Image Restoration Technologies Based On Generative Adversarial Network

The rapid development of digital images in recent years has brought considerable challenges to digital image processing technology,of which image recovery has become an urgent problem in digital image processing.The reason is that the noise information in the image can seriously affect the reader’s judgement and understanding of the image information.In order to solve the problem of easy loss of image detail information and edge texture in the denoising process,a cooperative processing strategy of denoising and super-resolution reconstruction of noisy images is proposed,and the main work is as follows:Firstly,to address the requirement that traditional image denoising methods are difficult to achieve for high-quality image processing,while existing deep learning-based image denoising methods have the problem of losing part of the edge texture information and detail information,a cooperative processing strategy for image denoising and super-resolution reconstruction based on generative adversarial networks is proposed.The strategy divides the image recovery into two stages,the first stage is the image denoising process based on generative adversarial network,and the second stage is the image super-resolution reconstruction process based on generative adversarial network,where the second stage image reconstruction process is mainly to compensate for the lost detail part of the denoised image in the first stage.Then,to address the problem that the De GAN denoising network can only denoise grey-scale images,the number of channels of the De GAN denoising network is changed by changing the output of the last layer of the generative network and the output of the last layer of the discriminative network from the original single-channel image to a three-channel image;the convolution and activation functions of the first two layers of the feature extraction network are removed,and the number of channels is changed so that the De GAN The change in the number of channels gives the De GAN network the ability to denoise colour images.However,as colour images contain more information than grey-scale images,the denoising effect of the De GAN network is poorer if only the network channels are changed.The structure is improved by using U-Net++ with cross-layer connections instead of the original U-Net,and the improved generative network is able to fuse image features at different scales;the loss function is optimised because only a single additive Gaussian white noise is added to colour images in this paper,whereas De GAN is designed for multiple mixed noise image removal.The optimization is based on the performance of the four loss functions in the ablation experiment,and the four loss functions are given different weight values and then linearly summed to achieve the improvement in denoising accuracy.Finally,improvements to the SRGAN image reconstruction network are proposed to address the problem that the image quality of the image after the first stage of denoising is sometimes slightly worse after the second stage of reconstruction than after only the first stage of denoising.The improvement treatments are: to avoid the loss of image information in the pre-processing process,the scaled-down sampling in the pre-processing part is eliminated and the original size image is directly used for reconstruction;a dataset is added to the discriminative network for saving the baseline comparison image,which is the target image of the discriminative network,to improve the upper limit of the original reconstructed image quality;the similarity structure degree loss function that performs better in the denoising process of the De GAN generation network is The similarity structure loss function,which has performed well in the denoising process of the De GAN generation network,is introduced into the loss function of the SRGAN generation network to improve the objective evaluation index of similarity structure of the reconstructed images.