Research On The Image Retrieval System Combining Multiple Content Features And Relevance Feedback Technology

With the rapid development of computer and network,as well as the popularization of image acquisition device, the type and quantity of image data are speedily increased.Current researches are focusing on the methods of finding required images from huge image information fast and accurately. In the early time, the image retrieval was a kind of key words researching technology based on artificial words noting of images. But artificial note is insufficient and subjective. What is more, it needs too much time and energy. And then, CBIR(Content-Based Image Retrieval) appears in recent years.CBIR combines several subjects: pattern recognition technology, image process,database management, computer vision and so on. The CBIR descriptors extract feature vectors in accordance with content features such as color, texture, shape, etc. The degree of similarity between two images is measured by the distance of their feature vector,when retrieving in an image library, images will be returned in the order of their similarities with the querying image. So the key of CBIR retrieval effect is selecting appropriate features to describe the content of an image, moreover, choosing good methods to extract and match them.With the deepening of research work, CBIR descriptors based on different content characteristics emerge in endlessly. After analysing on some popular CBIR descriptors,a new image retrieval descriptor FCTL(Fuzzy color、texture and location) is presented in this paper. In FCTL, the feature of an image is described by a 300 bin histogram containing fuzzy color, texture and location information. Matsushita distance between their histograms is used to represent the similarity of two images. ANMRR which is recommended by MPEG-7 Standard is used to evaluate experimental performance,comparing with the retrieval results of different descriptors on WANG and UCID image libraries, we can find that FCTL has higher retrieval accuracy than some frequently used descriptors.In the current, CBIR descriptors usually describe the image through its low-level content features, but not semantic information, it may different with the user’s understanding of image, and leading to unsatisfactory search results. Therefore, in order to achieve high customer satisfaction, it is not only important to choose an accurate retrieval descriptor, but also should add a relevance feedback mechanism to achievehuman-computer interaction. Keep learning user preferences according to his evaluation of searching results, and return a better result.This paper presents a relevance feedback mechanism based on query point movement, it adjusts the matching weights of each dimension of the feature histogram on the basis of feedback information, to make the relevant images more and more similar with querying image, the non-relevant images more and more dissimilar.In order to facilitate experimental analysis, this paper constructs an image retrieval system named Image Retrieval, it supports FCTL and serval other image retrieval algorithms, such as: FCH(Fussy Color Histogram)、EHD(Edge Histogram Descriptor)、FCTH(Fuzzy Color and Texture Histogram) and CEDD(Color and Edge Directivity Descriptor), which are based on different content features. Image Retrieval offers two commonly used image library: WANG and UCID. In order to further improve the retrieval efficiency, the system adds relevance feedback mechanism mentioned above.