Research On The Key Technology Of Content-Based Image Retrieval

Due to the development of multimedia and Internet techniques, rapid and effective searching for desired images from large-scale image databases becomes an important and challenging research topic. In this dissertation, lots of exploratory research work has been done around some key techniques of CBIR, which include low-level feature and semantic-level feature extraction, similarity measure, relevance feedback and so on. The research work is valuable both in theory and application. The main contributions of this dissertation are summarized as follows:(1) Fuzzy color histogram based color feature extraction method is analyzedand discussed. Based on image normalization, Zernike moment based shape feature is researched.(2) Based on the current pattern comparison theory, by analyzing the conceptsof joint entropy, conditional entropy and marginal entropy, the mutual information similarity measurement method is proposed. It proved to have the following properties: Non-negativity, symmetry, and triangle inequality. This theory extends the research field of information pattern similarity measurement.(3) Latent semantic indexing(LSI) is introduced in the research field of CBIR. A new approach of extraction of image semantic information is proposed. Latent semantic indexing is built based on the singular value decomposing. It circumvents the natural language processing. Experimental results show that latent semantic indexing can perform effective image semantic indexing.(4) A novel image semantic annotation method is reported. It incorporates thehuman visual perception in the retrieval procedure. It retrieves the images based on understanding the content of the image. This method overcomes the limit of the traditional statistical based semantic annotation method. It improved the accuracy and efficiency of image semantic annotation.(5) Currently, what the content based image retrieval is facing is the dimensionality disaster. It is an indispensable step to make dimensionality reduction through the feature selection. This dissertation is researching on the filter based and wrapper based feature selection method. The feature selection method combined the Relief and support vector machine is presented. It is effective for image retrieval and classification .(6) As any technique is promoted by the performance evaluation of corresponding research area, for the development of effective image retrieval applications it is imperative to study the standard of performance evaluation in content-based image retrieval. Three basic problems areexplained in the paper: establishing of well balanced large scale testbed;defining of objective evaluation criteria;means of getting relevance judgements for queries.