Improved Asymmetric Bagging Based Relevance Feedback Strategy For CBIR

With the discovery of internet and a widespread mobile terminal, image has become an indispensible part of real life activities. And the number of images is increasing day by day. How to quickly get useful images from these images becomes an urgent problem to solve. Content based image retrieval method is one of the essential technologies employed in this area, this technique involves many disciplines, and it is a comprehensive discipline. In C BIR system using low- level features to describe the image sometimes is different from that of high- level image characteristics in such a situation semantic gap is generated. Relevance feedback is a very popular technique for narrowing the semantic gap and it has a very good effect in CBIR application. Relevance feedback is a technology that is used to achieve customer satisfaction by collecting feedback information through interaction. It is known that small sample result to asymmetric problem in feedback samples and further lead to the wrong classification outcome. All these problems still need to be studied rigorously in an image classification task and be solved.To solve the above problems, this paper researches on the technology involved in relevance feedback, in order to solve the problems of small sample, sample asymmetry and real-time problems. The focus is on relevance feedback algorithms such as QPM, FW, BI, Single-SVM, and PSO and so on; they were conducted based in the analysis of advantages and disadvantages. It involves using vector components to describe image features combined with the classification model to realize the retrieval. The main research work of this paper can be divided into five parts, such as: 1) this paper first introduces the background and the status of the research content, and on the basis of relevant studies conducted a brief analysis. 2) Many kinds of relevance feedback technology of image retrieval have been studied theoretically and review is given. In this work the query point movement, weight adjustment and relevant feedback technology are researched, though, we mainly studied RF based on support vector machine. We described the theory of the asymmetric bagging and random subspace, at the same time we clearly expressed the principle and the main process involved. Then, in order to solve the small problem, asymmetric feedback sample and real-time problems, we proposed an improved asymmetric package named IAB. 3) Due to the fact that the improved algorithm involves the influence degree of different samples, we defined a function to measure sample of importance. For outstanding sample of importance of each round, we propose to introduce the sample weight that depends on each feedback layer. We further put forward the new framework IAB-FSVM-RF, and realize the successful image retrieval. In order to prove the effectiveness and generalization of the presented method in this paper, experiments were tested with the standard database and medical image data set.