Compressed Video Restoration

The School of EECS is hosting the followingHDR Progress Review 1 Confirmation Seminar:

Compressed Video Restoration

Speaker: Shuyun Wang 
Host/Chair: Prof Xue Li

Abstract

Video content has become the dominant form of data on the Internet, with streaming projected to account for over 80–90% of global traffic by the mid-2020s. To reduce bandwidth and storage overhead, videos are often heavily compressed (e.g., H.264, H.265). Compressed video restoration has thus become a crucial task to counter quality degradation caused by compression, transmission errors, and low-quality sources. It encompasses sub-tasks like super-resolution, enhancement, and inpainting, each targeting specific types of degradation. In practice, these artifacts often co-occur, demanding joint and adaptive restoration strategies.

Existing methods still struggle with real-world degradations. Many overlook the rich priors in codec metadata, rely on synthetic corruption assumptions, or require predefined masks, limiting their robustness. In cases of severe visual damage, models also fail to leverage semantic information from language, which can guide content-aware recovery when visual cues are insufficient. Furthermore, current models operate with static pipelines, lacking the ability to perceive complex degradation patterns and dynamically plan and execute specialized restoration procedures.

To bridge these gaps, we aim to develop a unified, autonomous framework for compressed video restoration that integrates codec metadata, visual priors, and natural language semantics. Our approach combines codec-aware super-resolution, blind bitstream-corrupted video recovery, language-guided video recovery, and an autonomous restoration agent capable of degradation-aware reasoning and task orchestration. These contributions collectively move toward scalable, generalizable, and semantically guided restoration pipelines suitable for real-world compressed video.

Bio

Shuyun Wang is a second-year Ph.D. student at the School of EECS, University of Queensland, under the supervision of Dr Xin Yu and Dr Miao Xu. He obtained his Bachelor's degree in Software Engineering from Hebei Normal University, and his Master's degree in Computer Technology from Hebei University of Technology. His research interests include video restoration and video inpainting.