A ROI image retrieval method based on CVAAO

Image and Vision Computing 26 (2008) 1540–1549

Contents lists available at ScienceDirect

Image and Vision Computing journal homepage: www.elsevier.com/locate/imavis

A ROI image retrieval method based on CVAAO Yung-Kuan Chan a,*, Yu-An Ho b, Yi-Tung Liu c, Rung-Ching Chen c a b c

Department of Management Information Systems, National Chung Hsing University, No. 250, Kuokuang Road, Taichung 402, Taiwan, ROC Department of Computer Science, National Chung Hsing University, No. 250, Kuokuang Road, Taichung 402, Taiwan, ROC Department of Information Management, Chaoyang University of Technology, No. 168, Gifeng E. Road, Wufeng, Taichung County, Taiwan, ROC

a r t i c l e

i n f o

Article history: Received 6 June 2006 Received in revised form 1 April 2008 Accepted 24 April 2008

Keywords: ROI image retrieval Color-based image retrieval Color histogram Fuzzy color histogram Hierarchical overlapping segmentation

a b s t r a c t A novel image feature called color variances among adjacent objects (CVAAO) is proposed in this study. Characterizing the color variances between contiguous objects in an image, CVAAO can effectively describe the principal colors and texture distribution of the image and is insensitive to distortion and scale variations of images. Based on CVAAO, a CVAAO-based image retrieval method is constructed. When given a full image, the CVAAO-based image retrieval method delivers the database images most similar to the full image to the user. This paper also presents a CVAAO-based ROI image retrieval method. When given a clip, the CVAAO-based ROI image retrieval method submits to the user a database image containing a target region most similar to the clip. The experimental results show that the CVAAO-based ROI image retrieval method can offer impressive results in finding out the database images that meet user requirements. Ó 2008 Elsevier B.V. All rights reserved.

1. Introduction In content-based image retrieval (CBIR) methods, queryby-example (QBE) is by far the most widely supported method in research prototypes and commercial products [16]. A user formulates a query by giving an example image IQ; the method then extracts the feature of IQ and compares it with the features of pre-feature-extracted images in the database. After that, the method delivers to the user the database images most similar to IQ. QBE image retrieval method has numerous advantages for many practical applications [1,2,5,7,11,15]. For instance, trademarks are specially designed marks used to identify companies, products or services. The imitation of a registered trademark is illegal. However, there are so many trademarks around the world; how to avoid designing a trademark that is similar to an existing one is very difficult. Developing an automatic and fast contentbased trademark image retrieval method is hence necessary. In addition to trademark copyright query, the QBE image retrieval method can still be used in many other application fields such as medical image archiving, computer aided design, and geographic information systems. The CVAAO (color variance among adjacent objects) proposed in this study is a color histogram of the color differences between two adjacent objects in an image. The CVAAO can not only effectively depict the principal color and the texture distribution of an image, * Corresponding author. Tel.: +886 4 22840422; fax: +886 4 22857173. E-mail addresses: [email protected] (Y.-K. Chan), [email protected] (Y.-A. Ho), [email protected] (Y.-T. Liu), [email protected] (R.-C. Chen). 0262-8856/$ - see front matter Ó 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.imavis.2008.04.019

but also distinguish the objects with inconsistent contours. Based on the CVAAO, this paper provides a CVAAO-based image retrieval method which submits to the user the database images most similar the query image assigned by the user. However, in many cases, users are interested in only a small region on a database image. We call this region ‘‘region-of-interest (ROI)”. For example, to automatically classify goods by a computer according to manufacturers, a supermarket can employ a camera to record the images of all articles and use the trademarks of their manufacturers as the query images. Then the articles, whose images are retrieved when given a query image, are made by the manufacturer. We call it ROI image retrieval in which a user may be interested in only finding the database images containing a required region disregarding their backgrounds. Unfortunately, the image of an article typically contains not only ROI but also irrelevant areas. To solve this problem, this paper proposes a CVAAObased ROI image retrieval method. In a ROI image retrieval method [3,10,15], a user may select a region image as a query image IQ; then the method delivers to the user the database images ID, and each of which contains a region R that is most similar to IQ. Here, we call R the ROI, and ID a target database image. R may appear on ID at different locations with varied sizes and rotation angles. For instance, Fig. 1 shows a query image and the target database images that contain shift, rotation, and scale variations and hold a ROI encircled by a red rectangle. An excellent image retrieval method should be insensitive to these variations. The distribution of pixel colors in an image generally contains a lot of interesting information. Recently, many researchers deter-

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

1541

Fig. 1. A query image and the target database images with scale, shift and rotation variant images.

mined the feature for image retrieval by analyzing the color attribute of an image [2,5–8,10,11,13]. Color histogram [2,6,7,10] is the most commonly used feature for color-based image retrieval. Dimai [3] proposed a CCH (conventional color histogram) image retrieval method. By modifying the CCH image retrieval method, Ju et al. [10] presented an FCH (fuzzy color histogram) image retrieval method. On the basis of CCH and FCH, Dimai [3] and Ju et al. [10] still presented CCH-based and FCH-based ROI image retrieval methods, respectively. To support ROI image retrieval, users must be permitted to query arbitrarily shaped images. In other words, a ROI image retrieval method must be able to identify ROI from database images. The CCH-based and FCH-based ROI image retrieval methods adopt hierarchical overlapping segmentation to partition a database image into several overlapping blocks. Then, both methods cut off ROI from the database image by comparing the color histogram of each overlapping block with the color histogram of the query image. Vu et al. [16] also adopted a sampling-based approach (called SamMatch) to implement the ROI query. The image features adopted by the methods mentioned above can describe only the distribution of the principle colors of an image, but cannot characterize the texture of the image. In addition, all these methods partition each database image into some overlapping regions of several sizes, and consider that the query image has the same size as one of the overlapping regions. Nevertheless, the users may give arbitrarily sized query images. Hence, these methods cannot precisely decide the size of ROI. To deal with the aforementioned problems, this paper first introduces the CVAAO which portrays the color variances between contiguous objects in an image. The CVAAO can characterize the principle colors and texture of the image. Based on the CVAAO, this paper presents a CVAAO-based image retrieval method and a CVAAO-based ROI image retrieval method. The CVAAO-based ROI image retrieval method decides the location and size of ROI on a database image via the shape, area, and position of the biggest object in IQ; then it cuts off ROI from the database image. The experimental results show that the proposed method can more accurately meet user‘s requirements. The rest of this paper is organized as follows. The next section briefly reviews the CCH-, FCH-, and SamMatch-based ROI image retrieval methods. Section 3 describes the CVAAO and the CVAAO-based image retrieval method. Section 4 introduces the CVAAO-based ROI image retrieval method. Section 5 investigates

