Research On Several Key Problems Of Content-based Image Retrieval

With the rapid development of multimedia technology and the growing popularity of theInternet, the storage and management of digital images has become a burdensome task. When thetext-based image retrieval cannot meet the huge network image library, there is an urgent needingto propose a solution. content-based image retrieval (CBIR) is proposed, it immediately becamethe focus of the field of the experts. This paper mainly discuss the research on several keyproblems of CBIR, mainly including the following:(1) In order to solve the single feature cannot adequately describe the image content, wepresent a fusion of multiple features of image retrieval method. The algorithm first establish apseudo Zernike chrominance distribution moment, and it extracts the color characteristics of theimage; then decompose using steerable pyramids, and calculate decomposition after eachsub-band standard deviation, mean and entropy as texture features, and finally use thesecharacteristics to calculate the similarity and sort the results, finally the sorted results are returnedto the user. The algorithm can effectively improve the retrieval performance of the system.(2) To solve the problem of the "gap" between image low-level visual features andhigh-level semantic, we use the ensemble classifier and feature reweighting relevance feedbackimage retrieval algorithm. Firstly, using three visual features: color, texture, and shape toconstructing SVM respectively; then integrating the three SVM classifiers; last, weighting thecharacteristic features in round feedback, getting the feedback results. Through experiments wecan find our method has significantly improved the efficiency of the image retrieval method.(3) To solve the unstable performance problems of traditional SVM classifier which iscaused by the number of training samples, we present a HMM based SVM kernel functionmethod instead of the original SVM kernel function; at the same time, in order to solve the SVMclassifier classification error is still large, we present a Semi-BMMA based reordering problemwhich map the features to a new sub-space. Recalculating the feature similarity getting thereordering retrieval results. Simulation experiment shows that it can get the ideal retrieval results.