Residue guided loop filter for HEVC post processing

The block-based coding structure in the hybrid coding framework gives rise to the obvious artifacts such as blocking, ringing .etc. Recently, some Convolutional Neural Network (CNN) based works apply reconstruction as the only input to reduce the artifacts. Though the performance of these works relying on powerful learning ability surpasses traditional loop-filter based methods in the High Efficiency Video Coding (HEVC) standard, how to enhance the high frequency signal is still not addressed. In addition to reconstruction, we first propose using the residue as the other input of our CNN-based loop filter. In essence, the residual signal as a high frequency indicator guides the CNN to augment the high frequency signal such as sharp shape and edge information. Second, we find out that the reconstruction and residue signals have different characteristics and should be handled with different network structures. For the reconstruction, we develop an All Frequency (reconstruction) CNN (AF-CNN) adopting the down sampling and up sampling pairs to learn all frequency signal with the global information. For the residue, we devise a High Frequency (residual) CNN (HF-CNN) customizing the Residual Blocks to adapt to the high frequency signal information. To the best of our knowledge, this is the first work that employs residual signal as a vital independent high frequency input to direct the learning of CNN- based loop filtering. We implement the proposed algorithms in the HEVC reference software. The experimental results show that our proposed approach of dual inputs of Residual and Reconstruction with HF-CNN and AF-CNN respectively (RRHA) presents significant BD-rate savings compared with the current CNN-based scheme.