Research On Extraction Of High-level Semantic And Low-level Visual Feature For Image Retrieval

Content-based image retrieval (CBIR) is a complex and challenging problem. The common research method is to represent image content using low-level visual feature, such as, color, texture and shape, etc. Here, "content" is some kind of objective statistic character of image, which couldn't be understood by human being directly. The nature of CBIR is to search the relevant or similar images based on low-level visual features, which implies relevant images have similar visual features, so it is possible to cluster or classify images according to low-level visual features. In other words, image classification is limited image understanding, the purpose is to group images into some semantic class, so that semantic feature of images could be extracted automatically, which will not only help organize image database semanticly, but also help label images automatically, this will drive CBIR from lab to industry.Motivated by these, this paper is mainly focused on how to extract image high-level semantic feature from low-level visual feature, on the other hand, CBIR related problems, such as, low-level feature extraction, representation and its similarity measurement, are also deeply discussed in this paper, finally, an experimental CBIR system has been built to validate the methods. We summarized the main work and contributions as follows:(1) Novel texture spectrum descriptorA novel texture description method is proposed. It describe intensity changes of neighboring pixels, which indicates that texture is a kind of change of pixel intensity; the texture unit obtained using this method is local texture, which makes it possible to get texture spectrum; it pay more attention to salient change of image pixels, this is more consistent with human vision perception for image texture, so it can describe the smoothness degree of images efficiently.(2) Image similarity measurement using weighted bipartite matchingMax weighted bipartite matching algorithm for multi-region image similarity measurement is proposed. It incorporates properties of all the regions in the images, which reduces the influence of inaccurate segmentation, furthermore, this algorithm takes the spatial information into consideration, so it can retrieve more relevant and more accurate images.(3) Measuring image similarity nonlinearly and non-metriclyIn this paper, the input space is mapped into feature space using nonlinear map Φ, then replace kernel function with dot production, and compute the similarity among images...