Study On Super-Resolution Reconstruction Algorithms Based On Generative Adversarial Networks

Super-resolution reconstruction is a technique for recovering corresponding high-resolution images from low-resolution images.The process of digital image acquisition and processing will reduce the resolution of the image and affect subsequent applications.How to convert a lower resolution image to a higher resolution image is the focus and difficulty in the field of image processing.The super-resolution reconstruction algorithm based on Generative Adversarial network is studied and analyzed,and three improved super-resolution reconstruction algorithms based on Generative Adversarial networks are given.The main work of this paper is as follows:A super-resolution reconstruction algorithm based on Gaussian coding feedback for Generative is presented.Aiming at the problem that the discriminative model in SRGAN model can not provide sufficient information for generating network to guide the Generative Model training,a Gaussian coding combined with leak network reconstruction method is proposed.The feature map of the discriminative model is encoded by Gaussian coding and then used by leak network.Information is transmitted to the generation network to guide it to training.Aiming at the problem that the single content loss can not effectively measure the reconstruction effect,a symmetric content loss that satisfies the cycle consistency is proposed.A super-resolution reconstruction algorithm based on feature fusion for Generative Adversarial network is presented.Aiming at the problem of generative model poor learning ability in SRGAN model and only using data-driven training,a reconstruction algorithm based on feature fusion is proposed.The recursive residual network is combined with prior knowledge to extract and fusion multiple features.The dense connection network makes full use of the feature information to highlight the edge information of the reconstructed image.The input of discriminant network for the SRGAN model is a whole image,which brings about a problem of high computational complexity.The residual discriminant method is proposed.The two difference maps are used as the input image of the discriminant network,the two discriminating methods are used for training.A super-resolution reconstruction algorithm based on the attention mechanism to Generative Adversarial networks is presented.Aiming at the problem that SRGAN extracts the local information of the input image and ignores the global information,a method of combining the spatial and channel attention mechanism on the generation network and the discriminant network is proposed.Based on the attention mechanism of the human eye,the image reconstruction effect is better and more in line with the human eyes observation mode.The above three super-resolution reconstruction algorithms based on generative adversarial network are combined to obtain an effective super-resolution reconstruction algorithm based on the generative adversarial network.Using the test data to verify the performance of the algorithm,the subjective index is significantly better than the original algorithm,the objective index PSNR is increased by 0.8db,and SSIM is increased by 3.3%,Finally,the CelebA face dataset is used to test the algorithms.Experiments show that the actual use of the algorithm is good and has certain feasibility and practical value.