Deep Multiframe Enhancement for Motion Prediction in Video Compression

This work proposes a novel Deep Learning technique to increase the efficiency of currently available video compression techniques based on motion compensation. The goal is to improve the frame prediction task, whereby a more accurate prediction of the motion from the reference frames to the target frame allows to reduce the rate needed to encode the residual. This is achieved by means of a convolutional neural network (CNN) architecture that processes the basic block-based motion-compensated prediction of the current frame as well as predictions from past reference frames. This method allows to reduce typical artifacts such as blockiness, and achieves a more accurate prediction of motion thanks to the representation capabilities of CNNs, leading to smaller prediction residuals. Preliminary results show that the proposed approach is capable of providing BD-rate gains up to 6%.