Image Super-resolution Reconstruction Based On An Improved Generative Adversarial Network

Digital image processing is a cross-disciplinary subject,and image super-resolution reconstruction is an important research direction.In real life,low-resolution images do not meet the needs of real life,so it is important to get high-resolution images.Image super-resolution reconstruction has been widely developed in many fields,such as communication field,medical detection,remote sensing radar satellite imaging and video processing.The traditional image super-resolution reconstruction technology has high requirements on the equipment and the super-resolution image is smooth and the detail information is not very good.Deep learning has a good advantage in generating images,so this paper used generative adversarial network which is an important research direction of deep learning for image super-resolution reconstruction technology research,around this problem,the main work of this paper is as follows:(1)Analyzing and studying the basic model and training algorithm of the generative adversarial network.Although the network has a good advantage in generating images,there are some problems in the network,such as the disappearance of the gradient,the excessive freedom of the model,and the inability to generate discrete data distribution problem.Aiming at some improved methods against the network research proposed by scholars,WGAN solves the problem of gradient disappearance and the inability to generate discrete data distribution by changing the loss function.CGAN solves the problem that the model is too free by changing the network structure.This paper was based on these two networks to improve the b,asic generative adversarial network.(2)At the same time,the residual network and the symmetric convolutional neural network are introduced.The generation model uses the residual network and the symmetric convolution network.The discriminant model uses a multi-layer convolutional neural network.Not only the network structure is improved,but also the loss function is improved,the perceptual loss is used for the loss function,and the deep convolutional neural network VGG19 is used to obtain the perceptual loss function,so that the training effect of the network is optimized and improved.(3)The proposed image super-resolution reconstruction experiment was carried out to improve the effectiveness of the proposed method.The four data sets were tested in general,and the image super-resolution experiments were carried out for different amplification factors.The traditional methods were used to compare with the method of this paper,such as Bicubic,SRCNN,VDSR and DRCN and SRGAN.In the experiment,the PSNR value of the evaluation index on the 2x,3x,4x amplification factor increased by 1.125dB,2.175dB and 2.075dB,respectively,and the evaluation index SSIM also had good results.It can be concluded that the experimental results of improvd generative adversarial network reflected the problem that can be well applied to image super-resolution reconstruction.