(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 2, February 2011
Content Based Image Retrieval using Dominant Colorand Texture features
M.Babu Rao Dr.B.Prabhakara Rao Dr.A.Govardhan
Associate professor, CSE department Professor&Director of Evaluation Professor&PrincipalGudlavalleru Engineering College JNTUK JNTUH college of EngineeringGudlavalleru, Krishna (Dist.), A.P, India Kakinada, A.P, India Jagtial, A.P, Indiababuraompd@yahoo.co.in
Abstract
Nowadays people are interested in using digitalimages. So the size of the image database is increasingenormously. Lot of interest is paid to find images in the database.There is a great need for developing an efficient technique forfinding the images. In order to find an image, image has to berepresented with certain features. Color and texture are twoimportant visual features of an image. In this paper we propose anefficient image retrieval technique which uses dominant color andtexture features of an image. An image is uniformly divided into 8coarse partitions as a first step. After the above coarse partition,the centroid of each partition (“color Bin” in MPEG-7) is selectedas its dominant color. Texture of an image is obtained by usingGray Level Co-occurrence Matrix (GLCM). Color and texturefeatures are normalized. Weighted Euclidean distance of colorand texture features is used in retrieving the similar images. Theefficiency of the method is demonstrated with the results.
Keywords- Image retrieval, dominant color, Gray level co-occurrence matrix.
I.
INTRODUCTION
Content-based image retrieval (CBIR) [1] has become aprominent research topic because of the proliferation of videoand image data in digital form. Increased bandwidthavailability to access the internet in the near future will allowthe users to search for and browse through video and imagedatabases located at remote sites. Therefore fast retrieval of images from large databases is an important problem that needsto be addressed.Image retrieval systems attempt to search through adatabase to find images that are perceptually similar to a queryimage. CBIR is an important alternative and complement totraditional text-based image searching and can greatly enhancethe accuracy of the information being returned. It aims todevelop an efficient visual-Content-based technique to search,browse and retrieve relevant images from large-scale digitalimage collections. Most proposed CBIR [2,3,4] techniquesautomatically extract low-level features (e.g. color, texture,shapes and layout of objects) to measure the similaritiesamong images by comparing the feature differences.Color is one of the most widely used low-level visualfeatures and is invariant to image size and orientation[1]. Asconventional color features used in CBIR, there are colorhistogram, color correlogram, and dominant color descriptor(DCD).Color histogram is the most commonly used colorrepresentation, but it does not include any spatial information.Color correlogram describes the probability of finding colorpairs at a fixed pixel distance and provides spatial information.Therefore color correlogram yields better retrieval accuracy incomparison to color histogram. Color autocorrelogram is asubset of color correlogram, which captures the spatialcorrelation between identical colors only. Since it providessignificant computational benefits over color correlogram, it ismore suitable for image retrieval. DCD is MPEG-7 colordescriptors[4]. DCD describes the salient color distributionsin an image or a region of interest, and provides an effective,compact, and intuitive representation of colors presented in animage. However, DCD similarity matching does not fit humanperception very well, and it will cause incorrect ranks forimages with similar color distribution[5, 6]. In [7], Yang et al.presented a color quantization method for dominant colorextraction, called the linear block algorithm (LBA), and it hasbeen shown that LBA is efficient in color quantization andcomputation. For the purpose of effectively retrieving moresimilar images from the digital image databases (DBs), Lu etal.[8] uses the color distributions, the mean value and the
standard deviation, to represent the global characteristics of the image, and the image bitmap is used to represent the localcharacteristics of the image for increasing the accuracy of theretrieval system. In [3,12] HSV color and GLCM texture are used as featuredescriptors of an image. Here HSV color space is quantizedwith non-equal intervals. H is quantized into 8-bins, S into 3-bins and v into 3-bins. So color is represented with onedimensional vector of size 72 (8X3X3). Instead of using 72color feature values to represent color of an image, it is betterto use compact representation of the feature vector. Forsimplicity and with out loss of generality the RGB color spaceis used in this paper.Texture is also an important visual feature that refers toinnate surface properties of an object and their relationship tothe surrounding environment. Many objects in an image can bedistinguished solely by their textures without any otherinformation. There is no universal definition of texture. Texture
118 http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 2, February 2011
may consist of some basic primitives, and may also describethe structural arrangement of a region and the relationship of the surrounding regions [5]. In our approach we have used thetexture features using gray-level co-occurrence matrix(GLCM).Our proposed CBIR system is based on Dominant color[21] and GLCM [17] texture. But there is a focus on globalfeatures. Because Low level visual features of the images suchas color and texture are especially useful to represent and tocompare images automatically. In the concrete selection of color and texture description, we use dominant colors, Gray-level co-occurrence matrix. The rest of the paper is organizedas follows. The section II outlines proposed method in termsof Algorithm. The section III deals with experimental setup.The section IV presents results. The section V presentsconclusions.II.