the performances of the CVAAO-based image retrieval method and the CVAAO-based ROI image retrieval method by experiments, and compares them with the CCH-, FCH-, and SamMatch-based image retrieval methods and the CCH-, FCH-, and SamMatch-based ROI image retrieval methods. The conclusions are given in the last section. 2. Related works This section briefly reviews the CCH-, FCH-, and SamMatchbased ROI image retrieval methods whose performances will be compared with that of the CVAAO-based ROI image retrieval method by experiments in this paper. The basic concept of the generic algorithm employed to decide the parameters used in the CVAAO-based ROI image retrieval method, is introduced in this section. This section also introduces ANMRR (average normalized modified retrieval rank) which will be used to measure the retrieval accuracy of an image retrieval method. 2.1. The ROI image retrieval methods reviewing Dimai [3] proposed an image retrieval method which combines the color histogram of an image with a set of IHD’s (inter hierarchical distances) based on a fixed partition of the image. In this paper, the color histogram is defined as the conventional color histogram (CCH), and the image retrieval method as the CCH-based ROI image retrieval method. Since CCH considers neither the color similarity across different bins nor the color difference in the same bin, it is sensitive to noise interference such as illumination changes and quantization errors. Therefore, Ju et al. [10] proposed a fuzzy color histogram (FCH) which associates the color similarity of each pixel’s color to all color histogram bins through fuzzy-set membership function. Based on the FCH, Ju et al. [10] also developed an FCH-based ROI image retrieval method. The FCH F(I) = [f1, f2, . . ., fn] of an image I with N pixels can be defined as

fi ¼

N X j¼1

lij Pj ¼

N 1 X l; N j¼1 ij

where n is the number of color bins in the FCH; Pj, is the probability of the pixel with color j in image I; lij (05lij 51) is the membership value of the jth pixel in the ith color bin.

1542

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

To speed up the computation, Ju et al. [10] performed the fine uniform quantization in RGB color space by mapping all pixel colors to n0 color histogram bins. n0 must be large enough so that the color difference between two adjacent bins is small enough. Then, they transform the n0 colors from RGB to CIELAB color space [8]. Finally, they classify these n0 colors in CIELAB color space into n clusters by using fuzzy c-means (FCM) clustering technique [4], with each cluster representing an FCH bin. During image retrieving, the FCH of a query image can be extracted by the above mentioned approach; then the method compares this FCH with the FCHs to all database images. The similarity between the two images with the FCHs FQ and FT, can be, respectively, measured by the following formula: 2

dE ðF Q ; F T Þ½F Q  F T Tn1 ½F Q  F T n1 : Ju et al. [10] employed FCM clustering algorithm to compute the FCH. The FCM clustering algorithm is used to classify the n0 fine colors to n; meanwhile, it is also used to compute the membership matrix. The algorithm minimizes the objective function Jm which is the weighted sum of squared errors within each group:

J m ðU; V; XÞ ¼

n X c X

lmik kxk  mi k2A

k¼1 j¼1

The algorithm terminates when Jm is smaller than a given threshold e. Here, V = [v1,v2,. . .,vc]T is a vector of unknown cluster prototypes, and the weighting exponent m controls the extent of membership shared by c clusters. The value of lik represents the membership of the data point xk from the set X = {x1, x2,. . .,xn} with respect to the ith cluster, and the membership matrix is represented as U = [lik], which satisfies the following expression:

lxk 2 ½0; 1;

n X

lxk ¼ 1; and

X

lxk > 0; for1 6 k 6 n; and x 2 X:

x2X

k¼1

The CCH- and FCH-based ROI image retrieval methods systematically partition each database image ID into overlapping blocks in three hierarchical levels. The highest level contains only one block, which is the image ID itself. On the middle level, ID is equally split into 3  3 overlapping rectangle blocks where the length of each side in every block is half the length of its corresponding side in ID. On the lowest level, the image is split into 5  5 equal rectangle blocks where the length of each side in a block is one-third of the length of its corresponding side in ID. For each of the 35 blocks in ID, a 64-bin FCH and a 64-bin CCH are computed respectively. In the FCH-based ROI image retrieval method, the similarity between the query region image and each block of the 35 blocks is measured by the Euclidean distance of their 64-bin FCHs. Similarly, the CCH-based ROI image retrieval method measures the similarity between the query image and each of the 35 blocks by the Euclidean distance of their 64-bin CCHs. Vu et al. [16] submits a SamMatch-based ROI image retrieval method which submits a region of interest as a query. It processes the query by sampling-based matching approach called SamMatch [9]. This method resizes each database image into 256  256, quantifies it into 256 colors, and considers the average color of the pixels in each 16  16 block to be the color of the block. Consider two arbitrary-shaped subimages Q and S, each represented by n sampled blocks. Their matching score is defined as follows [16]:

SimðQ; SÞ ¼

n X i¼1

wi 1þ

DðcSi ; cQi Þ

;

where DðcSi ; cQi Þ is the distance of cSi and cQi . cSi is the color of block i of S and cQi is that of Q. The parameter wi is a weight fac-

tor. Since SamMatch compares the corresponding sampled blocks of subimages, it involves implicitly the shape, size, and texture features of the image objects. As a result, SamMatch has the benefits of region-based techniques without reliance on the highly variable accuracy of segmentation methods. However, this occurs by having the user point out the object areas at the time of the query. SamMatch-based ROI image retrieval method takes samples of 16  16-pixel blocks at various locations in each image. The block size is chosen since the correlation between pixels tends to decrease beyond 15 to 20 pixels. It is then possible to significantly reduce storage overheads by representing each sampled block using its average color. To support general applications, SamMatchbased image retrieval method collects these blocks at uniform locations throughout the image. 2.2. The Generic algorithm The genetic algorithm (GA) [12] is a heuristic optimization method which operates through determined, randomized search. The set of possible solutions for the optimization problem is considered as a population of individuals. The degree of adaptation of an individual to its environment is specified by its fitness. The coordinates of an individual in the search space are represented by chromosomes, in essence, a set of character strings. A gene is a subsection of a chromosome which encodes the value of a single parameter being optimized. Typical encodings for a gene could be binary or integer. Genetic algorithms are derived from the evolution theory. According to the evolution theory, within a population, only the individuals well adapted to their environment can survive and transmit some of their characters to their descendants. Basically, a genetic algorithm consists of three major operations: selection, crossover, and mutation. The selection evaluates each individual and keeps only the fittest ones among them. In addition to those fittest individuals, some less fit ones could be selected according to a small probability. The others are removed from the current population. The crossover recombines two individuals to have new ones which might have a better performance. The mutation operator induces changes in a small number of chromosomes units. Its purpose is to maintain the population diversified enough during the optimization process. 2.3. ANMRR This paper will use the MPEG-7 retrieval metric NMRR (normalized modified retrieval rank) [14] to describe the effectiveness of an image retrieval method. NMRR not only indicates how many of the correct items are retrieved, but also how highly they are ranked among the retrieved items. NMRR is defined by

