基于DeblurGAN的运动模糊图像盲复原算法研究

DeblurGAN方法利用条件生成对抗网络解决了端到端的图像去模糊问题，但存在图像边缘细节恢复不足以及鲁棒性不高的问题，针对此问题，提出一种基于DeblurGAN的运动模糊图像盲复原方法。在生成网络中，采用多尺度卷积核神经网络提取特征，并使用级联空洞卷积扩大神经元的感受野；采用自适配归一化方法代替原来生成器中使用的实例归一化方法。其次，引入了梯度图像L1损失，结合对抗损失和感知损失，将其作为图像去模糊的正则约束，使得生成图像的边缘特征更加清晰。实验结果表明，提出方法复原的图像峰值信噪比数值较DeblurGAN算法高出5.4%，结构相似性指标高出1%；在主观上清晰化效果较好，且消除了网格效应。

Research on blind restoration algorithm of motion blur image based on DeblurGAN

The DeblurGAN method uses Conditional Generative Adversarial Networks (cGANs) to solve the end-to-end image deblurring problem, but there are problems of insufficient image edge detail recovery and low robustness. Aiming at this problem, a blind restoration method of motion blurred images based on DeblurGAN is proposed. In the generative network, a multi-scale convolution kernel neural network is used to extract features, and cascaded atrous convolution is used to expand the receptive field of neurons; an adaptive normalization method is used to replace the instance normalization method used in the original generator. Second, the gradient image L1 loss is introduced, combined with adversarial loss and perceptual loss, as a regular constraint for image deblurring, making the edge features of the generated image clearer. The experimental results show that the peak signal-to-noise ratio of the image restored by the proposed method is 5.4% higher than that of the DeblurGAN algorithm, and the structural similarity index is 1% higher; the subjective clearing effect is better, and the grid effect is eliminated.