A channel-wise contextual module for learned intra video compression.

In the multimedia era, exploding image and video data highlight the importance of video compression for storage and transmission. The All-Intra structure is a coding mode in HEVC and VVC, in which each frame is encoded using intra coding, and in this paper learned All-Intra coding is explored on the basis of the research of the learned image compression. A channel-wise contextual module based on channel segmentation is introduced to fully exploit non-local information. Then, two distinct attention mechanisms are designed for different feature layers to enhance the effectiveness of the transform network. Additionally, a post-processing module is employed to enhance the quality of decoded frames. Experimental results on the Kodak and Tecnick datasets demonstrate that the proposed method performs better than the majority of the recent learning-based methods and traditional image codecs (BPG, JPEG2000 and JPEG), and also perform better than traditional video codecs in terms of PSNR. • A novel video compression framework specially designed for I frame is introduced. • A channel-wise contextual module enhances context-based entropy modeling. • A post-processing module aims to enhance the quality of the reconstructed frames. • Two distinct attention mechanisms are designed for different feature layers. [ABSTRACT FROM AUTHOR]