Gradient-based conditional generative adversarial network for non-uniform blind deblurring via DenseResNet

Abstract

Abstract Blind image deblurring aims to recover the sharp image from a blurry image. The problem is seriously ill-conditioned and many existing algorithms based on kernel estimation require heuristic parameter adjustments and high computational cost, and cannot perform well on non-uniform motion blurs. To address this issue, image deblurring is viewed as an image translation problem in this paper. The authors solve it based on a conditional generative adversarial network (GAN), where the sharp image is restored by an end-to-end trainable neural network. Different from the generative network in basic conditional GAN, the proposed generator is based on dense blocks and residual network (DenseResNet), aiming to mitigate the problems of overfitting and vanishing gradient, and strengthen the blur feature propagation. To generate clear structure, the basic conditional GAN formulation is further revised by introducing joint VGG features and L 1 -based gradient loss. Extensive experimental results demonstrate the superior performance of the proposed method.