A Two-stage Network for Image Deblurring

Blind deblurring is a typical challenge in image processing, carried out to correct various complex types of distortions that occur in the real world. Although learning-based deblurring methods have substantially outperformed the traditional algorithms, they fail to make full use of image priors, leading to inconsistent data distributions between the restored images and sharp images. In this paper, we divide the deblurring process into two steps and propose a two-stage network. The first stage generates an initial deblurred image using a common convolutional network. The second stage converts the initial data distribution into a latent sharp image distribution to obtain sharp edges through a prior network. Moreover, we propose a relativistic training strategy to train the prior network, which aims to learn the latent sharp image priors. In addition, we use a coarse-to-fine multiscale framework that shares weights between the different scales. The experimental results show that compared with other state-of-the-art methods, our method achieves competitive performances on benchmark datasets.