Learning to Fuse Residual and Conditional Information for Video Compression and Reconstruction.

With the rapid development of the Internet, video compression and reconstruction have attracted more and more attention as the use and transmission frequency of video data have increased dramatically. Traditional methods rely on hand-crafted modules for inter-frame and intra-frame coding, but they often fail to fully exploit the redundant information of video frames. To address this problem, this paper proposes a deep learning video compression method which combines conditional context information and residual information to fully compress intra-frame and inter-frame redundancy. Specifically, the proposed algorithm uses conditional coding to provide rich context information for residual methods. At the same time, residual coding supports conditional coding in dealing with redundant information. By fusing the video frames generated by the two methods, information complementarity is achieved. Experimental results from two benchmark datasets show that our method can effectively remove redundancy between video frames and reconstruct video frames with low distortion to achieve better than state-of-the-art (SOTA) performance.