Content-based Image Retrieval has been drawing more and more research attention in the recent years. This is in response to the requirement in multimedia applications, such as digital libraries, where efficient image indexing and access tools are essential. In this thesis, we discuss most of the basic research issues and concepts in this field. We focus our discussions mainly on the definitions, representation, extraction and matching algorithms of low-level image features, including color, texture and shape. An efficient color-induced retrieval scheme is presented, followed by convincing test results. The technical implementation of our image retrieval system, as well as some key components in building such a practical system, is also covered in great detail. Feature representation and extraction are the basis of Content-based Image Retrieval. Unlike text-based features, visual features are the natural way to represent the content of the image, and can be generated automatically. The most frequently used features are color, texture and shape. This thesis provides a comprehensive survey of the representation and extraction of these low-level features, and also pays close attention to the matching criteria between a pair of features. Color information is the most intensively used visual feature in virtue of its strong correlation with the underlying image objects or scenes. Compared to other low-level visual information, color is more robust with respect to scaling, orientation, perspective and occlusion of images. Here we propose an efficient scheme of color-induced image retrieval, which consists of a new set of features and its specific matching strategy. We make use of the spatial knowledge of color distribution pattern and assume a rectifying process to refine the result from traditional color clustering. Experiment results show that our approach is more robust than conventional color clustering methods in terms of accuracy and efficiency. Since the 1990's many Image Retrieval systems have been built, both commercial and academic. We select a few representative systems and highlight their distinct characteristics. In this thesis, we also give a detailed picture of our own Image Retrieval system. The whole system architecture is illustrated thoroughly and the exact operating principle is explained accordingly.
修改评论