Neural Network-Based In-Loop Filter for CLIC 2022

Yonghua Wang, Jingchi Zhang, Zhengang Li, Xing Zeng, Zhen Zhang, Diankai Zhang, Yunlin Long, Ning Wang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 1774-1777


A hybrid video codec comprised of an optimized VVC codec and a convolutional neural network-based loop filter (CNNLF), was submitted in the video compression track in Challenge on Learned Image Compression (CLIC) 2022[1]. This paper presents the traditional methods and deep learning scheme in video coding optimization, which were adopted in the hybrid codec based on VTM-15.0. Traditional methods include QP adaptive adjustment of I frame and rate-distortion optimization based on SSIM. Meanwhile, the deep learning scheme proposes an adaptive CNNLF, which is turned on / off based on the rate-distortion optimization at CTU and frame level. The network architecture mainly consists of the attention residual module and the convolution feature maps module, which help extract image features and improve image quality. To balance performance and complexity, the proposed scheme sets different training parameters for 0.1 Mbps and 1 Mbps, respectively. The experimental results show that compared with VTM-15.0, the proposed traditional methods and adding CNNLF improve the PSNR by 0.4dB and 0.8dB at 0.1Mbps, respectively; 0.2dB and 0.5dB at 1Mbps, respectively, which proves the superiority of our method.

Related Material

@InProceedings{Wang_2022_CVPR, author = {Wang, Yonghua and Zhang, Jingchi and Li, Zhengang and Zeng, Xing and Zhang, Zhen and Zhang, Diankai and Long, Yunlin and Wang, Ning}, title = {Neural Network-Based In-Loop Filter for CLIC 2022}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2022}, pages = {1774-1777} }