Luma-only Resampling-based Video Coding with CNN-Based Super Resolution

Chaoyi Lin, Yue Li, Junru Li, Kai Zhang, Li Zhang

Abstract: Resampling-based video coding strategy is well-known for compressing high-resolution video with limited bandwidth. To keep the high frequency texture in chroma components, we propose a luma-only resampling-based video coding method, which includes three aspects. Firstly, the luma component is downsampled while the chroma components remain unchanged in the pre-processing module. Secondly, a convolutional neural network-based (CNN) module is adopted to remove coding artifacts and restore the reconstructed frame. Finally, a single channel based super resolution method, which utilizes CNN as basic structure, is proposed to restore the original color format. Extensive experimental results demonstrate the benefits of the proposed method. Compared with VTM-11.0, the proposed method achieves {8.63%, 7.82%, 8.34%} BD-rate reductions on 4K sequences for Y, U, and V components under all-intra configuration, respectively.