Research And Application Of Image Super-resolution Reconstruction Based On Generative Adversarial Network

In the process of image acquisition,it may be affected by many factors,such as the weather,the device with low pixel,the change of the background motion and so on,resulting in poor image quality and low resolution,which could not meet the requirements of subsequent image processing and analysis.Image Super-Resolution reconstruction is the process of reconstructing a corresponding high-resolution image from a single picture or a group of video sequences with low resolution,so as to gain higher resolution images for the fields of computer vision and image processing,such as Face Recognition,Video Surveillance,Unmanned Driving,etc.In recent years,with the rapid increase of computing power,Deep Learning has been applied to computer vision and other fields,and has achieved outstanding results.Inspired by this,the field of image Super-Resolution reconstruction(SR)hsa also adopted the idea of deep learning.Compared with traditional algorithms,the effect of the method has been greatly improved,so the method has became a mainstream technology in Super-Resolution.The images reconstructed by common deep learning networks lack reality and high-frequency details.Generative Adversarial network can overcome these shortcomings.Therefore,This paper mainly research the application of the Generative Adversarial Network(a kind of deep learning networks)in image super-resolution reconstruction.The details of the research work are as follows:1)Review the basic theories of super-resolution reconstruction and Generative Adversarial Network,and summarize their development status and research significance.Firstly,the super-resolution reconstruction algorithm is divided into three categories to present,then the super-resolution reconstruction based on deep learning is divided into two categories to describe,and then the principle of Generative Adversarial Network is elaborated..Finally,the technology of image super-resolution based on generative adversarial network is introduced.This will consolidate the theoretical basis for the follow-up research.2)An image super-resolution reconstruction scheme based on clique network is proposed: the clique block is improved and applied to image super-resolution reconstruction,the pooling layer and batch normalization layer of the clique block are removed;and in order to adapt to the image super-resolution task,the appropriate loss function and optimization method are adopted.After setting appropriate parameters,the training set is trained to obtain a model,then the corresponding high-resolution images can be reconstructed from the low-resolution images.Finally,the results of this network are compared with several existing networks by using the evaluation standard Peak Signal-to-Noise Ratio and Structural Similarity,it proves that our method is better.It is helpful to pave the way for image super-resolution based on Generative Adversarial Network.3)An image Super-Resolution reconstruction technique based on Iterative Back Projection-Generative Adversarial Network is proposed: It consists of a generator network and a discriminator network.In the generator network,Iterative Back Projection is combined with the residual network to apply to image super-resolution reconstruction.The up-sampling layers adopt sub-pixel convolution layer and the down-sampling layers adopt maximum pooling layer,what's more,the batch normalization layer of residual block;the loss function consists of a Huber loss function,a contextual loss function,and an adversarial loss function.A comparative experiment with SRGAN(Super-Resolution Using a Generative Adversarial Network)is implemented,our method can achieve a better reconstruction effect than SRGAN.In this paper,the image super-resolution reconstruction based on circle network is researched firstly,then the image super-resolution reconstruction based on Generative Adversarial Network with iterative back-projection is studied.Finally,the experiment shows that method can gain realistic high-resolution images with high-frequency details.