean values, the
images are clustered as Red, Green and Blue major component categories. These three
mean values for each image are stored and considered as features. Then the top ranked
images are re-grouped according to their texture features. In the texture-based approach
the parameters gathered are on the basis of statistical approach. Statistical features of grey
levels were one of the efficient methods to classify texture. The different texture
parameters like entropy, contrast, dissimilarity, homogeneity, standard deviation, mean,
and variance of both query image and target images are calculated. From the calculated
values the required image from the repository is extracted.
17
The Search efficiency of image retrieval relies on an efficient classification
scheme. Rather than searching an image over a huge collection of image database, if the
images are classified into a finite number of groups, then the search time is reduced by an
amount equal to the total number of groups. This plays a significant role when the present
day application falls into real time. These images, if analyzed, can reveal useful
information to the human users. Images deals with the extraction of implicit knowledge,
image data relationship, or other patterns not explicitly stored in the images. Image
mining is more than just an extension of data mining to image domain.
Most current CBIR systems work on a completely different principle. Fixed-
length real-valued multicomponent feature vectors typically characterize stored images,
each image having a value for every feature in the database. In this case, searching
consists of calculating the similarity between feature vectors from query and stored
images, a process of numerical computation.
Cluster-based retrieval of images by unsupervised learning (CLUE) was one such
method to tackle the semantic gap problem. The CLUE is built on a hypothesis that
images of the same semantics tend to be clustered [4]. It attempts to narrow the semantic
gap by retrieving image clusters based on not only the feature similarity of images to the
query, but also how images are similar to each other. But this is a general approach and
implementations seemed to be complex.
The current CBIR techniques assume certain mutual information between the
similarity measure and the semantics of the images. A typical CBIR system ranks target
images according to the similarities with respect to the query and neglects the similarities
between target images. The performance of a CBIR system is improved by including the
similarity information between target images. To achieve this new technique and for
improving user interaction with image retrieval systems by fully exploiting the similarity
information. Given a query, images in the database are firstly grouped using color
features i.e., color based image retrieval. In the color based image retrieval the RGB
model is used. Color images are in three dimensional, so RGB color components are
taken from each and every image. Then the average value of R, G and B values for both
query image and target images are calculated. These three average values for each image
18
are stored as features. By using these stored features the image from the repository is
retrieved with respect to the query image.
Then the top ranked images are re-grouped according to their texture features. In
the texture-based approach the parameters gathered are on the basis of statistical
approach. Statistical features of grey levels were one of the efficient methods to classify
texture. The different texture parameters like entropy, contrast, dissimilarity,
homogeneity, standard deviation, mean, and variance of both query image and target
images are calculated. From the calculated values the required image from the repository
is extracted.
19
CHAPTER – 4
PROPOSED METHODOLOGY4.1 C - MEANS CLUSTERINGDuring the process of clustering first we create the segmentation of the given image.
After performing the process of segmentation which is nothing but the process of
performing cluster. After cluster formation we find out the centre of each cluster and
grouping they based on the minimum distance. This process is continuing up to the
formation of cluster. If cluster formed on the basis of cluster then stop else continue the
process. This process play vital role for image retrieval system. The process of C-means
clustering as shown in figure 4.1
Figure 4.1: Dataflow Diagram of C means Clustering
At the initial phase in our project we create the clusters of the input image. We
take the input image create the number of cluster; in our project we create five clusters.
After creating the cluster of the image we find the center of each cluster. Centroid is
20
useful to find out the distance between all the clusters that are create from the input
image. Due to this we know the position of each cluster. After creating the cluster we
create the group of the cluster on the basis of minimum distance. Grouping of cluster is
important when we compare the input image and the directory image. The process of
cluster based image mining technique is based on two techniques
Mining By Color Content
4.1.1
MINING BY COLOR CONTENTIn our system, we are having two approaches for images retrieval that is mining
by color contents and mining by text. Images are mostly RGB, Gray Scale and Binary
types which are easily possible to segments as per their pixels values. RGB images are
segmented as per the color content, Gray Scale images are segmented as per their average
RGB pixels value and Binary Images as per their ON/OFF pixels status. Image mining
by color contents include search for images based on their segmented color contents. Let
see Dataflow Diagram for image mining technique based on their color contents as
follows
21
Figure 4.2 Data Flow Diagram for finding age and sex
In our project we take the image as an input image which is also called as query
image through which we get the resultant image. After selecting the input image we are
performing the process of segmentation which is done on the basis of color content.
