Gradient-guided GAN for dynamic scene deblurring

Abstract

Dynamic scene blur, mainly caused by camera shake and motions, is one of the most common causes of image degradation. Recent GAN-based strategies have performance on deblurring tasks. To further improve the performance of GAN-based approaches on deblurring tasks, we propose Gradient-guided GAN for dynamic scene deblurring, it includes image restoration branch and gradient branch, which uses the gradient as a guide to supervise the restoration process. In particular, perform an attention fusion of feature image generated by restoration branch and gradient feature image generated by gradient branch, which using gradient information to guide the network to fully learn the deep feature information. Extensive experiments on GOPRO dataset show that our method achieve state-of-the-art performance in dynamic scene deblurring.