Low-bit Rate Video And Image Coding Based On Matching Pursuit

Currently, the demands of video communication on low bit-rate channels (10kb/s~60kb/s), such as wireless channel and telephone network, are becoming more and more comprehensive. Although some existed international video coding standards, such as H.261, H.263, MPEG-1 and MPEG-2, has wined great success in all kinds of commercial applications. But for all these standards, the phenomena of sensible blocking artifacts or image dithering will emerge under narrow bands, which impair the visual quality of reconstructed images. So, these standards still couldn't satiate the demands of very low bit-rate applications (under 20kb/s). As to still image coding, even though the performance of wavelet image coding algorithms excel first generation image compression standard JPEG, but they will produce boring mosquito noise at low bit-rates, which are mainly located around image edges. Besides that, their coding efficiency will drop rapidly for those natural images with rich textures.In order to improve coding efficiency at low bit-rates, this thesis adopted wavelet and nonlinear matching pursuit techniques, combined with some characteristics of human visual system, aimed at improve the performance of video coding at low bit-rates. The major contributions of this thesis are included in the following:1. On a basis of the analysis of statistical properties of wavelet image, the problems existed in EZW are pointed out. At the same time, the deficits of zerotree coding will be discussed. Then, we give our modification of zerotree structure, and propose adaptive image coding based on high order statistical modeling, which aims at removing statistical redundancy existed in subbands. The results of experiments proved our scheme has better image compression quality.2. A new low bit-rate image-coding algorithm based on wavelet and matching pursuit is proposed. Such method divides the whole image into a series of image layers including different signal characteristics, such as low-freq&edge layer and texture layer. For the first, we adopt wavelet image algorithm; For the latter, matching pursuit based on redundant dictionary is utilized. In such way, we can consider the advantages of different transform methods and combine them together to code the image in an optimal way. Last experiments showed the performance of the proposed method outgoes the single wavelet image algorithm.3. A low bit-rate video-coding scheme based on matching pursuit is built in this section. The main idea lies in decomposing the difference frame of prediction in a redundant dictionary with much more basis functions, giving up traditional lattice DCT transform. By choosing non-hybrid dictionary, blocking artifacts of DCT system can be avoided ultimately. The cost of coding motion information is reduced, so the visual quality of images at low bit-rates is improved.4. Building a matching pursuit video coder based on visual fidelity. In this section, the characteristics of human visual system are introduced into bit allocation of coder, in order to improve subject visual quality of reconstructed images. At low bit-rates, because of limitation of the budget bits, we can make the quality of regions absorbing eye focus higher than other regions by adjusting bit-allocation.Finally, the key points of the thesis are concluded, some improvements to be done in the current research are analyzed, and some suggestions and expectations for future work are provided.