The Research On Key Technologies Of Semantic-Based Image Retrieval

With the development of digital image technology, particularly the rapid develop-ment of internet technology and its broadly usage, people are now facing more and more images in daily life, and the contents of images are more diverse as well. Therefore, the effective management on image resources is now a crucial application requirement in most domains. The research on image retrieval has become a hot spot in research community as there is such a large demand.The form of a digital image stored in computer is a pixel matrix, but the informa-tion that human obtains through observing the image is not merely the pixel matrix, but also the subjective, semantic meanings that closely relate to their experience and knowledge. Consequently, the visual features extracted by computer fail to describe the high-level semantic meaning accurately, which raised the problem of "semantic gap" in image retrieval. A lot of methods have been proposed in order to solve the "semantic gap" problem but none of them finally achieved this goal successfully.In this thesis, we addressed the problem in three aspects, from which further solu-tions have been proposed.Firstly, before the retrieval process a novel image segmentation method was de-signed to process the image, then high-level semantic features were obtained from the segmented result. The features were targeted quantized for the following retrieval phase. As the experiments show, the algorithm is able to obtain the semantic features of the image.Secondly, semi-supervised learning algorithm was introduced into the process of relevance feedback of the image retrieval to learn the semantic of the retrieved image. We proposed an enhanced relational graph algorithm by improving the existing seman-tic manifold learning algorithms based on the framework of manifold learning. Exper-imental results of the image retrieval and feedback image retrieval, show the efficiency of the new algorithm.Thirdly, in the thesis, we also studied on the technology of image semantic classi-fication which can be directly used in image retrieval. A new coding algorithm based on the differentiation maximization was proposed as an improved codebook model. Experiments have verified the effectiveness of the proposed algorithm.