Research On The Content-Based Image Retrieval

With the popularization and development of the Internet, large numbers of image data are produced fastly. How to retrieve the needed image data effectively is a hotspot for the present application field of image. To retrieve and identify image conveniently, the technology of the content-based image retrieval (CBIR) emerges as the times require. The paper researched on the technology of the CBIR, the main content is following:On the aspect of feature extraction, three features extraction methods are analyzed separately, like color, texture and shape features extraction; on the aspect of image similarity metric, the present similarity metric are researched on, such as Minkowsky distance metric and Mahalanobis distance metric and so on. The method about normalization of interior and exterior feature vectors is discussed at last.After analyzing the drawback of weighted multi-features image retrieval, a new method is putted forward, namely utilizing the feedback information from user to automatically adjust the weight of the multi-features image retrieval. The weight of Color Moments, co-occurrence matrix and Hu invariants moments and so on are adjusted automatically.Finally, an experiment system of the CBIR is designed. Some retrieval methods are implemented by the system: the retrieval method of color-based, texture-based of co-occurrence matrix and shape-based of Hu invariants moments; the retrieval methods of the random combination of "color+texture+shape"; the retrieval method of automatic adjusting the weight of multi-features according to the feedback information from the user's subjective evaluation. At last using the image database of UIUC and Caltech101, the validity of retrieval methods introduced above are analyzed and compared.The validity of the new method are proved by the result of the experiment, it not only improves the validity of image retrieval but also lightens the burden of users who need set the parameters of weight.