P  NGðqÞ RankðkÞ NMRRðqÞ ¼

k¼1

NGðqÞ

 0:5  NGðqÞ 2

KðqÞ þ 0:5  0:5  NGðqÞ

;

D

D

A C

B

A B C (b)

(a) Fig. 2. A color image and its objects.

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

where NG(q) is the size of the ground truth set for a query image q, Rank(k) is the ranking of the ground truth images by the retrieval algorithm and K(q) specifies the ‘‘relevance rank” for each query. As the size of the ground truth set is normally unequal, a suitable K(q) is determined by

KðqÞ ¼ min ð4  NGðqÞ; 2  GTMÞ; where GTM is the maximum of NG(q) for all queries. If Rank(k) > K(q), then Rank(k) is changed into K(q) + 1. The NMRR is in the range of [0, 1], and the smaller NMRR is, the better the retrieval performance will be. ANMRR is defined as the average NMRR over a range of queries, and is given by

ANMRR ¼

NQ 1 X NMRRðqÞ; NQ q¼1

where NQ is the number of query images. 3. CVAAO and CVAAO-based image retrieval method In a full color image, a pixel color is generally described by a 24bit memory space, so there are a total of 224 possible pixel colors. In the real world, a great number of images exist. Each image contains a group of large regions, and each region will have a uniform color when the pixel colors of the image are quantized down to a few representative colors. Many synthesized images like trademarks, cartoons, traffic signals, and flags, possess such a property. Segmentation of objects is very important for extracting the shape of an object on an image. However, extracting one object from an image is very difficult because of discretization, occlusions, poor contrasts, viewing conditions, noises, and etc. [17]. An image with a limited color palette is generally composed of unicolored regions. This paper views each unicolored region as an object. Fig. 2(b) shows the objects A, B, C, and D contained in the image in Fig. 2(a). Before extracting the CVAAO of an image, all pixels on database images are categorized into K clusters by K-means algorithm [15] according to the similarity of their colors. The mean value of all the pixel colors in each cluster is considered to be a color value in a color palette. The color palette containing K different colors is used as the common color palette CCP of all images (including all database images and query images). To extract the CVAAO of an image I, each pixel color C on I is replaced by one color in CCP that is more similar to C so as to create an image I0 which is as large as I, and all the colors in CCPare its possible pixel colors. This image I0 is called the color-reduced image of I. Each color Ch in CCP is given a corresponding variable vh. To compute vh, the proposed method visits each pixel on I0 beginning from the left-top pixel of I0 . It scans every pixel on I0 in the order from left to right and top to bottom. When visiting each pixel P0i;j located at the coordinates (i, j) on I0 , a color difference is computed and added to one variable vh for C h ¼ C 0i;j . Here, Pi,j is the pixel located at the coordinates (i, j) on I (we call Pi,j the corresponding pixel of P 0i;j ); C 0i;j and Ci,j are the colors of P 0i;j and Pi,j. vh can be calculated by the following statements. H and W are the height and width of I0 ; function arg (Ci,j = Ch) returns the index h of Ch when C 0i;j ¼ C h .

for i = 1 to H for j = 1 to W for (l, m) = (i, j + 1), (i + 1, j + 1), (i + 1, j), and (i + 1, j  1) if ðC 0l;m 6¼ C 0i;j Þ then h ¼ argðC 0i;j ¼ C h Þ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  2  2  2 vh ¼ vh þ Rl;m  Ri;j þ Gl;m  Gi;j þ Bl;m  Bi;j

1543

Here, (Rl,m, Gl,m, Bl,m) and (Ri,j, Gi,j, Bi,j), respectively, represent the three color components R, G, and B in Cl,m and Ci,j. The values of the K variables (v1, v2,. . .,vK) are considered to be the CVAAO of I; each vh maps to one certain color Ch in CCP. When using a tool like a scanner to input an image, the image may be enlarged or reduced because of different scanner resolution setups. We call this phenomenon a scale variation of the image. Since the CVAAO characterizes the color differences of the contour pixels between two adjacent objects, each variable vh can be divided by (H + W) to remedy the problem of scale variation. Hence, the CVAAO has great robustness in resisting the scale variation in images. The CVAAO is the color histogram of color differences between two adjacent objects in an image, so it is also indifferent to shift and rotation variations in images. Furthermore, the CVAAO can not only characterize the principle pixel colors and the texture of an image, but also distinguish the objects with inconsistent contours. Based on the CVAAO, this paper provides a CVAAO-based image retrieval method. In this method, when given a query image Iq, this method computes the CVAAO of Iq, compares the CVAAOs of Iq and each database image, and then delivers to the user the most similar database image which has the minimal image matching distance relative to Iq.     Let vq1 ; vq2 ; . . . ; vqK and vd1 ; vd2 ; . . . ; vdK be the CVAAOs of the query image Iq and one database image Id. One can use the following formula to compute the image matching distance D between Iq and Id:

