Asymmetric Convolutional Residual Network for AV1 Intra in-Loop Filtering

In video compression standards, in-loop filtering plays an important role in alleviating blocking, blurring and ringing artifacts caused by lossy compression, which enhances visual quality and benefits coding efficiency. The boom of neural network applications in super-resolution and image restoration brings insights into solutions of in-loop filtering in video codecs. In this paper, we design an asymmetric convolutional residual network (ACRN) for in-loop filtering in the state-of-the-art AV1 codec. With the asymmetric convolutional blocks, directional features can be extracted to restore textures and improve quality. The cascading structure of wide-activated residual blocks with pruned dense connections enables reflecting hierarchical coding unit (CU) partition characteristics of video coding without losing overall details. Experiments show that the proposed lightweight ACRN can bring up to 12.78% coding efficiency improvement in intra coding of AV1.