=Paper=
{{Paper
|id=Vol-1171/CLEF2005wn-ImageCLEF-MahmudurRahmanEt2005
|storemode=property
|title=Supervised Machine Learning based Medical Image Annotation and Retrieval
|pdfUrl=https://ceur-ws.org/Vol-1171/CLEF2005wn-ImageCLEF-MahmudurRahmanEt2005.pdf
|volume=Vol-1171
|dblpUrl=https://dblp.org/rec/conf/clef/RahmanDB05a
}}
==Supervised Machine Learning based Medical Image Annotation and Retrieval==
https://ceur-ws.org/Vol-1171/CLEF2005wn-ImageCLEF-MahmudurRahmanEt2005.pdf
Supervised Machine Learning based Medical
Image Annotation and Retrieval
Md. Mahmudur Rahman, Bipin C. Desai, Prabir Bhattacharya
CINDI Group, Concordia University
1455, De Maisonneuve Blvd. West,
Montreal, QC, H3G 1M8, Canada
mah rahm@cs.concordia.ca
Abstract
This paper presents the approaches and experimental results of image annotation and
retrieval in our first participation of ImageCLEFmed 2005. In this work, we investigate
a supervised learning approach to associate low-level global image features with their
high level visual and/or semantic categories for image annotation and retrieval. For
automatic image annotation, we represent input images through a large dimensional
feature vector of texture, edge and shape features. A multi-class classification system
based on pairwise coupling of several binary support vector machine (SVM) is trained
on this input to predict the categories of test images, which will be effective for later
annotation. For visual only retrieval, we utilize a low dimensional feature vector of
color, texture and edge features based on principal component analysis (PCA) and
category specific feature distribution information in a statistical similarity measure
function. Based on the online category prediction of query and database images by
the multi-class SVM classifier, pre-computed category specific first and second order
statistical parameters are utilized in Bhattacharyya distance measure on the assump-
tion that distributions are multivariate Gaussian. Experimental results of both image
annotation and retrieval are reported in this paper.
Categories and Subject Descriptors
H.3 [Information Storage and Retrieval]: H.3.1 Content Analysis and Indexing; H.3.3 Infor-
mation Search and Retrieval; H.3.7 Digital Libraries; I.4.8 [Image Processing and Computer
Vision]: Scene Analysis—Object Recognition
General Terms
Algorithms, Machine learning, Performance, Experimentation
Keywords
Image annotation, Classification, Content based image retrieval, Support vector machine, Statis-
tical distance measure
1 Introduction
During the last decade, there have been an overwhelming research interests in medical image re-
trieval and classification from different communities [11, 10]. Medical images of various modalities
(X-ray, CT, MRI, Ultrasound etc.) constitute an important source of anatomical and functional
information for the diagnosis of diseases, medical research and education. Effectively and efficiently
searching and annotating these large image collections poses significant technical challenges.
ImageCLEFmed is part of the Cross Language Evaluation Forum (CLEF), a benchmarking
event for multilingual information retrieval. The main goal of ImageCLEFmed is to improve the
retrieval of medical images from heterogeneous and multilingual document collections containing
images as well as text. This year (2005) in ImageCLEFmed, there are mainly two task: image
annotation and retrieval to be performed.
In automatic image annotation task, the main aim is to find out how well current techniques
can identify image modality, body orientation, body region, and biological system examined based
on the images. Here, we utilized a database of 9,000 fully classified images taken randomly from
medical routine to train a classification system. 1,000 images for which classification labels are not
available are used as test images and have been classified by a multi-class classification system.
The results of the classification step can be used for multilingual image annotations as well as for
DICOM header corrections.
In image retrieval task, we have experimented with a visual only approach, where an example
image is used to perform a search against a medical image database to find similar images based
on visual attributes (color, texture, etc.). Each medical image or a group of images represents an
illness, and case notes in English or French are associated with each illness to be used for diagnosis.
This task is based on the Casimage, MIR, PEIR, and PathoPIC datasets, containing about 50,000
images of different modalities (CT, MRI, X-ray etc.).
