Research On Super Resolution Image Reconstruction Method Based On Generative Adversarial Networks

With the advent of the digital age,images are widely used in various fields as the main medium for information dissemination.In reality,due to the limitation of imaging equipment and natural factors,the resolution of the acquired images is low and cannot meet the actual needs of people,so the demand for obtaining high-resolution images has become increasingly urgent.Image super-resolution reconstruction,as an important research direction of image processing,has great application value in military reconnaissance,urban planning,environmental monitoring,etc.It is a series of processing to reconstruct single-frame or multi-frame low-resolution pictures into Detail textures are sharper and richer in high-resolution images.Traditional super-resolution image reconstruction usually uses hardware to improve the resolution of the image,which requires high cost and faces huge technical challenges.Aiming at the above problems,Convolutional neural networks and generative adversarial networks are focuses on in deep learning,and studies the super-resolution reconstruction of remote sensing images based on improved conditional adversarial networks.The main work and innovations of this article are as follows:(1)Based on reading a lot of related literatures,the development and status of image super-resolution image reconstruction at home and abroad are briefly summarized and the significance of the research on image super-resolution reconstruction.In addition,the basic principles and mechanisms of deep learning convolutional neural networks and generative adversarial networks are introduced in detail,and their advantages and disadvantages are analyzed.Aiming at the improved models CGAN and WGAN proposed by scholars for generating adversarial networks,we will improve them based on these two network models.(2)An improved model for super-resolution reconstruction of remote sensing image based on conditional generation of countermeasure network is proposed,aiming at the instability,slow convergence speed,gradient disappearance or explosion when generating countermeasures network training.In order to accelerate the convergence speed of the model,the objective function of content loss and counter loss is used in the generator network.In order to improve the stability of network training,gradient penalty function is introduced to limit the discriminator gradient.The experimental results show that compared with SRCNN,FSRCNN and SRGAN,the improved model has significantly improved subjective visual effect and objective evaluation indexes.(3)We have improved the combination of SRGAN and WGAN needs to satisfy the Lipschitz continuity,the weights can only be compressed into a range to enforce the Lipschitz continuity.However,this will result in a great waste of the deep neural network fitting ability.In addition,the forced shearing of weights will easily lead to the disappearance or explosion of gradients.Therefore,by introducing Wasserstein divergence into SRGAN and maximizing it,the optimal scalar function T can be obtained.In this way,the Wassertein distance can be directly obtained by removing L constraints,and the objective function of generating network can be obtained by minimizing Wassertein distance,and finally the quality of reconstructed image can be improved.The experimental results show that method can generate high-resolution face images and is superior to DRCN,FSRCNN,SRGAN_WGAN,VDSR and DRRN models in both subjective vision and objective evaluation indexes.