Image Restoration Algorithm Based On Generative Adversarial Networks

Image restoration which restores degraded images to clear images has been a research hotspot in the field of digital images.Image deblurring and image super-resolution are classic problems in image restoration.Efficient and high-quality image restoration helps people to obtain richer data information,and plays a huge role in image recognition,image transmission,and video surveillance.Traditional method requires a lot of calculation and the quality of the generated images is not high.With the development of deep learning,CNN-based algorithms have been widely studied because they can restore images with high efficiency and high quality.In recent years,due to the superiority of GAN in image detail generation,it has begun to attract researchers' attention and has been applied to image deblurring and image super-resolution issues.How to improve the quality of the restored image to meet the practical application is the focus of current research.The thesis focuses on image deblurring and image super-resolution algorithms based on generative adversarial networks.The main work and innovations are as follows:(1)A new image deblurring algorithm based on generative adversarial network is proposed,which solves the problems of low quality and checkerboard effect of the generated image in the current deblurring algorithm based on generative adversarial network.The algorithm proposed in this paper uses the depth features extracted by VGG and DenseNet networks to calculate the perceptual loss,improves the network's edge detail expression ability,and eliminates the grid effect.The algorithm proposed in this paper adds sub-pixel convolutions as upsampling to the deep network which enhances the reuse of image features.Proposed algorithm also adds pixel loss to the loss function to improve the quality of generated images.The test results on the GOPRO dataset show that the algorithm proposed in this paper has improved subjective quality and objective PSNR and SSIM indicators compared to DeblurGAN,which currently has excellent performance.(2)A new image super-resolution algorithm based on generative adversarial networks is proposed,which solves the problems of current image generation with super-resolution algorithms based on generative adversarial networks that are not high-quality and insensitive to repetitive high-frequency details.In the proposed algorithm,densely connected hole convolutional network modules with different hole rates are added to enhance the network's ability to generate image features at different size levels;The channel attention mechanism is introduced in the network to adaptively select the generated image features,which improves the network's generated image quality.After experiments on the classic datasets of Set5,Set14,BSD100 and Urban100,the proposed algorithm has improved PSNR and SSIM indicators compared with ESRGAN.(3)An image restoration system based on generative adversarial network is designed and implemented.Based on the algorithm research,the paper designs and implements a prototype image restoration system.The system consists of an Android front end and a background system based on the LAMP architecture.The Android client interacts with the background system through the HTTP protocol,and users can easily restore the degraded image at any time through the Android phone.The design of the system reduces the computational pressure on the Android side,while ensuring the effect of image restoration.