Shape-based Image Retrieval

The paper focuses on the research of key problems to the shape-based image retrieval, such as automatic shape extraction, shape similarity measure, spatial relation representation, and fast similarity retrieval. By virtue of the concrete studies of shape-based retrieval, we also probe into much more general problems, such as visual information representation, shape perception model, perception organization and high dimension large-scale data processing.The contributions of the paper are as follows:(1) Image feature spectrum (IFS) method is proposed. By IFS, image contents in color, texture, shape and layout are represented as integrity. IFS is a partial implementation of MPEF-7, and provides a uniform frame for image content representation, extraction and similarity measure.(2) A force field based shape extracting method is proposed. First, curves in an image are extracted using edge detection and edge linking. Among these curves, there exist interactive forces. Within a curve, there exists a degenerating force. Second, by local interaction and global optimization, the secondary information disappears gradually, the interrelated curves that carry the primary information are combined, and at last the prominent shape emerges.(3) A new deformable template matching method is proposed. In order to overcome the shortcomings of template structure, energy formula and deformation method about ordinary template matching, we present a more reasonable template structure, simplify the inner-energy formula of the template, and propose the conception of quasi-similarity for the planar shape so that the template deformation has some psychological foundation.(4) A partial features based similarity K-NN retrieval method is proposed. In order to decrease complexity of computing similarity, we introduce the similarity calculation algorithm that is based on partial features or based on project.(5) A shape-based image retrieval system is implemented. When raw visual data is added to a visual database, prominent regions are extracted automatically by using image feature spectra.