D¼

K  X  vq  vd  i i

ð1Þ

i¼1

The smaller D is, the more Iq is similar to Id. 4. The CVAAO-based ROI image retrieval method Based on the CVAAO, this paper also presents a ROI image retrieval method, called CVAAO-based ROI image retrieval method. The method contains two aspects: database creating and image querying. This section describes these two aspects in detail. In this section, a generic algorithm will be provided to decide the most suitable parameters for the CVAAO-based ROI image retrieval method as well. 4.1. Database creating Let CCP1 be a common palette with K1 colors. The CVAAO-based ROI image retrieval method first transforms each database image ID into a color-reduced image I0 D containing only the colors in CCP1. Each set of the adjacent pixels with the same color C in I0 D constructs an object O. The number of pixels in O is the area of O, and C is the color of O. However, if the area of O is smaller than a given threshold THA, the method will regard O as a noise and ignore it. Additionally, the CVAAO-based ROI image retrieval method categorizes the directions from 0° to 360° into eight directions 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5° denoted by h0, h1,. . .,h7 respectively. For each object O, the method gives a minimal bounding rectangle (MBR) containing O, where the rectangle is oriented parallel to the X and Y axes. The method draws eight lines through the central pixel of O0 s MBR where the included angles between the eight lines and the X-axis are 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5 ° respectively. Fig. 3 shows the eight lines. Let Li be the i-th one of the eight lines, and P1, P2,. . .,Pe be the intersections of Li and the boundary of O. We define li as the length of Li, where li is the maximal distance between any two of P1, P2,. . .,Pe. For each object, the CVAAO-based ROI image retrieval

1544

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

6

5

4

3

7

2

1

0

iation, the CVAAO-based ROI image retrieval method adopts the included angle between OM and OC to be the included angle of RQ and the database image ID, and rotates one of them, so that they can be put in the same direction. Let AC be the area of OC, (xC, yC) be the central pixel coordinates C C C of the MBR of OC, and l0 , l1 , . . ., l7 be the lengths of OC in the directions 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°, respectively. The CVAAO-based ROI image retrieval method repeatedly rotates OC by (22.5  i)° for i = 0–7; each time it computes the difference between the lengths of the eight directions of OM and OC by the following formula:

di ¼

7  2 X C M lðiþjÞmod8  lj ;

for i ¼ 0 to 7:

j¼0 7

Fig. 3. Eight directions of an object.

method records its color, area, coordinates of the left-up and rightdown pixels of the MBR, and l0, l1,. . .,l7 in database. The area and l0, l1,. . .,l7 of an object can depict the shape of the object. 4.2. Image querying aspect In a ROI image retrieval method, a user is interested in only one region of a database image. Given an HQ  WQ query region image RQ, the CVAAO-based ROI image retrieval method first detaches all possible target region images from the database images. We named all the possible target region images as candidate region images. Then, this method computes the matching distance between each candidate region image and RQ, and takes the candidate region image with the minimal matching distance as the target region image RT. The image querying aspect contains two stages – candidate region image segmenting stage and region image matching stage. The purpose of the candidate region image segmenting stage is to sever the candidate region images from database images. In the region image matching stage, each candidate region image is compared with RQ, and the database image on which the target region image RT is located will be returned. This subsection introduces these two stages in detail. 4.2.1. The candidate region image segmenting stage Between two similar images, there generally exist some large objects with similar colors and shapes. Hence, the CVAAO-based ROI image retrieval method uses the object OM with the maximal area in RQ to determine the sizes and positions of the candidate regions on a database image. The CVAAO-based ROI image retrieval method first transforms RQ into a color-reduced image R0Q which only contains the K1 colors in CCP1. Then, the method cuts off the maximal object OM from R0Q and decides the color, area, the lengths M M M l0 , l1 , . . ., l7 , and the coordinates (xM, yM) of the central pixel on the MBR of OM. When a query region image is severed from a query image, the border district of the query region image may comprise some objects which are in expectation but partial. Therefore, the CVAAObased ROI image retrieval method uses the maximal object in R0Q without containing the boundary pixels of R0Q as OM. For example, the object D in Fig. 2 cannot be OM. Then the CVAAO-based ROI image retrieval method extracts all the objects OC from the database images, of which the colors are the same as of OM. The region imageRC including OC may be the target region image which we call a candidate region image. When one takes pictures, the lens may be adjusted to different positions, so for the same scene, the images with scale, rotation, and shift variations may be generated. To remedy the rotation var-

7

For h ¼ argðminðdi ÞÞ where argðminðdi ÞÞ returns the index i of i¼0 i¼0 the minimal di, hh = 22.5  h is the estimated included angle between OM and OC; in other words, hh is regarded as the included angle ofID and RQ. The CVAAO-based ROI image retrieval method then rotates ID around (xC, yC) through hh° to generate the image I00D , so that the direction of OM in I00D is close to the direction of OC. Consider a pixel P at coordinates (x, y) on ID. P00 is the corresponding pixel of P at coordinates (x00 , y00 ) on I00D . The relationship between (x, y) and (x00 , y00 ) is described as follows:



x00 ¼ ðx  xC Þ sin hh þ ðy  yC Þcoshh ; and y00 ¼ ðx  xC Þcoshh þ ðy  yC Þsinhh :

To reduce the incidence of scale variations, the CVAAO-based ROI image retrieval method takes the ratio of the areas of OC to OM to estimate the ratio of sizes of RC to RQ. Let RQ consist of qffiffiffiffiffi  HQ  WQ pixels. Accordingly, RC comprises HQ  AAMC  ðW Q  qffiffiffiffiffi AC Þ pixels. The upper-left and lower-right pixels of RC on I00D are qffiffiffiffiffi qffiffiffiffiffi qffiffiffiffiffi  AM   and ðW Q  xM Þ  AAMC þ xC ; xM  AAMC þ xC ; yM  AAMC þ yC qffiffiffiffiffi ðHQ  yM Þ  AAMC þ yC Þ, respectively. 4.2.2. The region image matching stage In this stage, from the candidate region imagesRC, the CVAAObased ROI image retrieval method intends to find out the target region image RT which is more similar to RQ, and to deliver the database image on which RT is located to the user. The method also transforms RQ and each RC into the color-reduced images containing only K2 colors in the other common color palette CCP2. Via these color-reduced images, the method extracts the CVAAOs of RQ and each RC. By comparing the CVAAOs, the matching distance between RQ and each RC can be obtained by formula (1). Finally, the method submits to the user the database image containing the candidate region image with the minimal matching distance. In this method, two different common color palettes CCP1 and CCP2 are used. CCP1 with K1 different colors is employed in detaching the candidate region image. The other common color palette CCP2, consisting of K2 different colors, is employed to extract the CVAAO feature of each region image. In order to get intact large objects in an image, a smaller K1 is required; otherwise, the large objects in some images may break into pieces so that the CVAAObased ROI image retrieval method may obtain erroneous OM. However, to compute a more precise matching distance between RQ and each RC, one needs to adopt a larger K2. 4.3. Suitable parameters decision The performance of the CVAAO-based ROI image retrieval method is deeply affected by the values of K1, K2, and THA. This

