Image Denoising Using An Improved Generative Adversarial Network With Wasserstein Distance

Digital image is an important carrier for human perception and dissemination of information in the era of big data.Different degrees of noise will be introduced in the process of digital image generation and preservation.Poor image quality will have a great impact on the acquisition of information,making subsequent images.There is great uncertainty in processing,which hinders the development of image processing technology.The widespread application of image technology in many fields has made the research of digital image quality enhancement technology a very realistic content in image processing.As a preprocessing problem for image research,image denoising has far-reaching significance.The image denoising discriminant model has received extensive attention in recent years due to its good denoising performance.Generative adversarial networks have excellent generation effects in unsupervised learning,but also have some defects,such as gradient disappearance,unstable training,easy overfitting,difficult to converge and other issues.In order to solve the above-mentioned shortcomings and the traditional denoising methods that destroy the visibility of important structural details after image denoising,this paper proposes an image denoising algorithm based on Wasserstein Generative Adversarial Network.which adds multi-level convolution of the generative network.Obtain more image features,and add multiple residual blocks and global residuals to extract and learn the features of the input noise image to avoid feature loss.The innovation of this algorithm is that it can obtain information from the deep network unsupervised to complete image denoising.When improving the traditional denoising algorithms such as blurring and unclear edge restoration,the combination of deep learning further improves the image generation and denoising effect.The main work of this paper is as follows:(1)Introduce a deep convolution structure to generate a confrontation network for image denoising processing.At the same time,this article improves the deep convolution structure,introduces an eight-layer residual structure into the generator,and proposes the Residual-GAN algorithm for network training,so that The main features of the image will not be lost in the process of convolution,which improves the clarity and realism of image generation.At the same time,the jump connection structure avoids the problem of gradient disappearance in the network training process.(2)In order to improve the problem that Residual-GAN is prone to instability in training,combine the Deep Convolution Generative Adversarial Network(DCGAN)and Wasserstein Generative Adversarial Network(WGAN),and use the GAN network with Wasserstein distance instead of the original network,which is effective through Wasserstein distance Measure the similarity between the generated distribution and the original distribution,improve the network structure and the number of residual blocks,and introduce a spectral normalization algorithm to make the network meet the Lipschitz continuity without changing the parameter matrix structure,and enhance the network training ability.(3)This paper defines the weighted sum of the confrontation loss and feature loss as the loss function of the network.First,randomly select images in the BSDS500 data set for preprocessing,and then input the confrontation network training,and input the denoised image into the improved network to generate a denoised image.The method in this paper inputs the generated image and the original image into the VGG19 network,uses a deep convolutional network to extract features,and calculates the Euclidean distance between the features,which can promote the acquisition of more image detail information,that is,use the features extracted from a specific level to calculate the target Loss,which further guides the training of the generation network,makes the texture details of the denoised image clearer,and finally,this method is applied to the denoising of CT images,and the denoising effect is excellent.Finally,the denoising effect is verified by the image evaluation index PSNR and SSIM value.After denoising,the average PSNR of the image is 30.71dB,and the average SSIM is 0.8473.Under the three noise levels ?=15,25,50,PSNR is compared with other image denoising methods.The values are increased by 1.2dB,1.275dB,2.415dB,and the SSIM values are increased by 0.125,0.175,and 0.275,respectively.At the same time,a clearer visual effect is obtained.Compared with the effect of using the Residual-GAN network for denoising,the experimental results prove that the network training is more stable and the denoising effect is better.