A Global Appearance and Local Coding Distortion Based Fusion Framework for CNN Based Filtering in Video Coding

In-loop filtering is used in video coding to process the reconstructed frame order remove blocking artifacts. With development of convolutional neural networks (CNNs), CNNs have been explored for in-loop considering it can be treated as an image de-noising task. However, addition being a distorted image, also obtained by fixed line block based encoding operations coding. It carries coding-unit distortion some similar characteristics. Therefore, this paper, we address problem from two aspects, (i) global appearance restoration disrupted texture and (ii) local caused pipeline Accordingly, three-stream fusion network developed with high-level feature stream, stream low-level stream. Ablation study conducted validate necessity different features, demonstrating that features complement each other achieve better performance when combined. To best our knowledge, are first one clearly characterizes above aspects experimental verification, providing clear pathway developing filter techniques. Experimental results demonstrate proposed method significantly outperforms existing single-frame methods achieves 13.5%, 11.3%, 11.7% BD-Rate saving on average AI, LDP RA configurations, respectively, compared HEVC reference software.