英文摘要 | Image retrieval all feature of image, such as color, texture, sketch, (shape) contour, spatial relation, objective feature, subjective feature, knowledge of domain etc. In this paper, we mainly study how to extract the lower feature of image, i.e. shape, texture, and color. We mainly studied several methods of shape matching, including the stochastic method and analytic method, developed methods based on coconcurrence matrix, improved the color space cluster algorithm and algorithm for region based image retrieval. Hopfield neural network is typical optimization method for shape match, we improved the structure of the network, update algorithm, and the energy function to adapt to the specification of problem of occluded shape matching. Genetic algorithm is an advanced optimization method, in some case, it perform better than neural network. We apply the genetic algorithm to solve problem of the partial occluded shape matching. The result is good. Each shape can be characterized by two one- dimensional signals. So if there exist a lot of shapes in same class, you can use them as train material and build a HMM to characterize shape. Schwarz representation had been studied in our lab for several years. In this paper we show that Schwarz integrals is same as Cauchy integral after studying the physical meaning of it. Based on this property, we build the Schwarz scale space and proposed some property of it. With the elegant result analytic function, the automatic scale selection algorithm is developed. Multi-scale Fourier descriptor is proposed to accelerate the speed of shape matching. We also utilize the Schwarz scale space to extract the dominant point of shape to achieve the simplification of shape. The key problem of image retrieval is effectiveness and efficiency. We designed a three layer hierarchical framework for shape based image retrieval. In the first layer we build the index for shape by their geometric feature. In the second layer Fourier descriptor is used to filter sub-database. In the third layer, we use multi-scale Fourierdescriptor to improve the result of second layer. In order to verify those algorithm, we built a demo system of image retrieval, on the platform the color, texture, region based image retrieval technique is implemented. The result is very good. |
修改评论