2 Image annotation
Automatic image annotation or image classification is an area of active research in the field of ma-
chine learning and pattern recognition. Retrieval systems have traditionally used manual image
annotation for indexing and then later retrieving their image collections. However, manual image
annotation is an expensive and labor intensive procedure [7]. Here, we investigate an automatic
approach to categorize or annotate images based on a supervised learning technique. A classifi-
cation task usually involves with training and testing data which consist of some data instances.
In supervised classification, we are given a collection of labeled images (or priori knowledge), and
the problem is to label a newly encountered, yet unlabeled image. Each instance in the training
set contains category or class specific labels and several image feature descriptors in the form of a
combined feature vector.
In this work, we present effective texture, edge and shape descriptors to represent image content
at global level and input in to a multi-class classifaction system based on several binary support
vector machine (SVM) classifier. The goal of SVM is to produce a model which will predict target
value or category of images with highetst probability or confidence in the testing set which are
given as input feature vectors to the classification system.
2.1 Feature Selection & Extraction
This section describes how our system characterizes images for efficient representation. The col-
lection contained mainly monochrome or gray level medical images with specific layout. Hence,
we characterize images by texture, edge and shape features, whereby ignoring color information
totally. We know that only one type of low-level features does not work well. So, we incorporate
these features in a combined feature vector and use it as input to the classification system. The
accuracy of our classification system depends greatly on the representation of these low-level vi-
sual features. These image features help propagate annotation from trained images to unlabeled
images. The more discriminative the low-level features, the more accurate the content-based
classification or annotation.
2.1.1 Texture Feature
Medical images of different categories can be distinguished via their homogeneousness or texture
characteristics. Therefore, it may be useful to extract texture features for image classfication.
We include a texture feature set approach to extract spatially localized texture information
from the gray level co-occurrence matrix [1]. A gray level co-occurrence matrix is defined as
a sample of the joint probability density of the gray levels of two pixels separated by a given
displacement d and angle θ. The G × G gray level co-occurrence matrix p for a displacement
vector d = (dx, dy) is defined as [16]:
p(i, j) = |((r, s), (t, v)) : (I(r, s) = i, I(t, v) = j|, (1)
where (r, s), (t, v) ∈ N × N, (t, v) = (r + dx)(s + dy).
Likewise image histograms, which can be thought of as an estimate of the probability distri-
bution of gray values or color levels in an image, a co-occurrence matrix is an estimate of the
joint probability distribution of pairs of gray-levels at some fixed relative position in the image.
Typically, the information stored in a co-occurrence matrix is sparse. Also, it is often useful to
consider a number of co-occurrence matrices, one for each relative position of interest, in order to
grasp different textual cues or the same textual cur at different scale. In order to obtain efficient
descriptors, the information contained in co-occurrence matrices is traditionally condensed in a
few statistical features. We obtained four co-occurrence matrices for four different orientations
(horizontal 0◦ ,vertical 90 ◦ , and two diagonals 45 ◦ and 135 ◦ ) and normalize the entries [0,1] by
dividing each entry by total number of pixels.
The co-occurrence matrix reveals certain properties about the spatial distribution of the gray
levels in the texture image. For example, if most of the entries in the co-occurrence matrix are
concentrated along the diagonals, then the texture is coarse with respect to the displacement
vector d. Haralick has proposed a number of useful texture features that can be computed from
the co-occurrence matrix. Higher order features, such as energy, entropy, contrast, homogeneity
and maximum probability are measured based on each gray level co-occurrence matrix to form a
five dimensional feature vector as follows [16]:
XX
Energy = p2 (i, j) (2)
i j
XX
Entropy = − p(i, j)logp(i, j) (3)
i j
XX
Contrast = (i − j)2 p(i, j) (4)
i j
XX p(i, j)
Homogeneity = (5)
i j
1 + |i − j|
M aximum probability = max p(i, j) (6)
Finally, we obtained a twenty dimensional feature vector by concatenating the feature vectors
of each co-occurrence matrix.
