Reduced complexity multiscale cnn for in-loop video restoration

Convolutional neural networks (CNNs) have shown promising improvements in video coding efficiency when included in traditional block-based codecs as a loop filter. Unfortunately, these coding gains are often accompanied by significant increases in complexity, measured by the number of multiply-accumulate (MAC) operations, that make them intractable in practice. As a result, there is considerable interest in reducing complexity for these CNN-based approaches. In previous work, we have shown that multiscale CNNs provide a path to reduce the associated MAC count. In this paper, we extend our work to consider channel grouping, spatial support limitations and shallower network depths to further reduce the MAC count of these multi-scale architectures. We demonstrate that the method can achieve an average VMAF bitrate reduction of 6.1% and 2.6% for all intra and random-access coding respectively, when compared to the evolving AV2 standard. Complexity is reduced to 1.85k MACs per pixel, which is a 390x reduction over previously published results.