Research On Image Super-Resolution Reconstruction Technology Based On Generative Adversarial Networks

As an image processing technology,Image Super-Resolution Reconstruction has been widely used in many fields.Generative Adversarial Networks as an important research hotspot of neural networks,has made outstanding progress in the application of Image Super-Resolution Reconstruction.It solves the problems of complex calculation and smooth reconstruction of image by traditional Super-Resolution Reconstruction algorithm,and the reconstructed image has good visual sense.This paper focuses on the research and discussion of Image Super-Resolution Reconstruction technology based on Generative Adversarial Networks.The main research contents are as follows:(1)In view of the difficulty in training the Super-Resolution Reconstruction of images based on Generative Adversarial Networks,the idea of spherical mapping is introduced to the discriminant,and the high-dimensional spatial distance of images is used to improve the stability of network training.Moreover,an additional feature discriminator is added to extract the high-frequency information of the image using the Visual Geometry Group(VGG)to improve the problem of lack of texture details in the reconstructed image.Firstly,the generated image was mapped to a high-dimensional spherical space through a discriminant network,and the training target was optimized based on the geometric moment distance between the spherical features.Then the reconstructed image and the real image were extracted from the VGG network,and the target was optimized to identify the source of the feature.Experimental results show that,compared with similar algorithms,the proposed algorithm improves the Peak Signal-to-Noise Ratio and Structural SIMilarity of reconstructed images,and improves the quality of reconstructed images.(2)Aiming at the problem that the non-differentiable image reconstruction index cannot be optimized by using gradient descent algorithm,the idea of sorting network is introduced into the Image Super-Resolution Reconstruction algorithm,and the evaluation of image quality by sorting network is used to improve the image evaluating index and improve the reconstructed image quality.The network consists of generator,discriminator and sorting network.First,the reconstructed image is generated by a generator,and then the reconstructed image is input into the sorting network to optimize the reconstructed image to have a better sorting index.Finally,the generated image and the original high-definition image are judged by a discriminant to be true and false.Experimental results show that the Natural image Quality Evaluator,Peak Signal-to-Noise Ratio and Structural SIMilarity of the proposed algorithm are improved,and the reconstructed image is consistent with human visual senses.(3)The reconstruction algorithm is optimized on the basis of the introduction of the sorting network.Since the introduction of the sorting network,the reconstructed image is consistent with human visual senses.However,the algorithm is based on the original Generative Adversarial Networks and the network training is not stable.In order to solve the above problems,this paper introduces penalty function and local penalty function respectively,and compares and selects the two models to improve the stability of the network.Experimental results show that,compared with similar reconstruction algorithms,the generation loss of the proposed algorithm can converge effectively,the reconstructed image has improved Peak Signal-to-Noise Ratio and Structural SIMilarity,and the image has rich texture details.