Research On Image Super-resolution Reconstruction Algorithm Based On Improved Generative Confrontation Network

Super resolution image reconstruction(SRIR or SR)is one of the important branches of computer vision and image processing,with a wide range of application value.The methods based on deep learning are more efficient than those with respect to high-resolution image reconstruction,and the quality of reconstructed image.However,the reconstruction methods based on deep learning show problems of the instability of network model and the fuzziness of high-frequency information.Recently,the image super-resolution algorithm based on the Generative Adversarial Network has achieved good reconstruction results.In this paper,the generated countermeasure network is studied in depth,and two improved image super-resolution methods for the generated countermeasure network model are proposed.The main contents are as follows:(1)Firstly,this paper surveys the research background of image super-resolution reconstruction and the theoretical basis of traditional image super-resolution reconstruction technology,including interpolation based,reconstruction based and learning based methods,and the advantages and disadvantages of these methods are compared and analyzed.Secondly,this paper discusses some existing methods of image super-resolution reconstruction based on deep learning,with emphasis on the reconstruction methods based on the generated countermeasure network.Based on SRCGAN model,two experiments are carried out:?using set5 and set14 data sets as test sets,the quality of the improved model reconstruction image details is verified from subjective and objective perspectives;?using Minist data set to further verify the quality of SRCGAN model reconstruction image.The experimental results show that the improved model effectively improves the quality of image reconstruction.(2)This paper proposes an improved method based on the Generative Adversarial Network.? Regarding the randomness of image generation in SRGAN,the paper introduces the condition class and puts forward a model SRCGAN.The model solves the defect of random processing in the process of image generation,guides the process of network training,and improves the efficiency of algorithm and the quality of reconstructed image.?In addition,the convolution layer and residual block parameters of the network are adjusted to solve the problems such as color normalization and the spatial representation of the destroyed original image.? The idea of PatchGAN is introduced to the discrimination network,which reduces the operation parameters.Under the same computing resources,a deeper network is used to capture more detailed features,which improves the problem of image reconstruction blur in SRCGAN network.(3)In this paper,an image super-resolution reconstruction algorithm based on multi-scale Gan is proposed.In order to solve the problem of the poor effect of image detail restoration in SRGAN network reconstruction,this paper integrates the idea of Laplacian pyramid and completes the task of multi-scale image reconstruction by staged reconstruction.The discrimination network is integrated with PatchGAN to improve the restoration effect of image details and the quality of image reconstruction.Using set5,set14,bsd100 and urban100 data sets as test sets,the performance of the improved algorithm proposed in this paper is validated by experimental analysis from objective and subjective evaluation indexes.