Multi-scale Fusion for Dynamic Scenes Deblurring*

This paper develops a method for addressing the dynamic scene deblurring task by proposing a Multi-Scale Feature Extraction and Fusion Module (MSFEFM). Despite image deblurring methods based on "coarse-to-fine" are developed and have made significant progress. However, the idea of multi-scale feature extraction and fusion at the feature level has not received much attention. This paper builds upon the foundation of multi-scale feature encoding using a U-shaped network architecture, achieving extraction and fusion of multi-scale features. Instead of focusing on extracting and training features at multiple scales from blurry inputs. Extensive experiments on both synthetic datasets and real-world images demonstrate that our model surpasses state-of-the-art (SOTA) methods.