2.1.2 Edge Feature
To represent the edge feature on a global level, a histogram of edge direction is constructed. The
edge information contained in the images is processed and generated by using the Canny edge
detection (with σ = 1, Gaussian masks of size = 9, low threshold = 1, and high threshold =
255) algorithm [4].The corresponding edge directions are quantized into 72 bins of 5◦ each. Scale
invariance is achieved by normalizing this histograms with respect to the number of edge points
in the image. Hence, the dimension of the edge feature vector is 72.
2.1.3 Shape Feature
Image classification for medical images with gary scale images have to use shape feature for efficient
representation. We represent the global shape of an image in terms of seven invariant moments
[8]. These features are invariant under rotation, scale, translation, and reflection of images and
have been widely used in a number of applications due to their invariance properties. For a 2-D
image, f (x, y), the central moment of order (p + q) is given by
XX
µpq = (x − x)p (y − y)q f (x, y). (7)
x y
Seven moment invariants (M1 − M7 ) based on the second and third order moments are given as
follows [8]:
M1 = (µ20 + µ02 )
M2 = (µ20 − µ02 )2 + µ211
M3 = (µ30 − 3µ12 )2 + (3µ21 − µ03 )2
M4 = (µ30 + µ12 )2 + (µ21 + µ03 )2
M5 = (µ30 + µ12 ) + (µ30 − 3µ12 )[(µ30 + µ12 )2 − 3(µ21 + µ03 )2 ]
+(3µ21 − µ03 ) × (µ21 + 3µ03 )[3(µ03 + µ21 )2 − (µ21 − µ03)2 ]
M6 = (µ20 + µ02 )[(µ30 + µ12 )2 − (µ21 + µ03 )2 + 4µ11 (µ30 + µ12 )(µ21 + µ03 )
M7 = (3µ21 − µ03 )(µ30 + µ12 )[(µ30 + µ12 )2 − 3(µ21 + µ03 )2 ]
−(µ30 − 3µ12 )(µ21 + µ03 )[3(µ03 + µ21 )2 − (µ21 − µ03)2 ] (8)
M1 − M6 are invariant under rotation and reflection. M7 is invariant only in its absolute
magnitude under a reflection.
Now, let us consider ft , fe , and fs be the texture, edge and shape feature vector respectively of
an image. Now the composite feature vector is formed by simple concatenation of each individual
feature vector as Fcombined = (ft + fe + fs ), where the dimension Rd of Fcombined is the sum of
individual feature vector dimension and d = (20 + 72 + 7) = 99. We use this vector to represent
the images. Thus, the input space for our SVM classifiers is a 99-dimensional space, and each
image in our database corresponds to a point in this space.
2.2 Category prediction with multi-class SVM
Support vector machine (SVM) is an emerging machine learning technology which has been suc-
cessfully used in content based image retrieval [5].
Given training data (~x1 , . . . , ~xn ) that are vectors in some space ~xi ∈ Rn and their labels
(y1 , . . . , yn ) where yi ∈ (+1, −1)n , the general form of the binary linear classification function is
g(x) = w
~ · ~x + b (9)
which corresponds to a separating hyper plane
w
~ · ~x + b = 0 (10)
where ~x is an input vector, w
~ is a weight vector, and b is a bias. The goal of SVM is to find the
2
parameters w ~ and b for the optimal hyper plane to maximize the geometric margin ||w||
~ between
the hyper planes, subject to the solution of the following optimization problem [3]:
n
min 1 T X
w
~ w
~ +C ξi (11)
w,
~ b, ξ 2 i=1
subject to
~ T φ(~xi ) + b) ≥ 1 − ξi
yi (w (12)
where ξi ≥ 0 and C > 0 is the penalty parameter of the error term. Here training vectors ~xi are
mapped into a high dimensional space by the non linear mapping function φ : Rn → Rf , where
f > n or f could even be infinite. Optimization problem and its solution can be represented by
the inner product. Hence,
~xi .~xj → φ(~xi )T φ(~xj ) = K(~xi , ~xj ) (13)
where K is a kernel function. The SVM classification function is given by [5]:
n
!
