# Task-Uncertainty-Aware Video Restoration for Time-varying Unknown Degradations
Wenrui Li, Hongtao Chen, Ruyi Zhang, Zhe Yang, Wangmeng Zuo · IEEE TMM · Accepted
# Problem
Real-world videos suffer from time-varying unknown degradations (TUD): the types and severities of corruption change unpredictably across frames. Uniformly averaging losses lets easy degradations dominate the gradients, so hard and composite degradations stay under-optimized.
# Method: DOVENet
- Restoration is formulated as MAP inference with spatial priors and temporal consistency, implemented as an unrolled network with degradation-aware alignment and restoration updates.
- Task-uncertainty regularization (TUR) is extended to video: frame-specific uncertainty plus degradation presence masks produce adaptive loss reweighting that emphasizes hard cases, without changing inference.
DOVENet consistently improves performance and generalization on multiple TUD benchmarks while staying efficient and scalable.
# My Contribution
Validation of the spatio-temporal uncertainty regularization and quantitative comparison across benchmarks.
# Task-Uncertainty-Aware Video Restoration for Time-varying Unknown Degradations
Wenrui Li, Hongtao Chen, Ruyi Zhang, Zhe Yang, Wangmeng Zuo · IEEE TMM・已接收
# 研究问题
真实场景中的视频常受到 ** 时变未知退化（TUD）** 的影响：退化的类型和程度在帧与帧之间不可预测地变化。若对各类损失简单取平均，梯度会被容易的退化主导，困难退化与复合退化则长期得不到充分优化。
# 方法：DOVENet
- 将复原问题建模为结合空间先验与时间一致性的 MAP 推断，并以展开式网络实现，其中包含退化感知的对齐与复原更新模块。
- 将 ** 任务不确定性正则化（TUR）** 推广到视频：利用帧级不确定性与退化存在掩码自适应地重新加权损失，突出困难样本，且不改变推理过程。
DOVENet 在多个 TUD 基准上持续提升了复原性能与泛化能力，同时保持高效、可扩展。
# 我的贡献
负责时空不确定性正则化的验证，以及在多个基准上的定量对比实验。
