GAN-based image deblurring: A comparison

In recent years, image deblurring algorithms based on Generative Adversarial Network (GAN) have been proposed. The motion deblurring has been studied for some time. The multi-scale architecture of convolutional neural networks (CNN) can clearly recover blurred images, but the processing time is long. Generative Adversarial Networks (GAN) performs well in style conversion and can clearly recover blurred images under a non-multi-scale architecture. The authors reviewed the popular papers related to DeblurGAN, compared the algorithm structure of different papers, related data sets and the accuracy of experimental results, which helped GAN based image deblur field to summarize and summarize, and enabled researchers to better promote the development of related fields. The authors review six papers, introduce their research motivation, research methods and experimental results, and show the differences of different deblur algorithms. Some excellent algorithms can work flexibly with a wide backbone, balancing performance and efficiency. They run faster than simple DeblurGAN while maintaining near-state of the art results.