Video Coding Research Based on Texture Characteristics
|School||Harbin Institute of Technology|
|Course||Computer Science and Technology|
|Keywords||Texture analyses Texture synthesis AR model JND Adaptive coding on MB level|
Since the 1990s, with the rapid development of Internet and mobile communications, the processing and transmission of video and multimedia on Internet have become a hot spot. At present, the video compression technologies mainly focus on three aspects of elimination of redundant data: spatial redundancy, temporal redundancy and statistics redundancy. The latest video compression standards MPEG-4 H.264/AVC and China’s own AVS standard have been greatly exceeded the standards of previous generations. These coding standards commonly use the framework which are integrated with motion compensation, intra prediction and transform to eliminate the redundancy in spatial and temporal space, and adopt rate distortion optical as a performance evaluation criterion. But video scenes always have a lot of texture regions, such as: grass, water and so on, these regions can be seen as part of the background. In most cases, observers only concern about the semantics of these regions while careless about their specific details. The traditional coding framework may use expensive bits in these areas, how can we ensure the image quality, but also expend less costly bits as the increasingly valuable of the network bandwidth is quite necessary.According to the texture analysis and synthesis coding framework, video sequence are first divided into texture frames and non-texture frames, then the former are coded with analysis and synthesis method, while the latter are still coded in the traditional way. In analysis part, we proposed a JND (just noticeable distortion) based texture segmentation algorithm to detect the texture region. Compared with previous methods, it has fully considered the HVS (human visual system) characteristics and the segmentation results are more accurate. In synthesis part, a texture synthesizer based on AR (auto regressive) model is utilized. Each texture frame has a group of AR parameters which are used for synthesis the texture region in decoder.In addition, the paper also proposed an adaptive coding framework based on self-feedback AR model, i.e. each macroblock is coding in two ways, traditional framework and AR model, and then the best mode is selected. In self-feedback AR model, the macroblock is synthesized without transmitting the AR parameter. The experiment results have shown that, when the ratio of AR model is much greater than the traditional ones, the compression efficiency of texture frames is improved in a certain extent.