1. Deformable multi-scale fusion network for non-uniform single image deblurring.
- Author
-
Zhang, Zhizhou, Chen, Yang, Zhu, Aichun, and Liu, Hanxi
- Abstract
Non-uniform image deblurring is an ill-posed problem. Previous research efforts attempt to solve this problem by increasing the number of scales processed in the model, including but not limited to multi-scale methods, multi-patch methods, and atrous convolution. However, these methods are still subject to the fixed geometric structures, which are inherently unable to adequately handle complex blur. This paper proposes a novel residual block called Deform-ResBlock that is composed of traditional convolution and deformable convolution to enhance the model's capability of modeling geometric transformations. Then, we design parallel multi-scale convolution streams composed of densely Deform-ResBlock for extracting multi-scale features. Finally, we apply the multi-patch approach stacking two stages to deblur images gradually. The overall method is named deformable multi-scale fusion network (DMSFN). Compared to the previous methods, our method combines the advantages of multi-scale and multi-patch approaches and has better modeling geometric transformation capability. Extensive experimental results on the GoPro, HIDE, and RealBlur datasets demonstrate that the proposed method performs favorably against the state-of-the-art in the non-uniform image deblurring. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF