Research On Content-based Image Retrieval Technology

With the development of network and multimedia technologies, the application of the image increasingly widespread. Content-Based Image Retrieval which becomes one of the hottest research topics in the field of image has emerged at the right moment. However, due to the limitation of CBIR, it is still unable to fully meet the user's requirements. Therefore, it brings very vital significance that the research on underlying content feature retrieval of the image.This thesis mainly researches the retrieval technology of the image color feature and the shape feature. First, a general overview of the development and research actualities of the content-based image retrieval system are introduced, and its key technology is researched, including the extraction of the image features (such as color, texture and shape), the similarity metric, relative feedback and retrieval performance evaluation. Secondly, as for the color feature extraction, in order to improve the retrieval efficiency, HSV color space is chosen and the color is quantized to 72-dimensions. According to the different attention degrees, the overall-and-aerial principal color method is adopted to divide the image, and the image is retrieved again with the different ratio so that the loss of space information is overcome. Thirdly, as for the extraction of the shape feature, after extracting image edge by Canny Operator, image's outer contour is gotten by technology of contour tracking. Then region filling is done for outer contour, and image's shape feature is extracted by using invariant moment after obtaining image region target.Finally, in order to design and implement the content-based image retrieval prototype system by synthesizing the feature of overall-and-aerial principal color and shape invariant moment, the Eigenvector Normalization Theory and Image Retrieval method integrated with multi-feature are researched. The experience results show that the retrieval precision of fusion multi-feature is much higher than traditional single feature retrieval, and achieve the desired goals.