Research On Image Denoising Algorithm Based On Generative Adversarial Network

Digital images are often affected by noise interference during the acquisition and transmission process.Image noise not only seriously degrades the visual quality of images,but also easily leads to the failure of methods such as edge detection and feature extraction,which seriously affects the subsequent analysis and processing of images.How to effectively remove image noise and accurately recover image edge details has become an important research goal of image filtering.This article studies the image denoising method based on generative adversarial networks based on the statistical properties of image noise,combined with the advantages of generative adversarial networks in both learning ability and data generation ability.The salt & pepper noise removal algorithm based on generative adversarial network is proposed based on the statistical characteristics of the image noise,combined with the data generation capability of generative adversarial network.The generator is designed as a parallel network structure,which is conducive to generating more texture details;By introducing residual modules with skip connections,the gradient disappearance of deep networks is avoided.By increasing the network width rather than depth,we can obtain more features without increasing the time cost.In addition,in order to improve the denoising ability of the network,the compound loss function that reflects both the stability of the network,the effectiveness of denoising and the stability of the original image content is constructed.The experimental results show that the denoising network proposed in this article has better performance compared with similar algorithms in terms of intuitive perception and objective evaluation indexes under the tests of Set14 and BSD300 datasets.For the complexity of the real image noise,a generation adversarial network-based denoising model is proposed.To address the lack of image texture detail generation capability of traditional generators,the generator is designed as a U-net derived network structure and adopts dense residual blocks to better extract image features and reduce detail loss.The discriminator network adopts a full convolutional network architecture to achieve pixel-level classification of images to improve the discriminator performance.In addition,an enhanced network structure is designed to enhance the denoising performance of the network.For the design of the loss function,firstly,WGAN-GP is used as the adversarial loss of this network to address the problem of instability in the training process of traditional GAN.Secondly,perceptual loss is added to improve the features of the generated images.Finally,the feature matching loss combined with multi-scale feature discrimination is used to constrain the different features between the generated image and the original image to construct a composite loss function that can reflect the network generation ability,denoising effect and training stability.Both the subjective and objective evaluation results of the simulation experiments show that the visual effect and denoising performance of the algorithm are better than those of similar algorithms.