跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2606.06176 · v5

RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision

首次提交 2026-06-04 21:48
版本更新 2026-11-17 17:43
雷达观测 2026-11-18 11:33
arXiv ↗PDF ↗

摘要

Learning-based underwater image enhancement (UIE) often relies on paired datasets whose reference images vary substantially in quality. Treating these references as equally reliable targets limits enhancement performance, while filtering pairs with modest corrections reduces the available supervision. We propose RQUL-UIE, a framework that uses estimated input-to-reference correction to determine where each reference supervises a shared recurrent enhancer. Specifically, level-wise latent space enhancement (LLSE) assigns smaller estimated corrections to shallower supervision stages and larger corrections to deeper stages, allowing all pairs to contribute without additional references or quality annotations. To compensate for detail loss caused by the variational Autoencoder (VAE)'s lossy compression, texture refinement via frequency modulation (TRFDM) draws on cues retained in the raw image. Evaluated against 12 baselines, the proposal achieves the best overall performance across four metrics on the three evaluation collections. Qualitative comparisons further show improved color balance and clearer structures in degraded underwater scenes. Project link https://github.com/Haru2022/RQUL-UIE

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

本站首次收录为 v5,暂无更早版本可对照。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v5 · 2026-11-17 17:43 · 最新观测版本

RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision

Learning-based underwater image enhancement (UIE) often relies on paired datasets whose reference images vary substantially in quality. Treating these references as equally reliable targets limits enhancement performance, while filtering pairs with modest corrections reduces the available supervision. We propose RQUL-UIE, a framework that uses estimated input-to-reference correction to determine where each reference supervises a shared recurrent enhancer. Specifically, level-wise latent space enhancement (LLSE) assigns smaller estimated corrections to shallower supervision stages and larger corrections to deeper stages, allowing all pairs to contribute without additional references or quality annotations. To compensate for detail loss caused by the variational Autoencoder (VAE)'s lossy compression, texture refinement via frequency modulation (TRFDM) draws on cues retained in the raw image. Evaluated against 12 baselines, the proposal achieves the best overall performance across four metrics on the three evaluation collections. Qualitative comparisons further show improved color balance and clearer structures in degraded underwater scenes. Project link https://github.com/Haru2022/RQUL-UIE

查看此版本解读 →查看此版本原文 ↗