Motion Compensation Based Video Artifacts Removal

After the video is compressed and decoded,its content quality often degrades and artifacts affecting people's perception appear in the picture,which is often brought about by the rough coefficient quantization process and motion compensation in the video coding process.In recent years,scholars have proposed single-frame and multi-frame algorithms for compressed video artifact removal based on deep convolutional neural networks,which have achieved good results,but the single-frame algorithm has the defect of not being able to handle temporal artifacts,and the multi-frame algorithm generally faces the problem of imperfect motion compensation schemes.In addition,most of the current algorithms do not take into account the characteristics of different video coding standards and do not deeply explore the useful information in the video compression coding and decoding process.In this paper,we deeply explore and study the H.264 video coding standard and propose a method for compressed video artifact removal using the macroblock segmentation information,motion vector and reference index calculated during the coding and decoding of H.264 video coding standard.We extract the macroblock segmentation information,motion vector and reference index of each frame in the compressed video from the code stream.We design a novel multi-frame compressed video artifact removal model that integrates the replication information we extract from the codestream.Our model consists of a spatio-temporal feature fusion network and a video quality enhancement network.The spatial-temporal feature fusion network accomplishes the goal of fusing the target frame with the reference frames information using the motion vector and the reference index information.The video quality enhancement network further improves the artifact removal effect on the fused features.Experimental results show that the method proposed in this paper can effectively reduce video compression artifacts.