Image Gradient inspired Generative Adversarial Network for Non-Uniform Deblurring*
Qing Qi · 2023
This paper presents a method to address the dynamic scene deblurring task by integrating image gradient priors into a generative adversarial network (GAN). Although image deblurring has made significant progress, deep learning-based methods have not fully utilized image gradient priors. Image gradient priors are used to regulate the image recovery process and serve as a quantitative evaluation metric for assessing the quality of deblurred images. The proposed model uses a data-driven approach to learn image gradients and integrates them throughout the design of network structures. The network architecture includes a Gradient Extractor Convolutional Layer (GECL) that computes the 1st-order spatial derivatives in a data-driven way rather than the traditional edge detection operators. Target loss functions are proposed to constrain the network training. This image deblurring strategy eliminates the need for solving optimization equations and leverages deep learning to learn massive data features. Extensive experiments on synthetic datasets and real-world images show that our model outperforms state-of-the-art methods.