EFFECT of REGION FILTERING on the PERFORMANCE of SEGMENTATION BASED CBIR SYSTEM Jagadeesh D.Pujari #, Arati S.Nayak * Department of Information Science and Engineering, S.D.M. College of Engineering and Technology, Dharwad, India # jaggudp@yahoo.com , * arati_nyk@yahoo.com Abstract- The Color and texture information have been the primitive image descriptors in content based image retrieval systems. This work describes an image retrieval method which uses color and texture approach for feature extraction. An image is represented by a set of regions, roughly corresponding to objects, which are characterized by color and texture. For segmenting images, JSEG (J-Segmentation) algorithm is used. The default settings of the algorithm are used. Feature extraction of every region of an image is done using wavelet decomposed coefficients of an image and its complement. Image retrieval is done using Region based image retrieval (RBIR) technique. A measure for the overall similarity between images is performed using a region matching method that integrates properties of all the regions in the images. The JSEG algorithm suffers from over segmentation. Because of the undesirable results from JSEG algorithm, region filtering is applied to the segmented images. Experiments are carried out on Wang’s dataset to study the performance of CBIR system after the application of region filtering. Key words: CBIR, segmentation, wavelet, decomposition, region filter, conditional co-occurrence histogram. 1.0 Introduction Content based image retrieval (CBIR) [1][2][3][4], also known as query by image content is the application of computer vision to the image retrieval problem. Content- based means, that the search makes use of the contents of the images themselves, rather than relying on human input meta-data such as captions or keywords. In text based approach, it is very difficult to answer user queries reliably, because different people may give different textual explanation for the same picture. Entering textual annotations manually is very expensive for large scale image databases. To overcome this difficulty, content based image retrieval systems have been developed, where image retrieval is based on the visual content. RBIR is a type of CBIR, in which the images are segmented into regions using a segmentation algorithm. A region is seen as a part of an image with homogeneous low level features. In RBIR, regions are used to represent and index images. In Whole-Image-as-Query RBIR, the user provides the example image and the system extracts feature information for the whole image for performing query [4]. In an Image-Region-as-Query RBIR, the user performs a query by selecting regions of significance from the example image, according to their necessities [1][2]. Segmentation based techniques for image retrieval have been used for obtaining better, texture and color descriptions of the image contents. It is difficult to find segmentation algorithms that match to the human perception. The aim of CBIR system is to produce a ranked list of images that are relevant to the query. The system starts with the pixel representation of the image. To rank images, we must first extract the visual features that are the core of CBIR systems. Features are the condensed representation of a visual aspect of the image, either globally for the entire image or locally for a small group of pixels or segments. Color and texture features can be extracted from every region of an image and a measure for the overall similarity between images, developed as a region matching method that integrates properties of all the regions in the image. Since segmentation process cannot be perfect, each region of a query is compared with every region of images in the database. In this paper, to analyze the effect of segmentation on retrieval performance of a CBIR system, experiments are carried out with different number of regions, eliminating all of regions below a cutoff size. In section 2, the wavelet based co-occurrence histogram method, with correlation information between the subbands obtained by the wavelet decomposition of the image and its complement, used to build the feature set is presented. Section 3, deals with algorithm for, Integrated Region Matching (IRM). In section 4, experimental results are discussed. Conclusion is given in section 5. 2.0 Material and methods 2.1 Image data The image set used for experiment consists of 1000 pictures which belong to WANG’s database [5]. It is a subset of Corel images of natural scenes divided into 10 labeled categories of 100 images each. The images are stored in JPEG format. The images are of resolution 256X384 or 384X256. A total of 100 queries with ground truth were formulated. All the images in the database are initially segmented into regions using the JSEG [6] segmentation algorithm proposed by Deng and Manjunath [7]. The default settings of the algorithm are used. 2.2 Feature computation Texture feature for each region of an image is calculated using wavelet based decomposition and cumulative 292 978-1-4244-8594-9/10/$26.00 c 2010 IEEE Authorized licensed use limited to: SDM COLLEGE OF ENGINEERING AND TECHNOLOGY. Downloaded on October 15,2024 at 07:14:15 UTC from IEEE Xplore. Restrictions apply.