Research And Design Of Super Resolution Image Reconstruction System Based On GAN

Image super-resolution reconstruction can promote the image quality and upgrade the image resolution,which gets most of attention in computer vision.In recent years,Deep Learning models,especially Generative Adversarial Network models,have achieved good results in the field of image super-resolution reconstruction.However,there are some problems,such as the missing of details,massive parameters,training difficulties,and so on.While the depth of networks is increasing,these models have more demands for hardware devices,which restricts their application in more conventional environments.The network models with recursive structure make full use of parameters,and are useful in model size reduction.With the changes in receptive fields,focusing on each patch of the image instead of the whole image is good for discriminating high frequency information.Based on the above thoughts,the optimization methods of generator and discriminator are proposed,which can improve the reconstruction quality and reduce the amount of network parameters.And the methods of feature layer selection for perceptual loss are compared.The specific work is as follows:(1)Trying to solve the lack of details and texture,as well as the large amount of model parameters and high hardware requirements,which are the problems of mainstream methods,a model which is named as Recursive Patch Generative Adversarial Network(RPGAN)is proposed based on the theory of recursive blocks and patches.The discriminator of the model limits attention to block areas of the image.The average authenticity probability of all parts is used to represent the authenticity probability of the whole image.In this way,the model can increase the details of reconstruction images;residual units of recursive structure are used to design the Generator,which increases the use of shallow features of the model.RPGAN reconstructs higher quality images with a smaller model.(2)In addition,the method of choosing feature layers which are used to calculate perceptual loss in the pre-trained vgg19 model is studied.Features before activation are used to replace those after activation to improve the perceptual loss of RPGAN,and further improve the reconstruction quality.(3)A super-resolution image reconstruction system based on GAN is designed with RPGAN proposed in the paper.The system can accomplish the task of super-resolution reconstruction of image using trained model automatically.It also allows users to set training parameters which are used to retrain the model interactively according to users' own needs and hardware environment.While the model is being trained,users can observe the training result in real time.In this way,the process of reconstruction can match users' requirement.Experiments were carried out separately on urban100,bsd100,set5 and set14.The data of experiments shows that RPGAN proposed in this paper reconstruct images of super-resolution with good qualities and more details.Network parameters of the model are reduced by 45.8%,and the time cost of training is reduced by about 12%.The super-resolution image system implemented by RPGAN can reconstruct high quality super-resolution images quickly.It can also achieve the goals of retraining models and reconstructing images with specified parameters given by users using small GPU memory and limited hardware devices.