When we perform the procedure of clustering on the input image we get the clustered
image. During this procedure the images in the directory are clustered on the basis of
color content. This is the important portion of our project. After the completion of
segmentation we compare the input image with the directory image. Then calculate the
deviation factor between the test image and the directory image. After calculate the
deviation factor we get the resultant image. After displaying the resultant images we
come to know the gender and age of the person in the photo. The dataflow diagram for
the process of mining by color content is shown in figure 4.2
Segmentation of images based on their color contents is a major area of research.
In this method we are using C-means clustering technique to clusterize an image based on
their RGB colors as per equation
Cn=∫1
n
∑i=1
M
¿ Pi−Pi+1∨¿………………. 4.1
Where
Cn = C means cluster.
n = No of Clusters.
Pi = Backward pixel
Pi+1= Forward Pixel
M= No of Pixels in an image.
In c-means clustering technique the pixels are scanned as per Raster Scan method
from left to right and top to bottom way. Whenever difference occurs with backward and
forward pixel we set it as OFF else we proceeds our scanning as per the equation
22
If |Pi-Pi+1| =0
Else
|Pi-Pi+1 |≠0
Pi+1=Pi
Pi=0 ….. 4.2
Deviation Factor Df is a difference between number of segments of directory
image and input test image. Deviation factor is a measure of similarity and difference
between two images which can be represented with an equation
Df =|Ctr−Cts| ….. 4.3
Where
Ctr = Directory Image Segments.
Cts = Input Test Image Segments.
An overall searching time for color based image is depended on number of color
segments in an image.
O (tc ) α n …… 4.4
Where O ( tc ) = searching time for color based image.
We use best fit searching technique to find out an image segment express with
O (C ) α O (n) …… 4.5
Where O (C ) = no of segments comparisons.
Complexity is a useful point in comparing software systems .This aspect is
normally obtained from the source code, but it is completely irrelevant for the user. The
mining result is a more important aspect for the user. The factor concerning to the mining
result are: the underlying color space used to represent the color features; the quantization
approach used; the number of bins on the histogram space (its dimension or digital color
23
resolution); the distance function used to represent the notion of nearness on the color
space (histogram representation); the fixed number of images to be retrieved and the
threshold used for matching similarity Several color spaces have been used for color
representation based on the perceptual concepts.
There is no agreement on which is the best choice. Anyway, its desirable
characteristics are completeness, uniformity, compactness, and user oriented.
Completeness means that it must include all perceptible different colors. Uniformity
means that the measured proximity among the colors must be directly related to the
psychological similarity among them. Compactness means that each color presents a
perceptual difference from the other colors. Color quantization transforms a continuous
tone picture into a discrete image. The digitalization process maps each component of a
continuous color signal into a series of limited number of (fewer) colors.
This process inevitably introduces distortion. The visible distortion is a subjective
and psychological notion. The questions are how to choose the colors to reproduce the
original (not necessarily colors that appear in the original image). A quantization
algorithm should distribute any visible distortion throughout the image so that none
stands out to be found particularly objectionable by an average human observer.
Empirical algorithms (as the popularity algorithms and the median-cut algorithms)
present cases where significant color shifts can be found. One of the numerical criteria
for color image quantization is to minimize the maximum variance between original pixel
color and the corresponding quantified color, which provides better results than empirical
algorithms. Another numerical criterion is to minimize the maximum discrepancy
between original and quantified pixel values. Recent works use adaptive quantifiers. The
basic strategy employed by these is a two-step approach.
The first step group original colors into clusters that are as small as possible. The
second step computes a quantified color for each cluster. This means that each image is
associated with two types of histograms in the mining process.
Mining in visual database is quite different from standard alphanumeric mining.
On current approaches, feature vectors per image are computed for evaluation distance
function on the feature space. Then this function is used to retrieve images from a given
24
set. Images with distance less than a predefined threshold or within a predefined number
are retrieved. These feature vectors facilitate mining by color, texture, geometric
properties, shape, volume, spatial constraints, etc.
We are going to implement the c means clustering algorithm.
Fuzzy c-means (FCM) is a method of clustering which allows one piece of data to
belong to two or more clusters.
With fuzzy c-means, the centroid of a cluster is computed as being the mean of
all points, weighted by their degree of belonging to the cluster.
By iteratively updating the cluster centers and the membership grades for each
data point, FCM iteratively moves the cluster centers to the "right” location within
a data set.
