Key Technologies Of Image Search With Cross-Media Information And Efficient Image Coding Algorithm

Along with the popularity of digital camera and camera phone, together with therapid growth of Internet, people nowadays is able to access a huge number of digitalimages via World Wide Web. To enable people to easily find the desired image fromsuch a big image database, developing a user-friendly, feature-rich and effective imagesearch system is critical.Image search system consists of three key modules: search interface, search algo-rithm and search results presentation, which respectively correspond to the three keyphases of conducting image search, i.e., inputting the query, performing search andshowing the results. This thesis proposes some key technologies of image search byexploiting the cross-media information and utilizing an effective image compressionalgorithm, which extend or improve the existing image search systems respectively inthese three key modules. On one hand, we exploit the cross-media information to re-spectively extend the scope of traditional image search systems in the point of view ofsearch interface, improve the search performance in the point of view of search algo-rithm, and generate effective summaries from the search results in the point of view ofsearch results presentation. On the other hand, for the search results presentation, wedevelop an effective image compression algorithm to speed up the presentation of thesearch results, so as to improve the user experience.Specifically, for the search interface, we propose a new image search system calledimage search by concept map, which can handle the image search intention concern-ing on not only the semantic concepts in the desired image but also the spatial layoutthe these concepts. Such search intention can hardly be handled by the existing imagesearch systems. for the search algorithm, we propose a regularized graph model toestimate the relevance of a text word and an image by exploiting the cross-media infor-mation. The estimated relevance is used to refine the textual description of the imagesand finally improve the performance of image search. For the search results presen-tation, on one hand, we propose a representative selection approach by cross-mediavoting, which can simultaneously generate visual summary and textual summary forthe image search results, so that the users can get a quick overview of the search re-sults. On the other hand, we develop a directional-DCT based image coding approachto speed up the delivery of the images in the search result. According to our experiments, the proposed key technologies of image search allexhibit promising profermance, and can significantly improve the user experience forimage search.