Research And Improvement Of Single Image Super Resolution Algorithm Based On Generative Adversarial Network

The single image super resolution algorithm(SISR)refers to the technique of processing the low resolution image with super resolution and reconstructing the resolution image with magnification factor.In recent years,generative adversarial networks(GAN)have been introduced into the super resolution task of single image.SRGAN algorithm is a groundbreaking work in this class of algorithms.However,the reconstruction of images of this class of algorithms may produce artifacts,the details are fuzzy,and the perception of reference images is low.In order to improve the above problems,this paper proposes a specific improvement method for the SRGAN algorithm through the in-depth study of the algorithm,and obtains a new single image super-resolution algorithm based on the generative confrontation network.The closer the degree of perceptual similarity between the reconstructed image and the reference image is,the closer the perception effect of the two images is.Therefore,this paper mainly uses LPIPS index to measure the perceived similarity between the reconstructed image and the reference image.This paper improves the network structure of SRGAN algorithm generator to improve the quality of reconstructed images.The methods to improve the network structure include:removing the batch normalization layer,improving the output convolution layer,increasing the long jump connection,improving the residual structure of the generator network,and so on.The experimental results verified that,when the magnification factor was 4 times,the "3×5" model in this paper was used to reconstruct the images of the PIRM test set.The corresponding LPIPS index decreased from 0.1713 to 0.1479,which decreased by 13.66%,and the quality of the reconstructed images was improved.In this paper,the loss function of "3×5" model is further improved.In the confrontation training stage,the standard generative confrontation network framework is used,combined with the improved high-frequency feature loss function,increase the feature discrimination loss function for training,the "Ours" model of this paper is obtained.The experimental results verify that when the magnification factor is 4 times,the "Ours" model in this paper is used to reconstruct the pictures of the PIRM test set.The corresponding LPIPS index drops from 0.1479 to 0.1458,which further reduces 1.42%.Improved perceived similarity between the reconstructed image and the reference image.In this paper,the unsharpening masking method is improved by the idea of filtering out useless high frequency noise and strengthening low frequency features,and an improved unsharpening masking method is obtained for a single image super-resolution task.According to the "Ours" model in this paper,the LPIPS index of reconstructed pictures under PIRM test set can be reduced from 0.1458 to 0.1438 by using the unsharpening masking method improved three times.In this paper,specific application environments are proposed for the improved algorithm,including network interpolation strategy,low resolution face recognition and target detection.In order to have a good human-machine interaction,this paper has written the related software interface application.