In our system we use c-means clustering to retrieve the image. The
process of clustering can be done with the help of raster scan image. The scanning
can be done from left to right and top to bottom. The process of clustering is
shown in figure 4.3
Figure 4.3: Example of C means clustering
25
4.1.2 IMAGE RETRIEVAL SYSTEMFollowing are the steps of project
Query Image
RGB Components processing
Cluster based on R,G,B component
Texture feature extraction
Similarity comparison
Target image selection
Pre-processing is the name used for operations on images at the lowest level of
abstraction. The aim of the pre-processing is an improvement of the image that
suppresses unwilling distortions or enhances some image features, which is important for
future processing of the images. This step focuses on image feature processing. The
process of image retrieval system is as shown in figure 4.4. Filtering is a technique for
modifying or enhancing an image. The image is filtered to emphasize certain features or
remove other features. The noise in the images is filtered using linear and non-linear
filtering techniques. Median filtering is used here to reduce the noise [8].
An RGB color images is an M*N*3 array of color pixels, where each color pixel
is a triplet corresponding to the red, green, and blue components of an image at a spatial
location. An RGB image can be viewed as the stack of three gray scale images that, when
fed into the red, green, blue inputs of a color monitor, produce the color image on the
screen. By convention the three images form an RGB images are called as red, green and
blue components. After calculating the mean values of Red, Blue and Green components,
the values are to be compared with each other in order to find the maximum value of the
components. For e.g. if the value of Red component is High than the rest of the two, then
we can conclude that the respective image is Red Intensity oriented image and which can
be clustered into Red Group of Images. Whenever the query image is given, calculate the
RGB components average values. Then compare this with the stored values. After the
procedure of feature extraction we perform the procedure of similarity comparison. Then
we get the target image.
26
Figure 4.4: Image Retrieval System
27
CHAPTER – 5
SYSTEM IMPLEMENTATIONIn this chapter we focus on the various parameters related with the project. i.e we discuss
on the various screenshots of the project.
Figure 5.1: Main Form
This is the main form of the project. It shows the name of the project and the process
button after clicking the process button we start our system work.
28
Figure 5.2: Input Image
After entering inside the project after clicking o the start process button we enter inside
the systems. After entering inside the system we have to provide username and password.
This can be done by using the username and password field. If the username and
password is not valid it displays the error message. The input image can be selected with
the help of input image button.
29
Figure 5.3: Resultant Images
Figure 5.3 indicates the result of the input image. In this form we come to know the age
and gender of the input image.
Load image button is useful to load the image.
Close button is used to close the window
Get Gender age button is used to get the age of the person.
Select Test image button is used to select the input image.
30
Figure 5.4: Result for the other Image
Figure 5.4 shows the same work as we discuss in the previous form. The previous result
can be done by taking the input image of the male. The above form also indicates the
number of training st used for the given image. It also shows the age and the sex of the
image. We also represent the deviation factor which is nothing but the matching
percentage of the image with their name.
31
Figure 5.5: Result for Female Image
Figure 5.5 shows the working of the system after taking the input image of the female.
32
CHAPTER – 6
RESULT ANALYSISImage basically made up of number of pixels. Any work that is related with the image
can be done on the basis of number of pixels in an image, Size of the image, color
intensity of the image and the shape of image. In this section we are focusing on the
various parameters i.e. size of image ,cropping the image , rotate the image and change
the color intensity of the image that are associated with the image. After performing
various operation of the image we get different results that we have to focus.
The table 6.1 indicates the comparative study of image parameters. From the
given table it is observed that when we change the size of the image the Clustering time,
mining turnaround time, and Overall image retrieval time also get decreases.
It also indicates that if the deviation factor get increases then the mining precision
get decreases.
Test 1 2 3 4
Image Type jpg jpg Jpg Jpg
Image Format 24 Bits RGB 32 Bits RGB 32 Bits RGB 32 Bits RGB
Clustering Time 00:00:38:26 00:00:02:56 00:00:0:93 00:00:00:07
Clustering Precision 98.00% 99.62% 98.25% 96.23%
Mining Turn Around time 00:00:14:65 00:00:04:01 00:00:03:02 00:00:03:07
Overall Image
Retrieval Time
00:00:53:682 00:00:06:58 00:00:03:95 00:00:03:14
Deviation Factor 0 9.26 11.49 48.52
Overall Mining
Precision100% 90.73% 88.50% 51.47%
Table 6.1: Result Analysis with respect to Deviation Factor
33
To analyze the table easily on the basis of overall mining precision and deviation
factor we create the graph. The figures 6.1 indicate the graph between precision and
deviation factor.
