In-Loop Filter with Dense Residual Convolutional Neural Network for VVC
Sijia Chen, Zhenzhong Chen, Yingbin Wang, Shan Ting Liu · 2020
In-loop filters are usually used to remove artifacts or further improve coding performance in lossy video compression. Encouraged by the superior performance of deep learning on image restoration, a dense residual convolutional neural network (DRN) based in-loop filter is proposed in this paper for Versatile Video Coding (VVC). The proposed network utilizes residual learning module, dense shortcuts and bottleneck layers, to solve the problems of gradient vanishing, encourage features reuse and save computational resources, respectively. We present the modifications of DRN in terms of network structure, training materials and training process for the trade off between computational complexity and coding efficiency, and valid the effectiveness of the modifications based on the experimental results. Besides, the proposed framework outperforms the VVC reference software with three coding configurations.