BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030653Z
LOCATION:Meeting Room S423+S424\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251216T135300
DTEND;TZID=Asia/Hong_Kong:20251216T140400
UID:siggraphasia_SIGGRAPH Asia 2025_sess121_papers_2401@linklings.com
SUMMARY:Self-supervised Underwater Color Restoration via Wavelet-Diffusion
  Model with Filtered Multi-Scale Feature Distillation
DESCRIPTION:Xin Zhang and zhuang Zhou (Beijing Institute of Technology; Be
 ijing Institute of Technology, Zhuhai); Yixiao Yang (National University o
 f Singapore); Haijun Xie (Beijing Institute of Technology, Zhuhai); and Ha
 owen Yan, Hexiang Zhai, and Binghua Su (Beijing Institute of Technology; B
 eijing Institute of Technology, Zhuhai)\n\nExisting underwater image proce
 ssing methods often struggle due to the limited availability of real paire
 d training data. Models trained on public datasets frequently fail to gene
 ralize across diverse underwater conditions and produce suboptimal color r
 estoration. To address these challenges, we propose a self-supervised unde
 rwater color restoration framework based on a Wavelet-Diffusion Model with
  Filtered Multi-Scale Feature Distillation. Specifically, we introduce a w
 avelet-diffusion training paradigm on terrestrial images, guided by a stoc
 hastic underwater imaging model prior. This randomized control enables the
  model to learn diverse underwater imaging processes, facilitating effecti
 ve generalization to real-world underwater images and achieving precise co
 lor restoration. Furthermore, to tackle feature entanglement in zero-shot 
 domain generalization and mitigate the slow sampling and partial corruptio
 n issues of diffusion models, We integrate a Mamba-based U-shaped student 
 network for multi-scale feature distillation. Additionally, we introduce a
  filtering mechanism to refine the diffusion sampled features, allowing th
 e student model to outperform the teacher in both performance and image qu
 ality. Extensive experiments across multiple underwater datasets demonstra
 te that our approach effectively restores natural colors, eliminating wate
 r-induced distortions while achieving state-of-the-art performance in both
  qualitative and quantitative evaluations. Code and data for this paper ar
 e at https://github.com/zx826/FMFD\n\nRegistration Category: Full Access, 
 Full Access Supporter\n\nSession Chair: Paul Debevec (Eyeline, USC Institu
 te for Creative Technologies (ICT))\n\n
END:VEVENT
END:VCALENDAR