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

With the continuous development of in-depth learning technology,great changes have taken place in the field of computer vision.In computer vision,there is a technology called image super-resolution,which can transform low-resolution image into high-resolution image,so as to improve the image quality.Image super-resolution technology plays an important role in human life,especially in satellite remote sensing,medical image,video surveillance and other aspects.However,traditional image super-resolution methods have some problems,such as limited image enhancement effect and unsatisfactory information processing for edge details.Researchers have found that image super-resolution can be achieved by using the generation antagonism network in depth learning.Therefore,the research of image super-resolution technology based on Generative Adversarial Network has important application value.The image super-resolution algorithm based on Generative Adversarial Network consists of two parts: the generative model and the discriminant model,in which the generative model plays a decisive role in the result of image super-resolution.The structure of the generated model is usually composed of Deep Residual Network.Therefore,the image super-resolution algorithm based on Deep Residual Network is studied and improved in this paper.Firstly,the method of image up-sampling is selected.Because the traditional method of image up-sampling is far lower than the method of image up-sampling based on depth learning,this paper makes an experimental comparison of the method of image up-sampling based on depth learning.Finally,the sub-pixel convolution layer is selected to realize the image in the image super-resolution algorithm based on Deep Residual Network.Up sampling function.Then,the factors affecting the image super-resolution algorithm based on Deep Residual Network are analyzed and experimented,and an improved image super-resolution algorithm model based on Deep Residual Network is obtained.Specific improvements include: eliminating the batch normalization layer in standard residual blocks;replacing L2 norm loss function with L1 norm loss function;increasing the number of residual blocks and adding residual scaling operation in the execution process of residual blocks.The improved image super-resolution model based on Deep Residual Network has good effect,which lays a foundation for the subsequent research and improvement of image super-resolution based on Generative Adversarial Network.In the process of research and improvement of image super-resolution algorithm based on Generative Adversarial Network,the improved image super-resolution model structure based on Deep Residual Network is adopted in the design of generative model.In the design of discriminant model,this paper draws lessons from the design idea of VGG network and uses several 3x3 convolution layers to form the main structure of discriminant model.The loss function of the generated model can not be directly used in the image super-resolution model based on Deep Residual Network because of the idea of Antagonistic Game in Generative Adversarial Network.Therefore,a mixed loss function based on perceptual content loss,L1 norm loss and antagonistic loss is proposed as the loss function of the generated model.Compared with the traditional bi-cubic interpolation method and the original image super-resolution method based on Generative Adversarial Network,the improved algorithm has the highest image quality.Compared with the original image super-resolution algorithm based on Generative Adversarial Network,the improved algorithm has different degrees of improvement in peak signal-to-noise ratio and structural similarity.