Compressed Domain Low Level Visual Descriptors

Content-based image retrieval and analysis have been developed for a long time, and various visual descriptors have been proposed. The need of multiple versions of an image spurs the development of image compression and descriptors based on compression domain. However, these descriptors are not able to achieve good performance in terms of quality and resolution scalability. As the appearance of JPEG 2000 compression standard, its coding algorithm and structure of bit stream make the scalability possible. The JPEG 2000 based descriptors can be developed to satisfy multiple compression levels, and keep a good performance even when the images are highly compressed. In this thesis, most existing famous and popular low level visual descriptors are reviewed. Image compression and some image analysis and retrieval approaches are introduced. Two JPEG 2000 based descriptors called state and context are proposed in this research, and an image retrieval system using these descriptors is constructed. Experiments are conducted and the results indicate the proposed descriptors have a good retrieval performance. State and context are further compared with industrial standard MPEG-7 descriptors and state-of-art SIFT method in multiple resolution and quality situations, and the proposed descriptors are proved to be more suitable in compression domain.