Content-Based Image Retrieval System Research

Content-based image retrieval technology, one of the hottest topics in the field of multimedia information retrieval recently, has become an emerging technology in the field of large-scale multimedia database management. And the good prospect for development and rising demand for application will continue to promote the advance of this technology. So-called content-based Image Retrieval technology (referred to CBIR, Content Based Image Retrieval or QBIC, Query by Image Content) takes advantage of content information of image, such as color, shape, texture and semantic features, for retrieving images which are similar with query instance user provided.In this paper, the background, significance and widespread application of content-based image retrieval technology are discussed. Meanwhile, the worldwide status of researching of CBIR is summarized according to researching emphases. And some existing and typical CBIR system prototypes are introduced. Then this paper comprehensively analyzes the key technologies of content-based image retrieval, mainly including:image content description, similarity measuring of features, feature index, query design, relevance feedback technology, and evaluation standards of performance of CBIR and so on.Based on the analysis above, the main contribution of this paper is that we propose and implement a CBIR system solution, and scientifically assess the retrieval performance of this system. Mainly includes:first of all, a density sampling SIFT descriptor is adopted for low-level feature extraction of image. The bridges are build through "Bag of visual words" model between low-level visual feature and high-level semantic. And then the two-dimensional spatial location information will be embedding into the feature vector using a pyramid matching algorithms. This kind of features is more stable and coincident with human visual perception. Secondly, flexible operation modes are adopted for user's Query instance. Furthermore, a histogram intersection method is used to measure the feature similarity. Thirdly, feedback mechanism, labeling positive and negative feedback image groups, can make the system learn the true intention of user, and then to optimize the result of retrieval. Finally, this system is test on two image databases, and the results of retrieval are evaluated and analysis scientifically. After that some defects which need further improve are found and solutions are proposed.