Research On Image Denoising Method Based On Resnet Structure And Generative Adversarial Network

Image is one of the main ways to obtain information and plays an important role in many fields.Image denoising technology refers to the process of reducing digital image noise and improving the visual effect of images,this is the basic premise for the visual system to understand and recognize images.How.to find a better balance between reducing noise and retaining details has always been a research hotspot.At present,various denoising methods have a good performance in a weak noise environment,but in a strong noise environment,the denoising effect is degraded due to insufficient internal available information.So,this paper proposes an Image Denoising Generative Adversarial Network model based on the research and analysis of two models of Super-Resolution Generative Adversarial Network(SRGAN)and Image De-raining Conditional Generative Adversarial Network(ID-CGAN).The model combines the ResNet structure and consists of five residual blocks to form a denoising core module.Each residual block contains two convolutional layers,batch normalization and Lrelu activation functions,and the skip connections are involvecd to make the network efficient in training and have better convergence performance.At the same time,in order to reduce the chessboard artifacts in the image denoising results,this paper also proposes a new smooth loss,which combines adversarial loss,pixel loss and feature loss to form a joint loss function.In the training process of the network model,the VOC2012 data set is used as the training data,and the generator learns the end-to-end mapping of the noise image to the ground truth through the multi-convolution and sub-pixel layers,and the discriminator is used to combat the training.When the similarity between the denoised image and the real image is low,the discriminator automatically supervises the generator to train in a better direction with a larger loss function value,and causes the network model to generate a denoising result image consistent with the ground truth distribution to achieve the purpose of denoising.In the experimental process,the network model of this paper is compared with the the loss function improvement and other image denoising methods.Judging by objective evaluation indicators and subjective visual evaluation:compared with before the loss function is improved,the PSNR value of the denoising result image of this model is increased by 1.59dB on average,and the phenomenon of artifacts is reduced visually;compared with the well-known BM3D and DnCNN models,the PSNR average values of the model denoising results are higher than 3.97dB and 2.79dB,respectively,and the image details and edge features can be well preserved.This proves the effectiveness of the network model denoising.