The Research Of Key Techniques In Content-based Image Retrieval

With the explosive growth of images, image retrieval has drawn more and more attention. Like textual information retrieval, existing methods for image retrieval are built upon the "query by keywords", which are referred to text-based image retrieval. In text-based image retrieval, images should be annotated by keywords. However, automatic image annotation is still chal-lenging. Moreover, as it is difficult to exactly describe the query image by several keywords, users prefer to give a sample image as the query to find similar images. To achieve this goal, content-based image retrieval is proposed, in which image retrieval is based on the vision fea-tures extracted from image content. There is a growing interest on content-based image retrieval from both academia and industry. In this dissertation, we focus on three key techniques in content-based image retrieval:shape-based retrieval, color-based retrieval, and user feedback processing. The main contributions of this dissertation are summarized as follows:(1) In the shape-based image retrieval, we propose an effective solution for trademark im-age retrieval by combining shape description and feature matching. First, we propose a new kind of contour-based shape descriptor, which can describe the relationship between two adja-cent boundary points more effectively than other existing descriptors. Second, we propose a new kind of region-based shape descriptor based on interest points and their spatial distribution to describe the region-based shape feature effectively. Finally, we propose a new kind of feature matching strategy to overcome the drawbacks of existing strategies, in which the parameter for correcting the dissimilarities between two images can be computed by a conditional probability based method. We conduct a large number of experiments to evaluate the performances of the proposed solution, the proposed shape description method and the proposed feature matching strategy, respectively. The experimental results show that the proposed solutions outperform existing solutions for trademark image retrieval.(2) In the color-based image retrieval, we propose two solutions to address the problem of isolated interest points in the images when building local color histogram. The first solution is to choose the new center of interest points as the basis of interest points grouping for building local color histogram. According to this idea, we present two methods based on the interest region and the minimal circle of interest points, respectively. The second solution is to group the inter-est points by clustering for building local color histogram. According to this idea, we present a method based on the weighted clustering of interest points, in which the weighted clustering al-gorithm is proposed to group the interest points more effectively. We conduct a large number of experiments on the public image database to evaluate the performances of these proposed meth-ods. The experimental results show that the proposed methods can address the existing problem in image retrieval based on the local color histogram to get more precise retrieval results.(3) In the user feedback processing, we propose a novel kernel function based on adja-cency matrix and local combined features to improve the support vector machine (SVM) based user feedback processing method. First, adjacency matrix and local combined features are in-corporated into the existing local-based representation method to describe the image content effectively. Second, the novel kernel function is proposed, which consists of two parts:one is the linear combination of traditional kernel functions, and the other is the computation of adja-cent best matched pairs. In this kernel function, the adjacent best matched pair is proposed to compute the similarities between images more effectively. Finally, we verify that the proposed kernel function can be used in SVM to process the user feedback. We conduct a large number of experiments on two public image databases to evaluate the performance of processing the user feedback by the proposed kernel function. The experimental results show that the proposed kernel function can be incorporated into SVM to process user feedback more effectively. After user feedback processing, we can get more precise retrieval results.