Adaptive CNN-Enhanced In-Loop Filter for Lossy Video Coding.

Efficient video compression techniques designed to reduce the costs of wireless video transmission play a critical role in meeting the growing demand for video content delivery. In this paper, we propose a data-driven and learning-based approach to deal with the various distortions introduced by conventional video coding frameworks. Specifically, we develop a lightweight convolutional neural network (CNN) that exploits multiple side information as in-loop filters to learn the implicit mapping relationship from reconstructed frames to lossless frames. Furthermore, we design three optional schemes for CNN models to take full advantage of both conventional and CNN filters without increasing the number of models, which facilitate more flexible filtering for various frames. The experimental results show that the proposed CNN-enhanced in-loop filtering module outperforms conventional filters and other CNN filters in H.266.