Image Super-resolution Reconstruction Based On Generative Adversarial Networks

In recent years,great progress has been made in image super-resolution reconstruction algorithms based on deep learning.However,the reconstructed image is too smooth and lacks authenticity.To this end,this dissertation focuses on how to restore the texture details of the reconstructed super-resolution image.The main work is as follows:First,this paper proposes an image super-resolution reconstruction model based on self-attention generative adversarial networks.First,the self-attention layer is introduced into the residual module of the generator,which can better use the global feature information to reconstruct the super-resolution image;Second,the model replaces the batch normalization layer in the generator with the instance normalization layer based on the deep network structure;Then,the model uses Charbonnier loss as the content loss to further evaluate the similarity between the reconstructed super-resolution image and the real super-resolution image;Finally,the model uses the feature values before activation of the VGG-19 network to calculate the perceptual loss,the features before activation can better represent the feature information of the image,so the texture of the reconstructed image and the original image can be better monitored,and ensure the authenticity of the reconstructed image.Second,this paper proposes a super-resolution reconstruction model for facial images based on generative adversarial networks.First,the model uses the bicubic difference to enlarge the low-resolution face image by a factor of 4 times,which is expand image pixels to target super-resolution image pixel size;Second,the model uses the encoder and decoder to design a new generator structure,and the encoder and decoder share features;Then,the model uses the intermediate features of the discriminator to construct the perceptual loss,instead of the perceptual loss based on the VGG16 classification network,a more robust image super-resolution perceptual loss can be obtained,and the facial details of the face can be better synthesized;Finally,the model uses the least squares loss to replace the traditional generation against the network loss,which optimizing GAN model training and improves the quality of images generated by the generator,the model constructs a general facial image super-resolution reconstruction model by combining perceptual loss,content loss and confrontation loss.The experimental results on Set5,Set14,BSD100,Vaild,CelebA and Helen datasets verify the effectiveness of the proposed image super-resolution reconstruction algorithm.There are 30 figures,6 tables,and 88 references in this dissertation.