Research Of Image Retrieval System Combining With Visual And Semantic Features

With the rapid development of computer technique and the wide use of Internet, more and more image files has been produced. So, how to find needed image has become a difficult problem and worth studying to resolve. Content-based image retrieval(CBIR) technique has been proposed to solve this problem. The research content of CBIR is how to extract low-level feature or semantic feature which can be fully describe image content from image and use it to support fast image retrieval.The research works in this dissertation include several aspects. One is define and extract low-level features which can be fit for image content representation. The second is study and obtain semantic information of images. The third is how to implement an image retrieval system. An efficient and practical image retrieval system is built based on the research work.In the representation of image feature, this dissertation summarize common image features briefly and image feature representation method based on the regional main color connect component is proposed. This method divide the image area into regular regions and extract the main color of each region, according to the adjacent region with the same main color, the whole image is divided into a number of connected components, the main color and size of each regions are treated as the region's features. This representation contains not only the image color information, and contains some information about the spatial distribution and shape. This is more accurate in description with various aspects of image content. The similarity measure, based on the image feature representation, treat the feature of all regions as a string, and measure the distance between two images by the string comparison.In the image semantic learning, the dissertation uses the semi-supervised learning approach to learn and propagate the image semantics, a large number of unlabeled images to get an initial semantics.Relevance feedback in retrieval process lets the user factors joined in the system, this can further capture user intent, and improve the system performance. In the retrieval process with visual features, we use a manifold learning approach to use user feedback information. By seeking sample and feedback images and k-nearest neighbor, and treating all the image as vertex in graph, obtaining the edge between the image points by local conformal mapping, using the shortest path between two points to measure the distance between the images, this representation is more effective than the Euclidean distance.In semantic improvement process, this dissertation integrates use the image of the hidden semantics and obvious semantic features. Hidden semantics generally appear in the conventional image retrieval based on visual characteristics, only know image with the same semantics, but do not know the specific semantics. Obvious semantic features generally appear in the Web-based image retrieval, this can obtained with the user inputs or the text around the image. This dissertation combines the advantages of both, using obvious semantics to improving the image semantics directly, using the correlation between the image with same hidden semantics and obvious semantics in the part images that we can take the obvious semantics spread to image which has no obvious semantics. By this way we can refine the image semantic annotations.In the final this dissertation describes in detail an image retrieval system implementation combined with visual and semantic features, the system's objectives, system structure and the actual performance evaluation are all introduced. Experimental results show that combination of visual and semantic image retrieval system can describe the user's query from the various intentions, has better query performance, to meet the needs of different users.