Super Resolution with Luma Component Enhancement for Down-up Sampling Based Video Coding in VVC

Recent down-up sampling video coding methods by CNN-based super resolution (SR) have demonstrated the potential of coding performance improvement. However, these methods generally process shallow features by using only a single layer of convolution, and without high resolution side information, resulting in lack of abundance of features fed into the backbone. In this paper, we propose a luma enhancement module (LEM), which internally consists of a multi-branch structure. Each branch is used for extracting features at different scales, enriching the diversity of shallow features of the model. For the input side information, we use reference picture re-sampling (RPR) images that is not used in the SR model adopted by NNVC-6.0. We consider RPR as a HR frame obtained by up-sampling with traditional conventional filters, which implicitly contains texture information at HR domain.