Blind Motion Deblurring Based on Generative Adversarial Networks

Abstract

In the past two years, GAN (Generative Adversarial Networks) has emerged and been applied to the image deblurring problem, showing good results, especially in restoring high-frequency texture details of the image. However, there are very few papers on GAN-based image deblurring so far, and it is not ideal for the processing of edge features. In this paper, an end-to-end blind image motion deblurring algorithm based on GAN is proposed. In the pixel domain, we use the weighted sum of cross entropy and L1 loss as the loss function. In the feature domain, the weighted sum of the features extracted by VGG and DenseNet is used to calculate the loss. And we add "deconvolution + PixelShuffle" module to the network. Experiments show that our method achieves the excellent performance in terms of PSNR (Peak Signal to Noise Ratio) and SSIM (Structural Similarity Index), simultaneously eliminates the checkerboards effectively.