X
f (~x) = sign αi yi K(~xi , ~x) + b (14)
i=1
A number of methods have been proposed for extension to multi-class problem to separate
L mutually exclusive classes essentially by solving many two-class problems and combining their
predictions in various ways [5]. One technique, commonly known as Pairwise Coupling (PWC)
or “one-vs.-one”is to construct SVMs between all possible pairs of classes. This method trains
L ∗ (L − 1)/2 binary classifiers, each of which provides a partial decision for classifying a data
point. PWC then combines the output of all classifiers to form a class prediction. During testing,
each of the L ∗ (L − 1)/2 classifier votes for one class. The winning class is the one with the largest
number of accumulated votes. We use this technique for the implementation of our multi-class
SVM by using the LIBSVM software package [6].
Multi-class image classification systems work by classifying an image into one of many pre-
defined categories. There are 57 categories of images provided in the training set for the image
annotation task. So, we define a set of 57 labels where each label characterizes the representative
semantics of an image category. Class labels along with feature vectors are generated from all
images at the training stage. In testing stage, each unannotated image is classified against the 57
categories using PWC or “one-vs.-one” technique. This produces a ranking of the 57 categories,
with each category assigned a confidence or probability score to each image. The labels of these
57 categories, along with their probabilities, become the annotation for the testing images as the
following format:
< imageno >< conf idence class1 >< conf idence class2 > · · · < conf idence class57 >
The confidence represents the weight of a label or category in the overall description of an
image and the class with the highest confidence is considered to be the class of the image. This
ranking of categories based on confidence score can be very useful in future annotation or search
purposes.
3 Image retrieval
Currently, most content based image retrieval (CBIR) systems are similarity-based, where similar-
ity between query and target images in a database is measured by some form of distance metrics in
feature space [15]. However, retrieval systems generally conduct this similarity matching on a very
high-dimensional feature space without any semantic interpretation or paying enough attention to
the underlying distribution of the feature space [14]. However, high dimensional feature vectors
not only increase the computational complexity in similarity matching and indexing, but also in-
crease the logical database size. For many frequently used visual features in medical images, often
their category specific distributions are also available. In this case, it is possible to extract a set
of low-level features (e.g., color, texture, shape, etc.) to predict semantic categories of each image
by identifying its class assignment using a classifier. Thus, an image can be best characterized by
exploiting the information of feature distribution from its semantic category.
In the image retrieval task, we have investigated a category based adaptive statistical similarity
measure technique on the low-dimensional feature space. For this, we utilized principal component
analysis (PCA) for dimension reduction and multi-class support vector machine (SVM) for online
category prediction of query and database images. Hence, category specific statistical parameters
in low-dimensional feature space can be exploited by the statistical distance measure in real time
similarity matching.
3.1 Feature extraction & representation in PCA sub-space
The performance of a CBIR system mainly depends on the particular image representation and
similarity matching function employed. Images in four datasets contain both color and gray level
images for retrieval evaluation. Hence, we have extracted color, texture and edge features for our
image representation at global level.
Color is the most useful low-level feature and its histogram based representation is one of the
earliest descriptors, which is widely used in CBIR [15]. For color feature, a 108 dimensional color
histogram is created in vector form on HSV (Hue, Saturation, Value) color space. In HSV space,
the colors correlates well and can be matched in a way that is consistent with human perception.
In this work, we uniformly quantized HSV space into 12 bins for hue(each bin consisting of a range
of 30◦ ), 3 bins for saturation and 3 bins for value, which results in 108 bins for color histogram.
Many medical images of different categories can be distinguished via their homogeneous texture
and global edge characteristics. Hence, we have extracted the same global texture and edge
features as measured for image annotation.
