Transcoded Video Restoration by Temporal‑Spatial Auxiliary Network – AAAI 2022 Paper Overview

At the AAAI Conference on Artificial Intelligence 2022, a joint paper by Kuaishou's audio‑video algorithm team and Xi'an University of Electronic Science and Technology titled "Transcoded Video Restoration by Temporal Spatial Auxiliary Network" was accepted, highlighting a novel video compression‑artifact restoration algorithm.

The paper addresses the growing importance of video technology in post‑pandemic life, where videos often undergo multiple encoding and transcoding steps, leading to severe compression artifacts that degrade user experience.

Leveraging deep learning, the authors propose a video compression‑restoration algorithm that treats the initial encoding information as an auxiliary supervision label, designing both auxiliary and global supervision loss functions to guide network training.

The network architecture incorporates a Temporal Deformable Alignment Module (TDAM) and a Pyramid Spatial Fusion Module (PSFM) within a convolutional neural network, jointly trained with Auxiliary Supervision (ASAM) and Global Supervision (GSAM) to enhance visual quality of transcoded videos.

Experimental results show that the proposed method improves average PSNR by 0.782 dB (a 52.4% gain over the previous STDF method) and yields noticeable subjective quality enhancements, such as removal of zebra‑striping artifacts and restoration of fine textures.

Figures illustrate the decreasing acceptance rate of AAAI over six years, the workflow comparison between single‑compression and transcoding scenarios, the network architecture, and visual comparisons between the proposed method and prior single‑compression restoration techniques.

For more information, see the conference link https://aaai.org/Conferences/AAAI-22/ and the paper on arXiv https://arxiv.org/abs/2112.07948 .