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

paper will employ a genetic algorithm to figure out the most suitable values for K1, K2, and THA. The CVAAO-based ROI image retrieval method makes use of a binary string, created by concatenating three binary substrings sk1 , sk2 , and sTHA , to represent a chromosome Ch where sk1 , sk2 , and sTHA encode a set of K1, K2, and THA. In this chromosome, the values of K1 (resp. K2 and THA) are calculated with pK 1 (resp. pK 2 and pTHA Þ multiplied by the number of 1 bits in sk1 (resp. sk2 , and sTHA Þ. For a certain application, one can accumulate some of its historic images (including database images and query images). After that, he can apply the accumulated historic images to train the most appropriate values for K1, K2, and THA via the generic algorithm. The fitness of a chromosome in this generic algorithm is defined as the ANMRR of querying by the CVAAO-based ROI image retrieval method based on the accumulated historic images and the values of K1, K2, and THA encoded by the chromosome. While running this genetic algorithm, the CVAAO-based ROI image retrieval method first generates N chromosomes at random, each containing three substrings sk1 , sk2 , and sTHA . The generic algorithm evolves the fittest solution by repeatedly executing mutation, crossover, and selection operations; in each time, only N chromosomes with the optimal fitness values are kept until the related fitness of the N reserved chromosomes are quite similar to one another. In the mutation operation, for each of N reserved chromosomes, the CTBIR system uses a random number generator to randomly specify three bits b1,b2, and b3, respectively, from sk1 , sk2 , and sTHA in the chromosome. Afterward, b1, b2, and b3 are replaced respectively by :1, :2, and :3 to generate a new chromosome, where : means the operator of ‘‘not”. In the crossover operation, the CTBIR system similarly uses a random number generator to randomly designate N0 pairs of chromosomes on the N reserved chromosomes. Let Ch[i. . .j] be the substring from the ith to jth bits in a chromosome Ch. For each pair of chromosomes Ch1and Ch2, the genetic algorithm concatenates 



         jsK j jsK 1 j sK  , Ch1 sK  þ 1 ... sK þ , Ch2 þ 1 ... Ch1 1... 21 1 1 1 2



             s s j K2 j jK j , Ch2 sK 1  þ 21 þ 1 . . . sK 1  þ sK 2  , Ch1 sK 1 þ 2

             sK  þ 1 . . . sK  þ sK  þ jsTHA j sK  þ sK þ , and Ch 1 2 1 2 1 2 2



           sTH j jsTHA j j A . . . sK 1 þ sK 2  þ . . . sK 1  þ sK 2  þ sTHA   into a 2 2 



 jsK j jsK 1 j ; Ch1 þ 1 ... new chromosome, and Ch2 1 . . . 21 2 



 

          sK  ; Ch2 sK  þ 1 . . . sK  þ jsK 2 j sK  þ jsK 1 j þ 1 ; Ch 1 1 1 1 1 2 2 



              jsTHA j ; ... sK 1  þ sK 2  ;Ch2 sK 1  þ sK 2  þ 1 ... sK 1  þ sK 2 þ 2 



           jsTH j jsTHA j ... sK 1  þ sK 2 þ ... sK 1 þ and Ch1 sK 1  þ sK 2  þ 2 A 2          sK  þ sTH  into other new chromosome, where sk1 , sk2 , and 2 A   sTH  are, respectively, the sizes of sk , sk , and sTH . A A 1 2 In the selection operation, the N optimal chromosomes are selected from the N chromosomes reserved in previous iteration, N and 2  N0 chromosomes created in the mutation and crossover operations according to their fitness. This genetic algorithm continuously performs the three operations – mutation, crossover, and selection until the related fitness of the reserved N chromosomes is very close. Finally, K1 and THA are respectively considered to be p  K 1  and pTHA  multiplied    by the   numbers    of 1-bits in Ch½1 . . . sK 1  and Ch½sK 1  þ sK 2 þ 1 . . . sK 1  +sK 2  þ sTHA , and K2   is  pTH A multiplied    by ð sK 1 þ the numbers of 1  bits in Ch½sK 1  þ 1 . . . sK 1  þ sK 2 Þ of the chromosome Ch with the best fitness in the N reserved chromosomes.

1545

5. Experiments The aims of this section is to investigate the performances of the CVAAO-based image retrieval method and the CVAAO-based ROI image retrieval method, and compare their performances with those of the FCH-, CCH-, SamMatch-based image retrieval methods as well as the FCH-, CCH-, SamMatch-based ROI image retrieval methods. This section will also explore their robustness in resisting the variances of images. 5.1. Performances of CVAAO-based and CVAAO-based ROI image retrieval methods     q d Let SetD ¼ f1d ; f2d ; f3d ; . . . ; f1000 and SetQ ¼ f1q ; f2q ; f3q ; . . . ; f1000 be two image sets, each of which contains 1000 full color images. The images in SetD are employed as the database images and those in SetQ are used as the query images. Some parts of them are drawn out from animations, where each image pair ðfid ; fiq Þ are randomly picked up from one same animation. The rest of images were scanned from natural images and trademark pictures. Fig. 4 shows some of the testing images. The first experiment is to test the performance of the CVAAObased image retrieval method in different K. At first, the images q d ff1d ; f2d ; . . . ; f100 g and ff1q ; f2q ; . . . ; f100 g are respectively specified to be the database images and query images, and different K is applied in the CVAAO-based image retrieval method. Table 1 lists the ANMRR obtained by the CVAAO-based image retrieval method with different K. The experimental results tell that the CVAAO-based image retrieval method a better given  can provide  performanceq when  d K = 21. Then, f1d ; f2d ; f3d ; . . . ; f1000 and f1q ; f2q ; f3q ; . . . ; f1000 are taken to be the database images and query images, and the CVAAO-, FCH-, CCH-, and SamMatch-based image retrieval methods are employed to answer the queries. Here, the K is set to 21 in the CVAAObased image retrieval method. Table 2 displays the experimental results. The second experiment is to investigate the performance of the CVAAO-based ROI image In this experiment,  retrieval method.  q from each of images in f1q ; f2q ; f3q ; . . . ; f1000 , a clip is manually cut out as the query region image. Fig. 5 shows three of the clips and their related fiq . The values of K1, K2, and THA significantly influence the performance of the CVAAO-based ROI image retrieval method. q d In this experiment, the images ff1d ; f2d ; . . . ; f100 g and ff1q ; f2q ; . . . ; f100 g are respectively specified to be the database images and the query images first, and the generic algorithm is used to determine the    most suitable values of K1, K2, and THA, where sk1 , sk2 , sTHA ;pK 1 , pK 2 , and pTHA are set to be 30, 40, 40, 1, 1, and 10. This experimental results show that the most suitable values of K1, K2, and THA are 14, 26, and 180.  d d d  d After that, the images f1 ; f2 ; f3 ; . . . ; f1000 and  q q q  q f1 ; f2 ; f3 ; . . . ; f1000 are used as the database images and the query images. Based on the most suitable values of K1, K2, and THA, this experiment adopts the CVAAO-based ROI image retrieval method to reply to the queries; meanwhile, the FCH-, CCH-, and SamMatch-based ROI image retrieval methods are also employed to respond to the queries. Table 3 demonstrates the experimental results. In this paper, the parameters n0 , n, and m used in the FCH-based image retrieval method and the FCHbased ROI image retrieval method are set to be 4,096, 64, and 1.9. The SamMatch-based ROI image retrieval method is very sensitive to the rotation and shift variations. In addition, its segmentation cannot precisely cut out the target region image. Hence, in the first experiment, the SamMatch-based ROI image retrieval method can offer an acceptable result, but in this experiment, the SamMatch-based ROI image retrieval method cannot give a

