Research On Image Retrieval Based On Semantic Feature

With the fast development and wide application of multimedia technology, computer technology, communication technology and Internet technology, people can collect and produce various kinds of multimedia data based on plenty of methods, and it is more and more important to organize and manage the multimedia information. Most frequent and maximal information of the multimedia data is the visual information. Therefore, the issue of the visual information retrieval has attracted many people's attention. Considering the application of the multimedia, we mainly research on the content-based image retrieval, which is a novel research direction of image retrieval. For the non-structured image data, the traditional text-based image retrieval is so inefficient and thereby the content-based image retrieval is bought forward and it has developed a lot. However, there has a big hurdle in content-based image retrieval till the present moment, i.e.,"semantic gap"."Semantic gap"is the gulf between the low-level image visual feature and high-level concepts, the images can be different of semantic concept while having similar visual feature, and they can also be different of visual feature while having the same concept. In this thesis, we research on the bridging"semantic gap"and extracting the semantic feature, and then we propose two effective algorithms to extract the semantic feature.Firstly, one image retrieval framework based on"Fuzzy Semantic Relevant Matrix"(FSRM) is proposed, which uses users'relevance feedback sufficiently. The method of initializing FSRM uses the low-level visual feature, while the initialization of FSRM using users'retrieval logs. Therefore, the performance of the system becomes more and more satisfying along with the use of the system. In the short time relevance feedback session, the images are clustered to catch the semantic requirement using the user's relevance feedback information. A multi-layer retrieval algorithm is also proposed to learn the hidden semantic information. The experiments, based on a standard database including 1000 images in the CORE CD, demonstrate the proposed framework has extraordinary improvement on the retrieval speed, relevance feedback, and long-term memory...