Image Deblurring Based on Aggregate Residual Adversary Networks

Abstract

In order to solve the problems that the existed deblurring algorithms canonlyprocess a single blur type and need long running time, an image deblurring algorithm based on the aggregation residual generation confrontation network is proposed. Firstly, the generation confrontation network is used to generate the reconstructed image discriminant label, so that the final image is closer to the clear image. Secondly, the feature extraction module is combined with the aggregate residual network and the channel attention module to extract the useful feature information of the intermediate layer. Finally, the combination of WGAN’s Wasserstein-1 distance and perceptual loss is used as the loss function to train the model to ensure the consistency of the generated image and the clear image. In the PyTorch environment, the proposed algorithm is tested with the GOPRO dataset and the Kohler dataset, and compared with other algorithms, such as<italic> L</italic>₀ norm prior, dark channel prior, specific deblurring, DeepDeblur and DeblurGAN. Experimental results show that when restoring motion blurred images and Gaussian blurred images, the proposed algorithm outperforms other algorithms in terms of evaluating indicator such as PSNR and the running time.