1546

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

Fig. 4. Partial testing images.

Table 1 The ANMRR obtained by the CVAAO-based image retrieval method with different K K

12

15

18

21

24

27

ANMRR

0.065

0.065

0.07

0.06

0.065

0.065

Table 2 The ANMRR obtained by the CVAAO-, FCH-, CCH-, and SamMatch-based image retrieval methods CVAAO

CCH

FCH

SamMatch

0.11

0.15

0.14

0.21 Fig. 5. Some of the clips and their related fiq .

good ANMRR. Table 3 also tells that the CVAAO-based ROI image retrieval method provides a better ANMRR than the CCH- and FCH-based ROI image retrieval methods. 5.2. The robustness in resisting the variations of images When one takes pictures, the lens may be adjusted to different positions and different illuminations may be used, so that for the same scene, the images with scale, rotation, hue, contrast, light, and noise variations may be generated. In the real world, the distortion variation is a very common phenomenon too. For example, putting a picture on an uneven plane may deform the objects on the picture. An excellent image retrieval method should be insensitive to these variations. The next three experiments are designed to explore the capacities of the CVAAO-based image retrieval method and the CVAAO-based ROI image retrieval method for

resisting the variations of rotation, distortion, noise, scale, hue, luminance, and contrast. In the next experiments, 100 full color images Iq,1, Iq,2,. . .,Iq,100 are used as the query images. These 100 full color images comprise an image set Sq. Besides, this paper employs the rotation, distortion, noise, scale, hue, luminance, and contrast functions in ADOBE PHOTOSHOP 7.0 to process each Iq,i, and, respectively, generates the variant images Ir,i, Id,i, In,i, Is,i, Ih,i, Il,i, and Ic,i. The group of images Ia,1, Ia,2,. . ., Ia,100 forms an image setSa, which is produced by the Table 3 The ANMRR obtained by the CVAAO-, FCH-, CCH-, and SamMatch-based ROI image retrieval methods CVAAO

CCH

FCH

SamMatch

0.32

0.36

0.34

0.97

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

same function where a=r, d, n, s, h, l, and c. Fig. 6 shows some images in Sq, Sr, Sd, Sn, Ss, Sh, Sl, and Sc respectively, where i is the image number in Sq, Sr, Sd, Sn, Ss, Sh, Sl, and Sc. In the next experiments, the values of K1, K2, and ThA are similarly set to be 14, 26, and 180. In the third experiment, all the images in Sr, Sd, Sn, Ss, Sh, Sl, and Sc are employed as the database images. In the database, the CVAAO, CCH, and FCH of each whole image are held in advance. For the ith query, the CVAAO, CCH, and FCH of the whole image of Iq,i are extracted. Based on the CVAAO, CCH, and FCH, the image matching distance of each database image and Iq,i can be determined. The image retrieval method individually submits to the user 14 database images with the minimal image matching distances relative to Iq,i. In this experiment, K(q), NG(q), GTM, and NQ are respectively given to be 14, 7, 7, and 100. The ANMRRtotal in Table 4 is the accuracy (ANMRR) which the experiment obtains. Furthermore, ANMRRr, ANMRRd, ANMRRn, ANMRRs, ANMRRh, ANMRRl, and ANMRRc in Table 4 are the accuracies (ANMRR) obtained by the experiment which takes the images in Sr, Sd, Sn, Ss, Sh, Sl, and Sc respectively as the database images, and K (q),

1547

NG(q), GTM, and NQ are set to be 2, 1, 1, and 100. These experiments compare the features of the whole query image and each whole database image too. In the experiments, the method based on the CVAAO needs to store the CVAAO of the whole database images in the database. Each database image owns one 21-dimension CVAAO, and each dimension of the CVAAO occupies a 2-byte memory space. Therefore, the method based on the CVAAO consumes 29,400 (=700  21  2) bytes to store the CVAAO of all the 700 database images. The method based on the CCH has to store the CCH of all the database images in the database. Each database image has one 64-dimension CCH, and every dimension of the CCH requires a 2-byte memory space. In this experiment, the method based on the CCH hence takes 89,600 (=700  64  2) bytes to keep the CCH of all the 700 database images. An FCH includes 64 dimensions, each of which is kept in a 2- byte memory space. Thus, the FCH-based image retrieval method takes 89,600 bytes memory space to hold the FCHs of all the

Fig. 6. The ith images in sets Sq, Sr, Sd, Sn, Ss, Sh, Sl, and Sc.

