Improving SRN-Deblur Performance Based on Blur Direction-based Classification

With the development of neural networks in the field of image processing, deblurring is one of the important tasks of image processing, more and more excellent models have appeared one after another. We propose a classification-based strategy that assumes and verifies that the blurred images can be divided into different categories, and then feeds the blur images to the corresponding deblur models after classification. Our classification-based method can generally improve the performance of the current deblur models. After experimental verification, for the same dataset, our method improves PSNR from 30.26 to 31.15 and SSIM from 0.9342 to 0.9374.