PROPOSED
METHODOnly simple features of image information can not getcomprehensive description of image content. We consider thecolor and texture features combining not only be able toexpress more image information, but also to describe imagefrom the different aspects for more detailed information inorder to obtain better search results. The proposed methodis based on dominant color and texture features of image.Retrieval algorithm is as follows:
Step1
: Uniformly divide each image in the database and thetarget image into 8-coarse partitionsas shown in Fig.1.
Step2
: For each partition,the centroid of each partition isselected as its dominant color.
Step3
: Obtain texture features (Energy, Contrast, Entropy andinverse difference) from GLCM.
Step4
: construct a combined feature vector for color andtexture.
Step5
: find the distances between feature vector of queryimage and the feature vectors of target images using weightedand normalized Euclidean distance.
Step6
: sort the Euclidean distances.
Step7
: retrieve first 20 most similar images with minimumdistance
A.
Color feature representation
In general, color is one of the most dominant anddistinguishable low-level visual features in describing image.Many CBIR systems employ color to retrieve images, such asQBIC system and Visual SEEK. In theory, it will lead tominimum error by extracting color feature for retrieval usingreal color image directly, but the problem is that thecomputation cost and storage required will expand rapidly. Soit goes against practical application. In fact, for a given colorimage, the number of actual colors only occupies a smallproportion of the total number of colors in the whole colorspace, and further observation shows that some dominantcolors cover a majority of pixels. Consequently, it won'tinfluence the understanding of image content though reducingthe quality of image if we use these dominant colors torepresent image.In the MPEG-7 Final Committee Draft, several colordescriptors have been approved including number of histogram descriptors and a dominant color descriptor (DCD)[4, 6]. DCD contains two main components: representativecolors and the percentage of each color. DCD can provide aneffective, compact, and intuitive salient color representation,and describe the color distribution in an image or a region of interesting. But, for the DCD in MPEG-7, the representativecolors depend on the color distribution, and the greater part of representative colors will be located in the higher colordistribution range with smaller color distance. It is may be notconsistent with human perception because human eyes cannotexactly distinguish the colors with close distance. Moreover,DCD similarity matching does not fit human perception verywell, and it will cause incorrect ranks for images with similarcolor distribution. We will adopt a new and efficient dominantcolor extraction scheme to address the above problems[7,8].According to numerous experiments, the selection of color space is not a critical issue for DCD extraction.Therefore, for simplicity and without loss of generality, theRGB color space is used.Firstly the image isuniformlydivided into 8 coarse partitions, as shown inFig. 2. If there areseveral colors located on the same partitioned block, they areassumed to be similar. After the above coarse partition, thecentroid of each partitionis selected as its quantized color. LetX=(XR, XG,XB) represent color components of a pixel withcolor components Red, Green, and Blue, and Ci be thequantized color for partition i.
Fig. 1 The coarse division of RGB color space.
B.
