Improving NN-based low-complexity loop filtering for video compression
Tong Shao, Peng Yin, Sean T. McCarthy, Jay N. Shingala, Ajay Shyam, Ajat Suneja, Siddarth P. Badya · 2025
Low-complexity neural network-based loop filters (NNLF) have been shown to improve the performance of state-of-the-art video coding. This paper proposes to further improve performance of the Low Operation Point (LOP) loop filter, LOP3, being studied in the JVET by replacing the sequential 1x3/3x1 backbone block design with a parallel 1x3/3x1 backbone design. The same number of kMac/Pixel is maintained. The proposed backbone design reduces the number of sequential convolutional layers while achieving some coding efficiency improvement. Implemented on top of NNVC-10 using SADL, the fast stage 3 training results show that the BD-Rate for fp32 is {-0.08%, -2.34%, -3.01%} under AI and {- 0.15%, -2.49%, -2.95%} under RA when compared with NNVC-10 anchor. The results show that the BD-Rate for int16 is {-0.12%, -2.19%, -2.70%} under AI and {-0.09%, -2.22%, -3.03%} under RA when compared with NNVC-10 anchor.