As the dimensions of color, texture and edge (108+20+72 = 200) feature vectors are high,
we need to apply some dimension reduction technique to reduce the computational complexity
and logical database size. Moreover, if the training samples used to estimate the statistical para-
meters are smaller compare to the size of feature dimension, then inaccuracy or singularity may
arise for second order (covariance matrix) parameter estimation. The problem of selecting most
representative feature attributes commonly known as dimension reduction, has been examined by
principal component analysis (PCA) [12] in our experiment. The basic idea of PCA is to find m
linearly transformed components so that they explain the maximum amount of variances in the
input data and mathmetical steps used to describe the method is as follows:
given a set of N feature vectors (training samples) ~xi ∈ Rd |i = (1 · · · N ), where the mean vector(µ)
and covariance matrix (C) is estimated as
N N
1 X 1 X
µ= ~xi & C= (~xi − µ)(~xi − µ)T (15)
N i=1 N i=1
Let νi and λi be the eigenvectors and the eigenvalues of C, then they satisfy the following:
N
X
λi = (νiT (~xi − µ))2 (16)
i=1
PN
Here, i=1 λi accounts for the total variance of the original feature vectors set. Now, PCA method
tries to approximate the original feature space using an m dimensional feature vector, that is using
m largest eigenvalues account for a large percentage of variance, where typically m << min(d, N ).
These m eigenvectors span a subspace, where V ~ = [~v1 , ~v2 , · · · , ~vm ] is the d × m-dimensional matrix
that contains orthogonal basis vectors of the feature space in its columns. The m×d transformation
~ T transforms the original feature vector from Rd → Rm ones. That is
V
~ T (~xi − µ) = ~yi , i = 1 · · · N
V (17)
where ~yi ∈ Rm and kth component of the ~yi vector is called the kth principal component (PC)
of the original feature vector ~xi . So, the feature vector in the original Rd space for query and
database images can be projected on to the Rm space via the transformation of V ~ T [12]. We have
applied this PCA technique to our composite feature vector and reduced the feature dimension
for subsequent SVM training and similarity matching.
3.2 Adaptive statistical distance measure
Statistical distance measure, defined as the distances between two probability distributions, finds
its uses in many research areas in pattern recognition, information theory, and communication. It
captures correlations or variations between attributes of the feature vectors and provides bounds
for probability of retrieval error of a two way classification problem. Recently, CBIR community
also adopted statistical distance measures for similarity matching [2, 14]. In this scheme query
image q and target image t are assumed to be in different classes and their respective density as
pq (~x) and pt (~x), both defined on Rd . When these densities are multivariate normal, they can be
approximated by mean vector µ and covariance matrix C as pq (~x) = N (~x; µq , Cq ) & pt (~x) =
N (~x; µt , Ct ) where,
1 1 T
exp− 2 (~x−µ) C (~x−µ)
−1
N (~x; µ, C) = p (18)
(2π)d|C|
here, x ∈ Rd and | · | is matrix determinant [9]. A popular measure of similarity between two
Gaussian distributions is the Bhattacharyya distance, which is equivalent to an upper bound of
the optimal Bayesian classification error probability [9] [13]. Bhattacharyya distance (DBhatt )
between query image q and target image t in the database is given by:
−1 (Cq +Ct )
1 (Cq + Ct ) 1 2
DBhatt (q, t) = (µq − µt )T (µq − µt ) + ln p (19)
8 2 2 |Cq ||Ct |
where µq and µt are the mean vectors, and Cq and Ct are the covariance matrices of query image
q and target image t respectively. Equation (19) is composed of two terms, the first one being
the distance between mean vectors of images, while the second term gives the class separability
due to the difference between class covariance matrices. In the retrieval experiment, we used
the Bhattacharyya distance measure for similarity matching. Here, it is called adaptive due to
the nature of online selection of mean (µ) and covariance matrix (C) by the multi-class SVM as
discussed in the next section.
3.3 Category prediction & parameter estimation
To utilize category specific distribution information in similarity matching, we utilize the multi-
class SVM classifer as described in section 2.2 to predict the category of query and databse images.
