Convolutional Capsule Network Based in-Loop Filter for HEVC
Mengqing Cao, Yue Yu, Jian Chen, Xiuzhi Yang, Zhifeng Chen · 2021
Recently, several convolutional neural network (CNN)-based in-loop filtering algorithms are proposed to improve the high efficiency video coding (HEVC). However, regular CNN-based filters can only apply a single model to the whole image, but a single model usually cannot adapt well to all local features in the image. To solve this problem, we propose an inloop filtering algorithm based on convolutional capsule network (CC-net), and adapt it into the HEVC hybrid video coding framework as a new in-loop filter. This article is the first to apply capsule to the filtering work of video encoding, which is proposed to use localized dynamic routing algorithm to improve the self-adaptability of the filter to different local features in the image, and we integrate the model into HEVC encoding loop. Experimentally, our proposed method brings average 7.9%, 5.4% and 4.1% BD-BR reductions under all intra, random access and low-delay P configurations, respectively, as well as, 0.4dB, 0.2dB and 0.2dB BD-PSNR gains respectively.