The Study Of The Content-based Image Retrieval Method

Aimed at the increasing application demands of multimedia information retrieval and the defects existed in content-based image retrieval (CBIR)systems, in the dissertation, some important questions of CBIR for teclmiques are discussed and a set of integrated solution is presented.The rationality of the solution is validated by experimental results.At first, the background, development and application of CBIR for universal teclmiques are summarized. Then, main questions are researched, which includes image feature analysls, image retrieval technique, system design and application. Firstly, some typical of image features in CBIR systems, such as color, texture, shape, etc, are detailedly analyzed, and a new method combined color and texture is presented with region feature extraction based clustering, in which image characteristics of color and texture are integrated , the feature descriptions of color, texture and structure are unified, and the invarlability of move, zoom and spin is met for improving the existenting feature descriptlon methods and still more according with human vision. Secondly, image retrieval teclniques in image database are researched, some key teclniques, methods and measures of CBIR are discussed, and a image retrieval model based on RBFNN(Radial Basis Function Neural Network) presented, in which low layer features based on machine vision and high layer features based on semantic description are combined, different retrieval algorithms and new retrieval functions are easily replanted and extended for meeting the demands of image classification and retrieval in large image database. Thirdly, the fused method is presented with PCA(Principal Components Analysis),which can improve the system performance.At last, some main questions of CBIR for universal techniques in System design and application are explored Based on the research fruits above, a universal structure of CBIR system is designed and some optimization measures of system performances are presented,which includes the optimization measure of system interface based on performance feature extraction, fast retrieval with classification. All those methods improve system efficiency. The main results are given and the research work in the future is projected.The main innovations in the dissertation are: (1)A new region-based feature extraction method is proposed,which more accords with human vision than existent similar methods do. (2) A universal retrieval model based on RBFNN classification is presented. (3) A fused method is proposed with different features, which availability is proved by the experimental result. (4) Some performance optimlzation measures in CBIR systems are presented.