Research On Content Based Image Retrieval

As the development of multimedia information technology and Internet technology, especially after the appearance of mass image database, people have paid more attention to image information. How to find a special image from the mass database efficiently and effectively has become a major issue to be solved. In this case, the research on content-based image retrieval was developed, and became a hot spot in the fields of multimedia information processing. This dissertation focuses on the research of image similarity analysis and the using of image local features in image retrieval. To make it clear, the main content and contribution are summarized as below:1. By studying graph-based data analysis methods, this dissertation proposed a twice ranking retrieval framework integrated manifold ranking and region matching. First, images are ranked according to the manifold structure of all image feature data. And then the proposed region matching graph (RMG) is constructed. By using RMG, the region-based similarity of images can be evaluated, which can modify the result of manifold ranking. In relevance feedback, by using the returned images, relevant and irrelevant, the ranking score in manifold ranking and the region weights of images are adjusted for improving the retrieval results.2. Two effective methods based on multiple instance learning for image retrieval are proposed. The first method extracts salient points of image by tracking wavelet coefficients of different scales, and then retrieval image using the global statistical feature of salient points. The images containing salient points are treated as multiple instance bags, and trained using EM-DD algorithm. A target feature representing the image content is learned in the training and used for retrieval. The second method segments each image into regions which are treated as instances, and constructs a bag-level weighted graph with the diverse density and similarity of instances. Then, a graph-based multiple instance learning framework for retrieval is carried out. Experimental results show that the two methods are helpful for image retrieval.3. Based on the research of probabilistic latent semantic analysis in fields of text information retrieval and classification, a method introducing the thoughts of probabilistic latent semantic analysis to region based image retrieval is presented. This method first defines a latent variable model for region-image co-occurrence probability. To obtained the latent semantic model, the expectation-maximization technique is used with a derived iterative procedure. The posterior probabilities of the region-image co-occurrence data for the latent variable are regarded as quantitative semantic measure to images. And then, all images can be retrieved by latent semantic feature using manifold ranking method.A test system of content based image retrieval is designed and realized. Encouraging experimental results from Corel image database illustrate the validity of the proposed methods in this dissertation.