Based on the online prediction, distance measure functions will be adjusted to accommodate cat-
egory specific parameters for query and reference images of database. To estimate the parameters
of the category specific distributions, feature vectors are extracted and reduced in dimension as
mentioned in section 3.1, from N selected training image samples. It is assumed that feature of
each category will have distinguishable normal distribution. Computing the statistical distance
measures between two multivariate normal distributions requires first and second order statistics
in the form of mean (µ) and covariance matrix (C) or parameter vector θ = (µ, C) as described
in previous section. Suppose that there are L different semantic categories in the database, each
assumed to have a multivariate normal distribution with mean vector µi and covariance matrix Ci ,
for i ∈ L. However, the true values of µ and C of each category usually are not known in advance
and must be estimated from a set of training samples N [9]. The µi and Ci of each category are
estimated as
Ni Ni
1 X 1 X
µi = ~xi,j & Ci = (~xi,j − µi )(~xi,j − µi )T (20)
Ni j=1 Ni − 1 j=1
where ~xi,j is sample j from category i, Ni is the number of training samples from category i and
N = (N1 + N2 + . . . + NL ).
4 Experimentation & Results
For annotation experiment, we perform the following procedure at the training stage:
Figure 1: Block diagram of the retrieval technique
• Consider the radial basis kernel function (RBF) K(xi , xj ) = exp(−γ||xi − xj ||2 ), γ > 0.
• Use cross-validation to find the best parameter C and γ
• Use the best parameter C and γ to train the whole training set
For image classification, recent work [5] shows that the RBF kernel works well with high-
dimensional image data and can handle the case when the relation between class labels and
attributes is nonlinear. Therefore, we use RBF kernel as a reasonable first choice. There are two
tunable parameters while using RBF kernels and soft-margin SVMs in the version space: C and
γ. The γ in the RBF kernel controls the shape of the kernel and C controls the trade-offs between
margin maximization and error minimization. Increasing C may decrease training error, but it
can also lead to poor generalization [7].
It is not known beforehand which C and γ are the best for our classification problem. In the
training stage, the goal is to identify good (C and γ ), so that the classifier can accurately predict
testing data. It may not be useful to achieve high training accuracy (i.e., classifiers accurately
predict training data whose class labels are indeed known). Therefore, we use a 10 fold cross-
validation with various combinations of γ and C to measure the classification accuracy.
In 10-fold cross-validation, we first divide the training set into 10 subsets of equal size. Se-
quentially one subset is tested using the classifier trained on the remaining 9 (10-1) subsets. Thus,
each instance of the whole training set is predicted once so the cross-validation accuracy is the
percentage of data which are correctly classified. We perform a grid-search on C and γ and using
cross-validation. Basically pairs of (C , γ ) are tried and the one with the best cross-validation
accuracy is picked. We find the best (C , γ ) is (200, 0.01) with the cross-validation rate 54.65%.
After the best (C , γ) is found, the whole training set of 9,000 images is trained again to generate
the final classifiers. Finally, the generated classifiers are tested on the 1000 image testing set of
unknown labels to generate the image annotation.
In ImageCLEFmed 2005 automatic image annotation experiment, we have submitted only one
run with the above parameters and the classification error rate is 43.3%. Which means, 433 images
were misclassified out of 1000 or accuracy of our system is 56.7% at this moment.
In image retrieval experiment, for statistical parameter estimation and SVM training, we ob-
served the images of four datasets closely and selected 33 different categories based on perceptual
and modality specific differences, each with 100 images for generating the training samples. How-
ever, for actual evaluation of similarity measure functions, we conducted our experiments on the
entire database (Around 50,000 images from four different collections). For SVM training, we used
Figure 2: A snapshot of the retreival result (most similar 15 images) of a Chest-CT (Image17.jpg)
query image
the reduced feature vector with radial basis kernel (RBF) function. After 10 fold cross validation,
we found the best parameters C = 100 and γ = .03 with an accuracy of 72.96% in our current
setting and finally trained the whole training set with these parameters. The dimensionality of
the feature vector is reduced to Rd → Rm , where d=200 and m=25 and accounted for 90.0% of
the total variances. In the ImageCLEFmed 2005 evaluation, we have submitted only one run with
the above parameters and achieved a mean average precision of 0.0072 across all queries, which is
very low at this moment compare to other systems.
5 Conclusion and future work
This is the first year for the CINDI group taking part in the ImageCLEF campaign and specially
in the ImageCLEFmed track. This year, our main goal was to participate and do some initial
experiment with the databases provided by the organizer.
