Study On Texture Analysis-synthesis Based Video Compression Techniques

Video Compression is a key issue for future communication service, where video storing and transmitting would play a very important role. Currently, the state-of-the-art H.264 compression standard has almost achieved a peak of compression efficiency. However, it is still within the traditional block based hybrid coding framework. The researchers now put more attention on using perception redundancy, of which one important trend is to implement techniques from computer vision into video coding framework.Comparing standard video coding, the new technology is not based on traditional signal processing theory that focuses on the exact reconstruction of the signals. The intention is to build natural scenes with acceptable subjective quality and understandable semantic meaning. In the natural video sequences, people may have very limited perception accuracy on some certain contents, such as repeated visual patterns, textures, from which we are only able to pick out the semantic meaning such as color, position, etc., but fails to identify each individual pattern. This means we are able to code the video contents beyond Shannon's information law, as long as we allow some uncertainty or distortion. Since human eyes are the final receiver of video content, as long as such distortion is hardly perceived, we can say that higher compression efficiency has been achieved. Based on this idea, we can purposely simplify the original video content in order to use fewer bits during coding, and then reconstruct the missing content with a visually plausible way based on the extracted visual pattern or other assistant information, which can be described as the decomposition—synthesis scheme. There have already been papers proving on feasibility of such a video coding framework. But these various schemes have their own draw backs and there is yet no widely accepted framework available. Currently, the key issue includes extract visual pattern in complex natural video scene, what assistant information is required and how to transmit them. While at the decoder side, the major problem is the detail reconstruction method and how to ensure temporal consistency after reconstruction.This paper will propose a new analysis-synthesis based video framework. In the new framework, the original video sequence will be pre-processed to simplify the complexity of the source signal, thus improving compression efficiency. Meanwhile, we will specifically design the pre-filter so that the different content in the video will be treated differently and important information such as edge and structure will be preserved. We also provide the corresponding detail completion scheme which fully takes advantage of the temporal correlation between frames in the video sequence. Through searching for detail information in the existing data, we will be able to retrieve the missing information in a visually plausible way. One scheme is 3D patch matching that we consider video content as space-time cubic. Under this scheme, the true motion trajectory of the objects in the video can be well preserved and hence avoid temporal flickering in the reconstructed video. Another scheme is to consider video frame as state vector in a dynamic system while we have simplified frames as observations. Kalman filter can be applied to solve the problem to give prediction on missing detail in an optimal manner.In summary, this thesis is about a new method in video coding, which can be considered as an option for low-bit rate video coding. The experiment shows significant improvement for compression efficiency and compatible subjective visual quality.