Research On Image Super-Resolution Reconstruction Algorithm Based On GAN Network

Images are one of the important ways for humans to visually obtain information,and there is a direct relationship between the size of their resolution and the amount of information which they contain.The higher resolution of an image,the more information which it carries.But low-resolution images are prevalent in daily life and don’t meet the needs of some scenarios,for example,in the fields of remote sensing satellites,medical images,old photographs,ancient murals and other fields,where the images acquired are often unsatisfactory due to factors such as the imaging environment,hardware conditions and historical human destruction.Image super-resolution reconstruction solves this problem by reconstructing low-resolution images into high-resolution images through algorithm.This paper focuses on image super-resolution reconstruction algorithms based on generative adversarial network(GAN),the research contents of this thesis are as follows:(1)Ancient murals have suffered from varying degrees and types of lesions due to human actions throughout history and damage from the natural environment.There is little textural detail and lack of clarity in the murals,resulting in a shortage of ornamental and research value.To solve the above problems,a super-resolution method for murals based on residual fusion generative adversarial network is proposed,which can better reconstruct detail textures.Firstly,in order to improve the utilization of shallow image features,information distillation blocks are introduced to extract shallow features from images.The information distillation block enhances the output of the network through the enhancement unit and the compression unit.Secondly,the residual dense blocks with residual scaling and feature fusion are used to extract the deep features of images,and local feature fusion and global feature fusion are applied in the generative network to adaptively fuse the features of different levels together,so that the reconstructed image contains rich detailed information.Finally,when calculating the perceived loss,the pre-activation features are used to enhance the consistency between the color brightness of the reconstructed mural image and the original mural image,because when the number of network layers is large,the pre-activation image features are richer,which provides more guarantee for network learning.The experimental results show that the algorithm has a clear outline and rich texture details in the subjective visual of mural images,which effectively improves the appearance of artifacts and achieves good results in objective evaluation.(2)Feature extraction in most networks uses a single size convolution kernel,which has limitations in extracting image features with different perceptual fields.Moreover,in lowresolution image,there is less high-frequency information,and the high-frequency features reconstructed by super-resolution are not ideal.To solve the above problems,an image superresolution reconstruction algorithm based on multi-scale generative adversarial network is proposed.Firstly,the multi-scale residual blocks are used to obtain the feature information of images under different receptive fields,and in order to avoid the loss of detail texture of images caused by too small or too large receptive fields,feature fusion is introduced in the multi-scale residual block.Feature fusion integrates information under different scales to obtain local feature information.Then,the channel attention module is introduced in the multi-scale residual block to enhance the reconstruction of high-frequency information and adaptively adjust the processing of feature details,overcoming the deficiency of lacking highfrequency information in the extraction of image features by larger convolution kernel,enabling the screening of some key information in the image reconstruction process and fully acquiring the channel feature information of the image.Tested on publicly available datasets,the experimental results show that an improvement in the overall visual effect of the reconstructed images and an advantage in objective evaluation metrics.