Self-supervised image blind deblurring using deep generator prior

Abstract

Deep generative prior (DGP) is recently proposed for image restoration and manipulation, obtaining compelling results for recovering missing semantics. In this paper, we exploit a general solution for single image deblurring using DGP as the image prior. To this end, two aspects of this object are investigated. One is modeling the process of latent image degradation, corresponding to the estimation of blur kernels in conventional deblurring methods. In this regard, a Reblur2Deblur network is proposed and trained on large-scale datasets. In this way, the proposed structure can simulate the degradation of latent sharp images. The other is encouraging deblurring results faithful to the content of latent images, and matching the appearance of blurry observations. As the generative adversarial network (GAN)-based methods often result in mismatched reconstruction, a deblurring framework with the relaxation strategy is implemented to tackle this problem. The pre-trained GAN and pre-trained ReblurNet are allowed to be fine-tuned on the fly in a self-supervised manner. Finally, we demonstrate empirically that the proposed model can perform favorably against the state-of-the-art methods.