Figure 6.1: Deviation Factor vs. Overall Mining Precision
Deviation factor is the difference between the two images. From this graph it is
observed that when the deviation factor is 0 at that time the overall mining precision is
100.After changing the size of the image the deviation factor get changed i.e. for original
image it shows 0 deviation factor and after change the size the deviation factor get
changed from 0 to 9.26.and when deviation factor is 11.49, the overall mining precision
is 88.50.
34
Test 1 2 3
Image Type jpg Jpg jpg
Image Format 32 Bits RGB 32 Bits RGB 32 Bits RGB
Clustering Time 00:00:56:23 00:00:40:31 00:00:54:40
Clustering Precision 98.944 97.186 98.2061
Mining Turn Around time 00:00:014:089 00:00:010:0281 00:00:03:054
Overall Image Retrieval Time 00:00:70:31 00:00:50:33 00:00:57:454
Deviation Factor 29.91 10.15 8.54
Overall Mining Precision 70.08% 89.84% 91.45%
Table 6.2: Result Analysis with respect to overall mining Precision
This graph indicates that once the deviation factor get increases the overall mining
precision get decreases. When the deviation factor is 10.15 the overall mining precision is
89.84.The value of the deviation factor and precision get changed as compare to the
previous value. The graph is shown in figure 6.2
From the above graph it is concluded that the when the deviation factor get
increases at that time the mining precision get decreases.
35
Figure 6.2: Deviation Factor vs. Overall Mining Precision
6.1 RECALL AND PRECISIONTesting the effectiveness of the content based image Retrieval about testing how the
CBIR can retrieve similar images to the query image and how well the system prevents
the return results that are not relevant to the source at all in the user point of view. A
sample query image must be selected from one of the image category in the database.
When the system is run and the result images are returned, the user needs to count how
many images are returned and how many of the returned images are similar to the query
image.
Determining whether or not two images are similar is purely up to the user’s
perception. Human perceptions can easily recognize the similarity between two images
although in some cases, different users can give different opinions. After images are
retrieved, the system’s effectiveness needs to be determined. To achieve this, two
evaluation measures are used. The first measure is called Recall. It is a measure of the
ability of a system to present all relevant items. The equation for calculating recall is
given below
36
Recall =Number of Relevant items Retrieved
Number of relevant items in Collection
The second measure is called Precision. It is a measure of the ability of a system
to present only relevant items. The equation for calculating Precision is given below
Precision =Number of Relevant items Retrieved
Total Number of items retrieved
The number of relevant items retrieved is the number of the returned images that
are similar to the query image in this case. The number of relevant items in collection is
the number of images that are in the same particular category with the query image. The
total number of items retrieved is the number of images that are returned by the system.
Table 6.3 indicates the result analysis on the basis of recall and precision. From
these it is observed that when we considered the test of the imageto find the gender of the
person in the image we get better recall as well as precision value. First test we get the
recall value 91.30 and the precision value 93.33 in our system. During second test the
recall value is 89.36 and the precision value is 91.34 in our system.
TestTotal Number
of Relevant Images
Number of Relevant Images
Retrieved
Total Number of Retrieved Images
Recall Precision
1 46 42 45 91.30 93.33
2 47 42 46 89.36 91.34
3 45 38 42 84.44 90.37
4 42 35 41 83.33 85.36
5 40 25 35 59.52 71.42
Table 6.3: Comparison of results with refer to Recall & Precision
37
Figure 6.3 indicates the graphical representation of the table 6.3.We plot the graph
between recall and the precision. Here we are plotting the Figure between number of
Images and Images retrieved.
0
10
20
30
40
50
60
70
80
90
100
RecallPrecision
Test on Number of Images
Imag
es R
etrie
ved
Figure 6.3: Graph on the basis of Recall & Precision
38
CHAPTER – 7
CONCLUSIONWe have developed a system which combines the Clustering and neural network
approach to identify humans’ genders and their predicted age after analyzing the images.
Our approach can be integrated with the face detection since they can share the same
integral image. An effective and efficient gender identifier can thus be realized.
We proposed an approach for age estimation using facial features based on neural
networks. We classified the ages firstly into four categories, and then each age range
category is also classified into two more specific age ranges. This had not been done
before elsewhere. We used five neural networks to achieve our task. The facial features
rely on 68 landmark points taken from face images. The development process includes
age classification, data collection, feature extraction by markup tool and finally training
and testing the system by Easy NN.
