Research On Image Super-resolution Reconstruction Of Conditional Generative Adversarial Network Based On Residual Learning

Image super-resolution reconstruction refers to enlarging low-resolution images to clear high-resolution images by algorithm.This technology can not only improve the visual effect of images and meet people's needs for high-definition image quality,but also be conducive to later image processing tasks.With the development of deep learning,a large number of superresolution reconstruction algorithms based on convolutional neural networks have appeared.Although the traditional reconstruction algorithm makes full use of related theoretical knowledge,the reconstructed image has poor results.The super-resolution reconstruction algorithm based on deep learning learns features from a large amount of data,which is beneficial for extracting high-frequency information,and its reconstruction effect exceeds that of traditional algorithms.This paper has made exploratory and innovative research on image super-resolution reconstruction based on conditional generative adversarial networks,and has achieved the following results:The image super-resolution method based on generative adversarial networks has a good perception effect,but the super-resolution image generated by its algorithm is very different from the ground truth,resulting in a low peak signal-to-noise ratio.Therefore,the image superresolution algorithm for conditional generative adversarial network based on residual learning of simplified residual network is proposed.First,the algorithm includes a generator network and a discriminator network.The generator network generates super-resolution images.The discriminator network is used to distinguish the super-resolution image and the ground truth.Secondly,the ground truth is added as a conditional input to the discriminator network,which can give a guide to the discriminator model,which can effectively improve the discrimination efficiency and optimize the training process.In addition,in the generator network,a method of deep residual learning is introduced to solve the problem of loss of detailed information as the depth increases.Finally,a total variation loss function is added to the perceptual loss function to continuously optimize the network and generate high-quality images.Compared with the SRGAN algorithm,the super-resolution image generated by the improved algorithm not only has no artifacts and mosaics,but also improves the PSNR and SSIM values by 2.5d B and 5.3%,respectively.Experiments show that our algorithm performs well on visual perception and evaluation indicators.In order to improve the performance of the generator network in the first work,fully utilize the context information of the network,fully learn the information of each convolutional layer,and improve the ability of feature learning,the image super-resolution algorithm of conditional generative adversarial network for residual learning based on dense residual network is proposed.A dense residual module is designed by combining dense connections and residual networks.In order to reduce the parameters of network training,group convolution is applied to the dense residual module,which makes the dense residual module lighter.In addition,the up-sampling module up-samples the shallow feature map through parameter sharing,so that the reconstructed image has rich high-frequency and low-frequency information without adding parameters.Experiments show that the reconstructed image is rich in details and has good visual effects.