Scholarlink

<p>Real-world video degradation shows pronounced spatial non-uniformity and temporal dynamics. Conventional unified models struggle with time-varying unknown degradations (TUD), often causing inaccurate frame alignment, insufficient degradation adaptation, and temporal inconsistencies. To address these challenges, we propose Prompt-Guided Dynamic Expert Network (PGDENet), a prompt-driven All-in-One Video Restoration network. By integrating prompt learning and a mixture of experts mechanism into a recurrent propagation paradigm, PGDENet achieves adaptive and efficient restoration of complex TUD. To reduce alignment errors, we design the Prompt-Guided Deformable Alignment (PGDA) module, which uses content-adaptive dynamic prompts to jointly constrain alignment and modulate features. For diverse degradations, we introduce the Mixture of Dimensions Experts (MoDE) system, employing dual-branch (spatial and high-frequency) routing to sparsely activate optimal sub-networks from a heterogeneous expert pool, balancing efficiency and performance. We further propose the Dynamic Prompt Expert Modulator (DPEM), which generates input-adaptive modulation signals from a learnable visual prompt pool, enhancing semantic representation while providing reliable degradation-aware priors for expert routing. A key-frame guided strategy is also incorporated to maintain robust global temporal consistency in long sequences. Extensive experiments validate the effectiveness of PGDENet on TUD restoration. Comprehensive evaluations are conducted on two synthetic datasets, each featuring seven degradation types with randomly varying corruption levels. For details, please visit: <a href="https://github.com/ycity16/PGDENet">https://github.com/ycity16/PGDENet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prompt-guided dynamic expert network for all-in-one video restoration

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Pengcheng Yu,
  • Jinhua Wang,
  • Ning He,
  • Guangmei Xu,
  • Xiaoyue Ma,
  • Jinxiu Chen

摘要

Real-world video degradation shows pronounced spatial non-uniformity and temporal dynamics. Conventional unified models struggle with time-varying unknown degradations (TUD), often causing inaccurate frame alignment, insufficient degradation adaptation, and temporal inconsistencies. To address these challenges, we propose Prompt-Guided Dynamic Expert Network (PGDENet), a prompt-driven All-in-One Video Restoration network. By integrating prompt learning and a mixture of experts mechanism into a recurrent propagation paradigm, PGDENet achieves adaptive and efficient restoration of complex TUD. To reduce alignment errors, we design the Prompt-Guided Deformable Alignment (PGDA) module, which uses content-adaptive dynamic prompts to jointly constrain alignment and modulate features. For diverse degradations, we introduce the Mixture of Dimensions Experts (MoDE) system, employing dual-branch (spatial and high-frequency) routing to sparsely activate optimal sub-networks from a heterogeneous expert pool, balancing efficiency and performance. We further propose the Dynamic Prompt Expert Modulator (DPEM), which generates input-adaptive modulation signals from a learnable visual prompt pool, enhancing semantic representation while providing reliable degradation-aware priors for expert routing. A key-frame guided strategy is also incorporated to maintain robust global temporal consistency in long sequences. Extensive experiments validate the effectiveness of PGDENet on TUD restoration. Comprehensive evaluations are conducted on two synthetic datasets, each featuring seven degradation types with randomly varying corruption levels. For details, please visit: https://github.com/ycity16/PGDENet.