Deep Video Compression using Compressed P-Frame Resampling

Abstract

The Convolutional Neural Network has emerged as one of the major players in the field of deep video compression. Many deep learning models relying on convolutional layers have outperformed the state-of-the-art compression standards by a huge margin. Although their work is still at infancy level, they seem to be the future of video coding. The proposed approach uses a frame resampling based video compression approach using Temporal 3-D CNN based encoder and Y-style CNN based decoder concatenated with High Fidelity GAN based entropy coding for frame compression. The proposed architecture employs frame downsampling method over the residual frame to control the bitrate of the compressed data and is trained through a simplified stagewise training procedure. The extensive experiments are conducted with different datasets and different colorspaces. The study shows that the proposed model outperforms the H.265 by 0.255 dB in terms of PSNR and nearly 0.1 in terms of MS-SSIM.