Extraction of dominant color of an image
The procedure to extract dominant color of an image is asfollows:According to numerous experiments, the selection of colorspace is not a critical issue for DCD extraction. Therefore, forsimplicity and without loss of generality, the RGB color spaceis used. Firstly, the RGB color space is uniformly divided into8 coarse partitions, as shown inFig. 2. If there are several
119 http://sites.google.com/site/ijcsis/ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 9, No. 2, February 2011
colors located on the same partitioned block, they are assumedto be similar. After the above coarse partition, the centroidof each partition (“color Bin” in MPEG-7) is selected as itsquantized color.Let X=(XR, XG,XB) represent color components of a pixelwith color components Red, Green, and Blue, and Ci be thequantized color for partition i. Theaverage value of colordistribution for each partition center can be calculated byAfter the average values are obtained, each quantized colorcan be determined by using In this way, the dominant colors of an image will be obtained.
C.
Extraction of texture of an image
Most natural surfaces exhibit texture, which is animportant low level visual feature. Texture recognition willtherefore be a natural part of many computer vision systems.In this paper, we propose a texture representation for imageretrieval based on GLCM. GLCM [11, 13] is created in four directions with thedistance between pixels as one. Texture features are extractedfrom the statistics of this matrix. Four GLCM texture featuresare commonly used which are given below:GLCM is composed of the probability value, it is definedby which expresses the probability of the couplepixels at direction and d interval. When and d isdetermined, is showed by P
i, j
. Distinctly GLCMis a symmetry matrix and its level is determined by the imagegray-level. Elements in the matrix are computed by theequation shown below:
)4(),,(
),,(
),,(
i j
d jiP
d jiP
d jiP
GLCM expresses the texture feature according thecorrelation of the couple pixels gray-level value at differentpositions. It quantificationally describes the texture feature. Inthis paper, four texture features are considered. They includeenergy, contrast, entropy, inverse difference.Energy
x y
yxPE
)5(,
2
It is a texture measure of gray-scale image representshomogeneity changing, reflecting the distribution of imagegray-scale uniformity of weight and texture.Contrast I =
)6(,
2
yxPyx
Contrast is the main diagonal near the moment of inertia,which measures how the values of the matrix are distributedand number of images of local changes reflecting the imageclarity and texture of shadow depth. Large Contrast representsdeeper texture.Entropy
x y
yxPyxPS
)7(),(log),(
Entropy measures randomness in the image texture. Entropy isminimum when the co-occurrence matrix for all values isequal. On the other hand, if the value of co-occurrence matrixis very uneven, its value is greater. Therefore, the maximumentropy implied by the image gray distribution is random.Inverse difference
)8(),(
)(11
2
yxPyxH
x y
It measures number of local changes in image texture. Itsvalue in large is illustrated that image texture between thedifferent regions of the lack of change and partial very evenly.Here p(x, y) is the gray-level value at the Coordinate (x, y).The texture features are computed for an image when d=1and =0
0
, 45
0
, 90
0
, 135
0
. In each direction four texture featuresare calculated. They are used as texture feature descriptor.Combined feature vector of Color and texture is formulated
.
III.
EXPERIMENTAL SETUP
A.
Data set
Wang’s [15] dataset comprising of 1000 Corel imageswith ground truth. The image set comprises 100 images in eachof 10 categories. The images are of the size 256 x 384 or384X256. But the images with 384X256 are resized to256X384.
B.
Feature set
The feature set comprises color and texture descriptorscomputed for an image as we discussed in section 2.
C.
Computation of similarity
The similarity between query and target image ismeasured from two types of characteristic features whichincludes dominant color and texture features. Two types of characteristics of images represent different aspects of property. So during the Euclidean similarity measure, whennecessary the appropriate weights to combine them are alsoconsidered. Therefore, in carrying out Euclidean similaritymeasure we should consider necessary appropriate weights tocombine them. We construct the Euclidean calculation modelas follows:D(A, B
) =ω
1
D(F
CA
, F
CB
) + ω
2
D(F
TA
, F
TB
) (13)
),,(
d jiP
),,(
d jiP
120 http://sites.google.com/site/ijcsis/ISSN 1947-5500
Search History:
Result 00 of 00
00 results for result for
p.
Content Based Image Retrieval using Dominant Color and Texture features
Nowadays people are interested in using digital images. So the size of the image database is increasing enormously. Lot of interest is paid to find images in the database. There is a great need for developing…
(More)
1.5K
Reads
4
Readcasts
17
Embed Views
More From This User
Notes
Load more
