关键词: 去运动模糊, 生成对抗网络, 互信息, 低秩分解, 对称跳跃连接, 互信息损失, 梯度图像 L1 损失

Abstract: An end-to-end image motion deblurring method based on DeblurGAN was proposed with conditional generative adversarial network． In this method，the standard convolution layers in DeblurGAN were changed into bottleneck structures，and low-rank decomposition was further performed on the convolution layers in the bottleneck structures． Then two residual symmetric skip connections were added to accelerate the convergence of the network． In order to solve the problem that the restored images in the DeblurGAN are not clear，mutual information loss and the gradient image L1 loss were added to the network loss function． By maximizing the mutual information between the input image and its hidden feature，the extracted hidden feature can well represent the input information， thereby obtaining a clear restored image from the hidden feature，and the L1 loss helps to make the restored image with more significant edge． At the same time，the effectiveness of the proposed method was verified by experiments and compared with other existing similar algorithms． The results show that compared with DeblurGAN，the peak signal-to-noise ratio of the proposed method is higher; the structural similarity measure of the two methods is equiv- alent; the parameter quantity of our model is compressed to 3. 25% of DeblurGAN; the deblurring processing speed is increased by 3 times，and our model outperforms other existing similar algorithms．

Key words: motion deblurring, generative adversarial network, mutual information, low-rank decomposition, symmetric skip connection, mutual information loss, gradient image L1 loss