Research On Image Super-resolution Based Or Generative Adversarial Network

Image super-resolution reconstruction aims at the reconstructions of clear and high-resolution images based on the low-resolution images.With the rapid development and wide application of deep learning,many deep learning based image super-resolution algorithms have been proposed and obtained remarkable achievements.By analyzing the principles and ideas of several representative super-resolution algorithms,the problem of image super-resolution is studied from the two aspects of network structure and loss function to improve the performance of the algorithm in this paper.Firstly,on account of the network structure design,an image super-resolution algorithm based on multi-scale feature mapping network was proposed according to the idea of multi-scale mapping and the combination with the network structure of U-net.This algorithm improved the structure of U-net network to solve the problem of image super-resolution.Furthermore,a multi-scale feature mapping module was used to up-sample and non-linear mapping image features under multiple scales.The output features of multi-scale feature mapping module were fused with reconstruction feature by the net through concatenate operation.The underlying information was directly transmitted to the high level of the network to compensate the loss of image information caused by the down-sampling and up-sampling operations,and help the reconstructed image to recover more details information and improve the quality of the reconstructed image.The experimental results show that the performance of the algorithm has been significantly improved.Moreover,in order to study the effect of loss function on the quality of reconstructed image,a generative antagonism network was constructed in this paper for the image super-resolution combining with generating countermeasure network.The overall framework of the algorithm was composed of generator network and discriminator network.To control the influence of network structure on experimental results,the aforementioned multi-scale feature mapping network was used as generator network,and a discriminator network with fully convolutional structure was designed to better learn the local texture of images.The algorithm used a joint loss function that combined the MSE loss,the perceptual loss and adversarial loss to train the generator network.Relevant experimental results show that the proposed algorithm not only improves the perceptual quality of reconstructed images,but also eliminates the sharp noise in the smoothed areas.In summary,this paper studied the image super-resolution problem from the two aspects of network structure and loss function based on the related methods of deep learning.In addition,the higher perception quality and objective quantitative score of the proposed algorithm have been demonstrated through the comparison with other deep learning based image super-resolution algorithms.