The main objective of the images is to find sex and age of the person in our
system. The images are preprocessed with various techniques and the texture calculation
is highly focused. Here, images are clustered based on RGB Components, Texture values
and Fuzzy C mean algorithm. Entropy is used to compare the images with some threshold
constraints. This application can be used in future to classify the medical images in order
to diagnose the right disease verified earlier.
The dramatic rise in the sizes of images databases has stirred the development of
effective and efficient retrieval systems. The development of these systems started with
retrieving images using textual annotations but later introduced image retrieval based on
color content. This came to be known as Cluster Oriented Image Retrieval. Systems,
using this we can retrieve images based on visual features such as color, texture and
shape, as opposed to depending on image descriptions or textual indexing. In this project
we create an image retrieval system that evaluates the similarity of each image in its data
store to a query image in terms of color and textural characteristics, and returns the
images within a desired range of similarity. From among the existing approaches to color
and texture analysis within the domain of image processing, we have adopted the
39
histogram to extract texture and color features from both the query images and the images
of the data store.
REFERENCES
40
[1] Fukai, H., Nishie, Y., Abiko , K., Mitsukura, Y., Fukumi, M. and Tanaka, M. "An
Age Estimation System on the AIBO", International Conference On Control, Automation
And Systems, pp.2551-2554, 2008.
[2] Geng, X., Zhou, Z. and Smith-Miles, K. "Automatic Age Estimation Based on Facial
Aging Patterns", IEEE Transaction On Pattern Analysis And Machine Intelligence, Vol.
29, No. 12, pp.2234-2240,December 2007.
[3] Geng, X., Smith-Miles, K. and Zhou, Z. "Facial Age Estimation by Nonlinear Aging
Pattern Subspace", Proceedings Of The 16 the ACM International Conference on
Multimedia , pp. 721-724, 2008.
[4] Horng, W., Lee, C. and Chen, C. "Classification of Age Groups Based on Facial
Features", Journal of Science and Engineering, Vol.4, No. 3, pp. 183-192, 2001
[5] Cootes , F., C. J. Taylor, D. H. Copper, and J. Graham, "Active Shape Models – Their
training and Application", Computer Vision Graphics And Image Understanding, Vol.
61, No 1, pp. 38–59, 1995.
[6] Cootes ,F. , Gareth J. Edwards , Christopher J. Taylor, "Active Appearance Models",
Proceedings of the 5th European Conference on Computer Vision-Volume II, p.484-498,
June 02-06, 1998.
[7] http://www.fgnet.rsunit.com
[8] Luu, K., Ricanek Jr., K. and Y. Suen, C. "Age Estimation using Active Appearance
Models and Support Vector Machine Regression", Proceedings Of The IEEE
International Conference On Biometrics :Theory Applications And System , pp.314-
318 ,2009.
[9] Ricanek Jr., K and Tesafaye, T., “ MORPH: A Longitudinal Image Database of
Normal Adult Age-Progression", Proceedings of the 7th International Conference on
Automatic Face and Gesture Recognition, p.341-345, April 10-12, 2006.
[10] http://personalpages.manchester.ac.uk/staff/timothy.
[11] http://www.scholarpedia.org/article/Facial_Age_Estimation Y. H. Kwon and N. da
Vitoria Lobo (1999). ‘Age Classification from Facial Images’, Computer Vision and
Image Understanding Journal 74(1), pp. 1–21.N
41
[12] Ramanathan, R. Chellappa, Face verification across age Progression, in: Proceedings
of IEEE Conference on Computer Vision and Pattern Recognition, vol. 2, 2005, pp. 462–
469.
[13] B.W. Hwang, H. Byun, M.C. Roh, S.W. Lee, Performance evaluation of face
recognition algorithms on the Asian face database, KFDB, in: Proceedings of Audio-and
Video-Based Biometric Person Authentication (AVBPA), 2003, pp. 557–565.
[14] Luu, K., Ricanek Jr., K. and Y. Suen, C. "Age Estimation using Active Appearance
Models and Support Vector Machine Regression", Proceedings Of The IEEE
International Conference On Biometrics :Theory Applications And System 2010.
[15] Ricanek Jr., K and Tesafaye, T., “MORPH: A Longitudinal Image Database of
Normal Adult Age-Progression", Proceedings of the 7th International Conference on
Automatic Face and Gesture Recognition, p.241-245, Aug 10-12, 2008.
42