Candidate: Wenbo (Paul) Yang
Date: May 19, 2025
Time: 10:00 AM
Location: Online
Supervisor: Zhou Wang
All are welcome!
Abstract:
Image degradations arise throughout the image formation and processing pipeline, from scene capture and camera hardware to post-processing, compression, and transmission. Although degradation modelling is central to image restoration, data augmentation, and visual effect synthesis, existing methods are often designed for specific degradation types, require strong domain knowledge, and depend on degradation parameters that are unavailable in real-world settings. This seminar presents a learning-based framework for universal image degradation modelling, with an emphasis on downstream vision tasks. Through the disentangle-by-compression principle, this framework learns degradation representation without requiring explicit supervision on degradation parameters. The learned model excels at degradation reproduction, transfer, and manipulations, enabling multiple downstream tasks, such as artistic effect synthesis, blind inversion-based image restoration, and dataset augmentation.