Visual Information Based Texture Coding Technique

Recent years, image and video coding has made great improvement, in terms of the compression efficiency, international standard such as JPEG2000,H.264/AVC,MPEG-4 and our national AVS are twice better than the pervious generation. Whereas, all main video coding standards take Rate Distortion as the criterion in judgment, consider little about the most authoritative judge of video--human. From the import of judgment by view, future video coding system will consider more about the quality based on human's eyes. So, this paper does some research on texture region in image based on human's visual, proposed a texture coding framework based on visual information. It works from the texture analyzer, which picks up the texture region where human's eye is not sensitive, and using visual quality judge standard to deal with by the best texture synthesize method, other region still use the traditional ways for coding.In our proposed coding framework, the first step if divide texture region where human's eye is not sensitive from video image, this paper use two segmentation methods, one is texture segmentation based on block clustering, we take SCC descriptor mentioned in MPEG-7 as the comparability rule, and combined with EMD distance to partition the texture region. Another is texture segmentation based on K-means clustering, using texture and color as the character vector, for the better result of texture synthesis.Based on research and analysis of the affine model found on global motion estimation as well as texture synthesis method based on graph cut, we proposed a amelioration texture synthesis method based on graph cut, which improved the quality of synthesis in some extent. Then we proposed a texture coding framework based on visual information, first judging the processing macroblock whether to be a texture one on the encode side, If it is a texture macroblock, the block is processed by both two methods, then use JND model to choose the better method, and using side information and reference frame to synthesize the texture blocks on the decode side. Or, the traditional encoding and decoding method is applied to the current block. Experiment results show that our texture coding framework can achieve very good visual impression on different kinds of texture image.Besides, another texture synthesis method based on Inpainting is proposed in this thesis. We use patches to fill in the texture regions, got rid of the pixels in the overlapped regions between patches, so there were some gaps between patches. We imposed Inpainting technology to synthesize these gaps. Experiments indicate that our proposed method can reach good visual impression.