Codec-Simulation Network for Joint Optimization of Video Coding With Pre- and Post-Processing

In real video coding systems, there might be pre-processing before the encoding and post-processing after the decoding. Those pre- and post-processing should be jointly optimized considering the existing of distortions introduced by lossy coding in between. This paper proposes a differentiable network to simulate the non-differentiable traditional coding to make the joint optimization of pre- and post-processing possible. The proposed codec-simulation network can achieve up to 49.4dB simulation accuracy for HEVC codec in intra mode. When we use down-sampling and up-sampling as examples of pre- and post-processing of a video codec, the joint optimization of these two processing modules in a video coding system can result in 46.88% and 55.15% BD-rate reduction based on PSNR and SSIM compared with that without joint optimization.