Research And Practice Of Deep Learning Applied To Super-Resolution Reconstruction

With the rapid development of intelligent devices,pictures have become an indispensable part of life and production.It has become a big demand in various fields to convert old photos and low-resolution images into high-resolution images.Therefore,superresolution technology has become active and appeared in people's field of vision.Super resolution reconstruction is the reconstruction of low-resolution images into high-resolution images with more details.The super resolution reconstruction technology can be divided into two major categories: hardware reconstruction and software reconstruction.Due to many limitations in hardware,the method of software reconstruction has become a hot topic in current research and has been applied in many important fields,which has great practical value.Through the understanding of the development course of super-resolution reconstruction,as well as to the current development situation of investigation and study,detailed introduces the related theoretical basis of super-resolution reconstruction technology.On this basis,in-depth study of current popular super-resolution reconstruction technology based on the deep learning,learning the basic principle and structure of the generated adversarial network,and to generate adversarial network based super-resolution reconstruction technique(SRGAN)optimization improvement,a new improved algorithm has been proposed.The improved algorithm introduces Wasserstein distance into SRGAN network,and explains how to apply Wasserstein distance to SRGAN in detail,and proves the effectiveness of the improved algorithm through experiments.In the original SRGAN,the training instability of the SRGAN network is caused by some problems of the generated adversarial network itself,so the training degree of the generator and discriminator should be carefully balanced to prevent the serious gradient loss of the generator caused by overtraining of discriminator.The introduction of Wasserstein distance solves the problem that the original network training is unstable,the training situation cannot be monitored,and the number of training times needs to be carefully balanced.Using RMSProp optimization algorithm instead of the original in the network optimization algorithm,greatly improve the convergence speed of the network.