This report presents our approaches to automatic image annotation and retrieval based on
global image feature contents and multi-class SVM classifier. Despite having 57 categories for
annotation, many of them closely-related, our classification system is still able to provide moderate
classification performance. In future, we will investigate with region based local image features and
statistical methods that can deal with the challenges of an unbalanced training set, as provided
in the current experimental setting.
The retrieval performance of our system is not good enough due to the following main reasons.
First of all, it is very difficult to select a reasonable training set of images with predifined categories
from four different datasets with a huge amount of variability in image size, resolution, modality
etc. The performance of our system is critical to the appropriate training of multi-class SVM as
parameter selection of statistical distance measure is depended on the online category prediction.
Secondly, we have only used global image feature, which may not suitable for medical images as
large unwanted background dominated in many of these images. In future we will try to resolve
these issues and incorporate text based search approach in addition with the visual based approach.
References
[1] Aksoy, S., Haralick, R. M.: Texture Analysis in Machine Vision., Chapter Using Texture
in Image Similarity and Retrieval,Series on Machine Perception and Artificial Intelligence.,
World Scientific (2000)
[2] Aksoy, S., Haralick, R.M.: Probabilistic vs. geometric similarity measures for image retrieval.,
Proceedings. IEEE Conference on Computer Vision and Pattern Recognition,2(2000) 357–362
[3] Burges, C.: A tutorial on support vector machines for pattern recognition., Data Mining and
Knowledge Discovery, 2, (1998) 121–167
[4] Canny, J. : A computational approach to edge detection., IEEE Trans. Pattern Anal. Machine
Intell., 8 (1986) 679–698
[5] Chapelle, O., Haffner, P., Vapnik, V.: SVMs for histogram-based image classification.,IEEE
Transaction on Neural Networks (1999)
[6] Chang, C. C., Lin, C.J.: LIBSVM : a library for support vector machines., (2001) Software
available at http://www.csie.ntu.edu.tw/ cjlin/libsvm
[7] Chang, E. Kingshy, G. Sychay, G. Gang, W. : CBSA: Content-based Soft Annotation for
Multimodal Image Retrieval Using Bayes Point Machines., IEEE Transactions on Circuits
and Systems for Video Technology., 13(2003) 26–38
[8] Dudani, S. A., Breeding, K. J. and McGhee, R. B. : Aircraft identification by moment
invariants., IEEE Trans. Comput. C-26 (1977) 39–45
[9] Fukunaga,K.: Introduction to Statistical Pattern Recognition., Second ed. Academic Press,
(1990)
[10] H. D. Tagare, C. Jafe, J. Duncan. :Medical image databases: A content-based retrieval ap-
proach., Journal of the American Medical Informatics Association., 4 (3) (1997) 184–198
[11] H. Muller, N. Michoux, D. Bandon, A. Geissbuhler :A review of content-based image re-
trieval applications -clinical benefits and future directions., International Journal of Medical
Informatics, 73:1(2004) 1–23
[12] Jain,A.K., Bhandrasekaran, B. : Dimensionality and sample size considerations in pattern
recognition practice., Handbook of Statistics, 2 (1987) 835–855
[13] Kailath,T.: The divergence and Bhattacharyya distance measures in signal selection.,IEEE
Trans. Commun. Technol, COM-15(1967) 52–60
[14] Puzicha, J., Buhmann, J., Rubner, Y., Tomasi, C.: Empirical evaluation of dissimilarity
measures for color and texture., Intern. Conf. on Computer Vision, (1999)
[15] Smeulder, A., Worring, M., Santini, S., Gupta, A., Jain, R.: Content-Based Image Retrieval
at the End of the Early Years., IEEE Trans. on Pattern Anal. and Machine Intell., 22, (2000)
1349–1380
[16] Tuceryan, M., Jain, A. K. : The Handbook of Pattern Recognition and Computer Vision
(2nd Edition), by C. H. Chen, L. F. Pau, P. S. P. Wang (eds.), World Scientific Publishing
Co., (1998) pp. 207–248