1548

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549

Table 4 The ANMRRs of whole image retrieval

ANMRRtotal ANMRRr ANMRRd ANMRRn ANMRRs ANMRRh ANMRRl ANMRRc

CVAAO

CCH

FCH

SamMatch

0.09 0.36 0.09 0.11 0.01 0.16 0.61 0.57

0.35 0.31 0.16 0.48 0.25 0.51 0.53 0.52

0.25 0.63 0.02 0.15 0.65 0.11 0.41 0.13

0.28 0.95 0.44 0.21 0.28 0.11 0.25 0.02

database images. However, in this method, each pixel color in a database image is mapped to one of 4096 histogram bins, and every histogram bin is classified to one of 64 clusters. We call the average pixel color in one of the 4096 histogram bins a fine color, and the average pixel color in each of the 64 clusters is named as a coarse color. This method takes up a 2-byte memory space to express each color. The method also needs to maintain these extra data. Therefore, the memory space required to hold the image features of all the database images for this method is 89,600+4,096  64  2 = 613,888 bytes. The SamMatch-based image retrieval method resizes each image to 256  256 pixels, reduces all pixel colors in the image into only 256 colors, divides the image into 16  16 blocks, and uses the average colors of the blocks as the feature of the image. Hence, 113 samples were evenly spread out in each image, and each dimension is held by a 2-byte memory space. It takes 158,200 (=700  113  2) bytes to hold the features of all database images. Since the CVAAO-based image retrieval method normalizes the image features (dividing by H + W), the experimental results also indicate that it is indifferent to the scale variation of images. Besides, the CVAAO-based image retrieval method can resist the distortion variation in images. After an object is processed by distortion operation, its shape is changed a lot, but the area of each color in the image is almost unchanged. Therefore, the CVAAO-, CCH-, and FCH-based image retrieval methods perform well for the images with distortion variation. In the real world, the distortion variation is a very common phenomenon. For example, putting a picture on an uneven plane may deform the objects on the image. When one takes a picture with a focus on certain objects, the background in the picture is generally different if the lenses are adjusted to different positions. Fig. 7 shows one of these image pairs. The CVAAO is hence susceptible to a great rotation variation in images. For two identical images with different hues, any two pixels located in the same position of the two images may respectively correspond to different colors in the common color palette. Similarly, for the images with large luminance and contrast variations, since the colors of the images are changed a lot, their CVAAOs are also significantly altered. Especially, human eyes are highly sensitive to the variations of luminance and contrast in images. The CVAAO is hence more sensitive to the luminance and contrast variations. The experimental result also shows that the SamMatchbased image retrieval method is insensitive to hue and contrast variations. However, the SamMatch-based image retrieval method can’t resist rotation and distortion variations because block positions are greatly changed. The next experiment is to scrutinize the capacity of the CVAAO-, CCH-, FCH- and SamMatch-based ROI image retrieval methods for precisely cutting off the target region image. This experiment takes the images in Sq as the query images; meanwhile, the same images are also being used as the database images. In this experiment, the user utilizes a mouse to extract the query region image from a query image as a query region image. In Fig. 6, the image Rq,i is the query region image severed from the query image Iq,i. The methods then search for the target region image more similar to

Fig. 7. The image pair with rotation variant.

the query region image and submit to the user only one database image on which the target region image is located. If the submitted database image is a related query image, we can state that the methods correctly reply to this query. In this experiment, the CVAAO-based ROI image retrieval method can accurately answer each query. However, the CCH-, FCH-, and SamMatch-based ROI image retrieval methods only obtain 29, 24, and 16 correct answers for the 100 queries. The reason is that the segmenting method, which the CCH-, FCH-, and SamMatch-based ROI image retrieval methods adopt, cannot precisely detach the target region image from the database image. The next experiment is to investigate the capacities of the CVAAO-, CCH-, FCH-, and SamMatch-based ROI image retrieval methods for precisely retrieving the database images which satisfy the users’ requirement. In this experiment, all the images in Sr, Sd, Sn, Ss, Sh, Sl, and Sc are used as the database images first. Similarly, in each query, the user applies a mouse to draw out an image region from a query image as the query region image. In the experiment, K(q), NG (q), GTM, and NQ are, respectively, given to be 14, 7, 7, and 700. The ANMRRtotal in Table 5 states the results of this experiment. Moreover, ANMRRr, ANMRRd, ANMRRn, ANMRRs, ANMRRh, ANMRRl, and ANMRRc in Table 5 are the ANMRR obtained by the experiments which respectively take the images in Sr, Sd, Sn, Ss, Sh, Sl, and Sc as the database images, and K(q), NG(q), GTM, and NQ are set to be 2, 1, 1, and 100. The experimental results indicate that the CVAAO-based ROI image retrieval method totally performs better than the CCH-, FCH-, and SamMatch-based ROI image retrieval methods since none of the latter three methods can precisely cut off the target region image. The CCH-, FCH-, and SamMatch-based ROI image retrieval methods partition each database image into some overlapping blocks. When a database image is rotated, an overlapping block on the rotated image may hold only a part of the target region image or some extra regions which does not belong to the target region image. Therefore, in these experiments, the ANMRRr are only 0.35, 0.41, and 0.83 for the CCH-, FCH-, and SamMatchbased ROI image retrieval methods. The CVAAO-based ROI image retrieval method can precisely detect the included angle between the target region image and the query region image. Hence, the CVAAO-based ROI image retrieval method is indifferent to the rotation variation in images. It gets ANMRRr = 0.03. For the images with the distortion variation, the shapes of their objects may change, but the colors of the objects and color variances among adjacent objects are invariant. Therefore, the CVAAO-based ROI image retrieval method still acquires that ANMRRd = 0.24. Similarly, in the distorted image, the overlapping block may only hold a part of the target region image. Thus, the CCH-, FCH-, and SamMatch-based ROI image retrieval methods are highly sensitive to the distortion variation in images. The ANMRRd which the CCH- and FCH-, and SamMatch-based ROI image retrieval methods obtain are 0.69, 0.68, and 0.91. When some noises are added to an image, its objects may be segmented into fragments. The CVAAO-based ROI image retrieval method adopts a smaller K1 to abate the shattered phenomenon; hence its

Y.-K. Chan et al. / Image and Vision Computing 26 (2008) 1540–1549 Table 5 The ANMRRs in the four experiments

