On the backward refinement in video coding

Nowadays, most of the video coding standards are based on the hybrid coding framework, whose information theoretic model was studied by Viswanathan and Berger more than ten years ago. The theoretic model, now known as the sequential coding model, abstracts consecutive frames as correlated random variables. Even the most efficient High Efficiency Video Coding (HEVC) standard can be roughly described by this sequential coding model. Searching for methods and algorithms which can improve the coding efficiency of HEVC is an attractive yet formidable task. In this paper, we propose a sequential coding model with backward refinement (BR). In this modified sequential coding model, after each frame is reconstructed, it can be backward refined by the frames encoded afterwards. Note that the proposed BR model does not depend on the coding tools and thus can be implemented on any current video coding standard. So, using this BR model, we can further improve the coding efficiency of any video coding standard, with only slight modification of the coding structure. Theoretical analysis is given on the performance of the BR model. We also propose a BR algorithm for the practical purpose, which is implemented on the HEVC standard. Experiment results show a maximum of 2.2% BD-rate saving compared with original HEVC codec.