Data-driven gradient priors integrated into blind image deblurring

Abstract

Blind image deblurring is a severely ill-posed task. Most existing methods focus on deep learning to learn massive data features while ignoring the vital significance of classic image structure priors. We make extensive use of the image gradient information in a data-driven way. In this paper, we present a Generative Adversarial Network (GAN) architecture based on image structure priors for blind non-uniform image deblurring. Previous image deblurring methods employ Convolutional Neural Networks (CNNs) and non-blind deconvolution algorithms to predict kernel estimations and obtain deblurred images, respectively. We permeate the structure prior of images throughout the design of network architectures and target loss functions. To facilitate network optimization, we propose multi-term target loss functions aimed to supervise the generator to have images with significant structure attributes. In addition, we design a dual-discriminant mechanism for discriminating whether the image edge is clear or not. Not only image content but also the sharpness of image structures need to be discriminated. To learn image gradient features, we develop a dual-flow network that considers both the image and gradient domains to learn image gradient features. Our model directly avoids the accumulated errors caused by two steps of “kernel estimation-non-blind deconvolution”. Extensive experiments on both synthetic datasets and real-world images demonstrate that our model outperforms state-of-the-art methods.