Deep motion-compensation enhancement in video compression

This work introduces the multiframe motion-compensation enhancement network (MMCE-Net), a deep-learning tool aimed at improving the performance of current video coding standards based on motion-compensation, such as H.265/HEVC. The proposed method improves the inter-prediction coding efficiency by enhancing the accuracy of the motion-compensated frame and thereby improving the rate-distortion performance. MMCE-Net is a neural network that jointly exploits the predicted coding unit and two co-located coding units from previous reference frames to improve the estimation of the temporal evolution of the scene. This letter describes the architecture of MMCE-Net, how it is integrated into H.265/HEVC and the corresponding performance.