Motion Image Deblurring using AS-Cycle Generative Adversarial Network
Xiaoming Zhu, Lijun Yao, Fan Luo, Kejun Wang, Zhou Che, Jing‐Kun Yan, Min Zhou, Yongchang Cai, Lingling Wang, Zelong Cao, Lan Peng, Fengqing Bai, Zifang You, Xiao Hongqiu, Haocheng Qi · International Journal of Computer Applications · 2021
To improve the problem of poor generalization ability of image deblurring model in real scenes, this paper proposes a model named AS-CycleGAN (Cycle Generative Adversarial Network based on Asymmetric Samples).The model trains on unpaired images by using two "dual form" Conditional Generation Adversarial Networks, adopting global residual connection and ResNetv2 residual module.To enhance the texture effect, the SFT layer is integrated.The experimental results on the data set of Gopro show that the SSIM and PSNR values of our algorithm are 15.97% and 0.75% higher than those of the benchmark model CycleGAN, respectively.By improving the residual structure and adding the SFT layer, the effect is even better.AS-CycleGAN provides a powerful help to solve the motion blur problem in the actual scene.