Enhanced multi-scale feature progressive network for image Deblurring

This paper tackles the problem of single image motion blur removal. Recently methods have achieved state-of-the-art results owe to multi-scale, scale-recurrent and coarse-to-fine architecture, however, the problem of image feature information extraction and information transfer between different stages has not been well solved. In this paper, first, an efficient Enhanced Multi-scale Feature Progressive Network (EMFPNet) was proposed, in order to solve the above problem, a multi-scale feature extraction module is applied in each stage to enrich the spatial features of the maps. Second, introducing a Cross-stage Feature Fusion module to solve the problem of information transmission in different stages. Third, a cross-stage attention mechanism is used to monitor and help the transmission of information. Compared to SOTA method, our method achieve 0.6% and 0.2% improvement in PSNR respectively on GoPro and HIDE datasets.