Motion deblurring method based on Improved DeblurGAN
Xinyan Zhang, Xiaofeng Wang · Academic Journal of Computing & Information Science · 2020
Generative Adversarial Networks(GANs) is a generation model that learns data distribution through the mutual against between generator network and discriminantor network.It has the advantage of generating clear and sharp samples, and has made progress in the application of image super-resolution and image repair. DeblurGAN solves the problem of end-to-end image deblurring by using conditional Generative Adversarial Networks(cGANs).In order to obtain better deblurring effect, this paper proposes an improvement based on DeblurGAN.Firstly, the method in this paper uses DenseBlock to replace the ResBlock in DeblurGAN, and adds two skip-connections. Finally, depthwise separable convolution is used to replace the common convolution block in the network, so as to reduce the network model, reduce parameters and accelerate the convergence speed of the network. The loss function uses the perceptual loss to ensure content consistency between the generated image and the clear image.