Scholarlink

<p>Real-world deblurring encompasses heterogeneous degradation patterns, ranging from spatially varying motion blur in natural images to temporally correlated artifacts in video sequences. Existing unified architectures often fail to model these domain-specific dynamics, while specialized networks do not scale effectively across modalities. To overcome these limitations, we introduce OmniDeblur-MoE, a unified Mixture-of-Experts (MoE) framework designed for modality-aware image and video deblurring. The system employs two specialized experts: RestFormer, which utilizes dynamic spatial attention for single-image reconstruction, and SwinFormer, which incorporates deformable temporal attention with hierarchical windowed transformers for video reconstruction. A lightweight Data Classification Network (DCN) performs modality prediction and activates the appropriate expert, ensuring targeted feature extraction. A gated fusion module aggregates expert outputs, followed by a multi-scale refinement decoder that enforces cross-resolution consistency in the final reconstruction. By eliminating redundant shared encoders and enabling expert-level specialisation, OmniDeblur-MoE achieves improved computational efficiency and stronger generalization across modalities. Experimental results demonstrate significant performance gains, achieving approximately + 5.1% PSNR improvement on GoPro for image deblurring and ~ 24% reduction in temporal error on REDS for video deblurring, establishing the effectiveness of the proposed framework for multimodal reconstruction.</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

OmniDeblur-MoE: Cross-Modality Deblurring Using a Lightweight Transformer-Based Mixture-of-Experts (MoE) Framework

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Arti Ranjan,
  • Ashish Arya,
  • Ravinder M.,
  • Rajiv Kumar Nath

摘要

Real-world deblurring encompasses heterogeneous degradation patterns, ranging from spatially varying motion blur in natural images to temporally correlated artifacts in video sequences. Existing unified architectures often fail to model these domain-specific dynamics, while specialized networks do not scale effectively across modalities. To overcome these limitations, we introduce OmniDeblur-MoE, a unified Mixture-of-Experts (MoE) framework designed for modality-aware image and video deblurring. The system employs two specialized experts: RestFormer, which utilizes dynamic spatial attention for single-image reconstruction, and SwinFormer, which incorporates deformable temporal attention with hierarchical windowed transformers for video reconstruction. A lightweight Data Classification Network (DCN) performs modality prediction and activates the appropriate expert, ensuring targeted feature extraction. A gated fusion module aggregates expert outputs, followed by a multi-scale refinement decoder that enforces cross-resolution consistency in the final reconstruction. By eliminating redundant shared encoders and enabling expert-level specialisation, OmniDeblur-MoE achieves improved computational efficiency and stronger generalization across modalities. Experimental results demonstrate significant performance gains, achieving approximately + 5.1% PSNR improvement on GoPro for image deblurring and ~ 24% reduction in temporal error on REDS for video deblurring, establishing the effectiveness of the proposed framework for multimodal reconstruction.