Research On Image Denoising Algorithm Based On GAN

The quality of digital image is inevitably affected by acquisition equipment and environment in the process of its generation,as well as various factors during its storage and transmission,which produces some random noise in the acquired digital image and makes it more difficult to conduct the follow-up procedures.Therefore,the image produced requires to be denoised.Up to now,the image denoising algorithm mainly adopts the approach of filtering,in which convolution is conducted between the image and the set filter and the pixels in the image be filtered.However,the image is inevitably rendered too smooth after noise filtering,whose edge details and texture have declined.In terms of the problems above,this paper mainly explores a sort of image denoising method based on the combination of generative adversarial networks and wavelet transform that the image is divided into different subbands by wavelet transform where generative adversarial training is adopted.The denoising networks obtained from different subbands can respectively realize the mapping from noise image subbands to denoising image subbands,and then the images of different subbands are restored to the original ones by inverse wavelet transform.In the selection and training of the network model,the hyperparameters including the number of network layers and the number of neurons in the traditional neural network model are artificially set in advance,while the network training can only update the parameters such as weight and bias.Based on the optimization of hyperparameter setting,this paper adopts the method of neural architecture search for the purpose of the acquisition of the optimal generator network structure hyperparameters.By the establishment of the search space and the application of the search strategy based on genetic algorithm,the network model is updated after a certain amount of training and the following search direction is determined by the denoising quality of the network model.Finally,the most appropriate network structure is selected for the generator network.The training,representing the idea of adversarial training in the cycle generative adversarial network,adds the adversarial process to the training of the traditional neural network model,as well as the adversarial loss,consistency loss,cycle consistency loss and denoising loss with different weights as loss functions,to form a condition of continuous conflict between the generator and discriminator,which improves the denoising performance of the model.This paper conducted the method of transfer learning than the search of the model structure directly in the process of generative adversarial training.The complete structure search process was put into the traditional neural network denoising model and the model structure and parameters applicable for image denoising are obtained from the traditional training model.Then the generative adversarial training,as the initialization parameters of the model structure and the corresponding model in the generation adversarial training,was carried out based on the results above.The part of experiment launched the research of the wavelet basis function and wavelet decomposition scale to select the optimal parameters.Through comparative experiments,it is proved that wavelet transform and network structure search can improve the denoising performance of the model.In comparison with other traditional denoising algorithms and the ones based on neural network model,the average peak signal-to-noise ratio of the images which are denoised by the algorithm in this paper turns out to be 26.45 dB with the average structure similarity of 0.86,which is higher than that of other filtering algorithms.From perspective of the subjective evaluation,the image clarity and the protection of fine structure after denoising by this algorithm are better than those of other neural network models.And the results show that this algorithm has a certain ability to remove strong noise and real environment noise.