A Fusion Network for Non-Uniform Deblurring*

In the field of computer vision, non-uniform image deblurring is a crucial and difficult task. By learning features from receptive fields, existing image deblurring algorithms have made advanced progress. However, non-local feature representations, which depict the global data distribution of blurry images are not taken into account. In this paper, we investigate non-uniform task by integrating local and non-local features. Specifically, we develop a DRDDBU (Dense-in-Residual Dense Dilation Blocks Unit) and a SAB (Scale Attention Block) to implement local and non-local, respectively. DRDDBU has the virtue of dense connections embodied in locally dense blocks and globally dense connections, which reuses and enhances all intermediate features. SAB is developed to preserve significant and suppress irrelevant responses for generating latent images. In addition, multiple loss functions are proposed to enhance network training and encourage convergence. Subjective and objective comparison experiments on various datasets are done to illustrate the efficiency of the suggested strategy. On synthetic datasets and real images, our non-uniform deblurring method outperforms state-of-the-art (SOTA) methods.