ANMRRtotal ANMRRr ANMRRd ANMRRn ANMRRs ANMRRh ANMRRl ANMRRc

1549

6. Conclusions

CVAAO

CCH

FCH

SamMatch

0.10 0.03 0.24 0.22 0.10 0.10 0.13 0.12

0.40 0.35 0.69 0.64 0.33 0.32 0.38 0.41

0.43 0.41 0.68 0.59 0.38 0.35 0.50 0.45

0.88 0.83 0.91 0.90 0.92 0.91 0.94 0.86

ANMRRn = 0.22. In the CVAAO-based ROI image retrieval method, if the hue of an image is converted, the pixels in this image may be assigned to different colors in a common color palette from those to which the pixels belong in its original image. Its ANMRRh is 0.10. Similarly, the CCH- and FCH-based ROI image retrieval methods have the same problem; their ANMRRh are 0.35 and 0.32 respectively. For the luminance and contrast variations in images, a pixel in the images may be specified to a color bin different from the color bin to which the pixel in the original image is specified. Therefore, the CVAAO-, CCH- and FCH-based ROI image retrieval methods are seriously vulnerable to the luminance and contrast variations. In the experiments, the CVAAO-based ROI image retrieval method stored in the database the area (2 bytes/object), color (1 bytes/ object), upper-left and lower-right coordinates (4  2 bytes/object) of MBR, and eight direction lengths (8  2 bytes/object) of each object on database images. In these experiments, each image in Sr, Sd, Sn, Ss, Sh, Sl, and Sc mostly contains about 8–12 objects. The method takes about 21,600 to 32,400 bytes of memory space to hold the CVAAOs of the images in each of the sets Sr, Sd, Sn, Ss, Sh, Sl, and Sc. For CCH and FCH, every database image is divided into 35 overlapping blocks, and 64 color bins are extracted from each block. Here, every color bin occupies 2 bytes. In the CCH-based ROI image retrieval method, each set of database images in Sr, Sd, Sn, Ss, Sh, Sl, and Sc holds totally 448,000 bytes memory space. The FCH-based ROI image retrieval method consumes 448,000 bytes to keep the FCHs of the images in each database image set, too. However, this method still has to save 4,096fine colors for each of the 64 coarse color bins. Here, each color takes up a 2-byte memory space. Hence, the method consumes 448,000bytes to store the FCHs of each set of database images. Besides, the method uses extra 4096  64  2 = 524,288 bytes of memory space to hold the 4096fine colors. The SamMatchbased image retrieval method is still a 113-dimension feature for each image, and each dimension is held by a 2-byte memory space. It consumes 158,200 (=700  113  2) bytes to hold the features of all database images. However, since the CVAAO-based ROI image retrieval method extracts the CVAAOs of the database images in image querying, it consumes much time on query process compared to the CCH-, FCH-, SamMatch-based ROI image retrieval methods. In this experiment, the CVAAO-based ROI image retrieval method takes 2074.64 s to process 700 queries; the CCH-, FCH-, and SamMatch-based ROI image retrieval methods respectively spends 927.50, 1435.00, and 1245.02 s on executing the 700 queries.

This paper provides a CVAAO-based image retrieval method and a CVAAO-based ROI image retrieval method. The CVAAO feature can not only describe the principle pixel colors and color complexity of an image, but also distinguish the objects with inconsistent contours. Furthermore, it is insensitive to the scale, shift, rotation, and distortion variations. Concerning the image querying aspect, the CVAAO-based ROI image retrieval method computes the location and size of the target region image RT on a database image via the shape, area, and position of OM on the query region image RQ, where RT is more similar to RQ. The experimental results demonstrate that this segmentation method can precisely sever the target region image from the database image compared to the segmentation adopted by CCH-, FCH-, and SamMatch-based ROI image retrieval methods. Besides, the CVAAO-based ROI image retrieval method is can not only accurately and efficiently find out the desired database images but also require less memory space for storing the features of database images. However, it takes more querying time than the CCH-, FCH-, and the SamMatch-based ROI image retrieval methods. References [1] J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum Press, Kluwer Academic/Plenum Publishers, New York, 1981. [2] R. Brunelli, O. Mich, Histograms analysis for image retrieval, Pattern Recognition 34 (8) (2001) 1625–1637. [3] A. Dimai, Differences of global features for region indexing, The Technical Report 177 of Swiss Federal Institute of Technology Lausanne, 1997. [4] J.C. Dunn, A fuzzy relative of the ISODATA process and its use in detecting compact, well separated clusters, Journal Cybernet 3 (3) (1974) 32–57. [5] J.M. Fuertes, M. Lucena, N. Pérez de la Blanca, J. Chamorro-Martı´nez, A scheme of colour image retrieval from databases, Pattern Recognition 22 (3–4) (2001) 323–337. [6] T. Gevers, H. Stokman, Robust histogram construction from color invariants, IEEE Transactions on Pattern Analysis and Machine Intelligence 26 (1) (2001) 113–117. [7] D. Guillamet, J. Vitria, A comparison of global versus local color histograms for object recognition, in: The Proceedings of 15th International Conference on Pattern Recognition, vol. 2, Barcelona, Spain, September 2000, pp. 422–425. [8] B. Hill, T. Roger, F.W. Vorhagen, Comparative analysis of the quantization of color spaces on the basis of the CIELAB color-difference formula, ACM Transaction on Graphics 16 (2) (1997) 109–154. [9] K.A. Hua, K. Vu, and J.H. Oh, Sammatch: a flexible and efficient sampling-based image retrieval technique for large image database, in: The Proceedings of ACM International Conferences on Multimedia, October 1999, pp. 225–234. [10] H. Ju, K.K. Ma, Fuzzy color histogram and its use in color image retrieval, IEEE Transactions on Image Processing 11 (8) (2002) 944–952. [11] M.S. Kankanhalli, B.M. Mehtre, H.Y. Huang, Color and spatial feature for content-based image retrieval, Pattern Recognition 22 (3–4) (2001) 323–337. [12] K.F. Man, K.S. Tang, S. Kwong, Genetic Algorithms: Concepts and Designs, Springer-Verlag, New York, 1999. [13] B.S. Manjunath, J.R. Ohm, V.V. Vasudevan, A. Yamada, Color and texture descriptors, IEEE Transactions on Circuits and Systems for Video Technology 11 (6) (2001) 703–715. [14] R.O. Stehling, M.A. Nascimento, A.X. Falcao, An Adaptive and efficient clustering-based approach for content-based image retrieval in image databases, in: The Proceedings of International Symposium on Database Engineering & Application, Grenoble, France, July 2001, pp. 356–365. [15] M.C. Su, C.H. Chou, A modified version of the K-means algorithm with a distance based on cluster symmetry, IEEE Transactions on Pattern Analysis and Machine Intelligence 23 (6) (2001) 674–680. [16] K. Vu, K.A. Hua, W. Tavanapong, Image retrieval based on regions of interest, IEEE Transactions on Knowledge and Data Engineering 15 (4) (2003) 1045– 1049. [17] J.Z. Wang, Y.P. Du, A scalable integrated region image retrieval system, in: The Proceedings of international conference on Image Processing, vol. 1, Thessaloniki, Greece, June 2001, pp. 22–25.