DeblurGAN+: Revisiting blind motion deblurring using conditional adversarial networks

Abstract

This work studies dynamic scene deblurring (DSD) of a single photograph, mainly motivated by the very recent DeblurGAN method. It is discovered that training the generator alone of DeblurGAN will result in both regular checkerboard effects and irregular block color excursions unexpectedly. In this paper, two aspects of endeavors are made for a more effective and robust adversarial learning approach to DSD. On the one hand, a kind of opposite-channel-based discriminative priors is developed, improving the deblurring performance of DeblurGAN without additional computational burden in the testing phase. On the other hand, a computationally more efficient while architecturally more robust auto-encoder is developed as a substitute of the original generator in DeblurGAN, promoting DeblurGAN to a new state-of-the-art method for DSD. For brevity, the proposed approach is dubbed as DeblurGAN+. Experimental results on the benchmark GoPro dataset validate that DeblurGAN+ achieves more than 1.5 dB improvement than DeblurGAN in terms of PSNR as trained utilizing the same amount of data. More importantly, the results on realistic non-uniform blurred images demonstrate that DeblurGAN+ is really more effective than DeblurGAN as well as most of variational model-based methods in terms of both blur removal and detail recovery.