Image Retrieval System Based On New Algorithm Of Similarity Measurement And Relevance Feedback

In recent years, content-based image retrieval (CBIR) has been an increasingly active direction. CBIR is very useful in various areas. But there are still many problems in measuring similarity and capturing semantics of images.Segmentation is needed in capturing the spatial information of images. Watershed algorithm based on image domain fails to capture the global color distribution information, while proper models are needed for method based on feature domain. A two-step segmentation framework is adopted in this paper. In the first step, the watershed algorithm is applied on 3-D L*a*b* color histogram. The Highest Confidence First (HCF) algorithm for Markov random fields is taken as the second step to refine the raw result to get continuous regions with relatively low computational load. Experiments on classic images show the effectiveness. This algorithm is also applied on CBIR and medical image processingImage description and similarity measurement are the main problems in CBIR systems. How to alleviate the influence of inaccurate segmentation result is the key. The proposed two-level description (TLD) framework can avoid improper spatial constraint caused by one-level description. Similarity measurement based on unbalanced region matching (URM) is introduced to construct the many-to-one mapping between regions, thus get the image similarity without solving the complex linear programming problem. A novel spatial descriptor is also proposed to integrate various spatial features. Extensive experiments show the effectiveness and robustness of the proposed algorithm.Relevance feedback can bridge the gap between low level image features and high level semantics. In this paper, region-based AdaBoost (RBA) algorithm is proposed. The region-based framework utilizes the segmentation result to capture the higher-level concept of images. Statistics of experimental results and comparison of particular results show that the proposed algorithm can capture the user's intention effectively. Other works on image database visualization are also presented. A practical CBIR system is constructed based on the above research results.