Research On Image Super-Resolution Reconstruction Algorithm Based On Improved Generative Adversarial Network

In the era of rapid development of information technology,images have become the most direct way for people to understand things and obtain information,and people’s demand for image quality is increasing day by day.However,in the process of image dissemination,due to the influence of various environmental factors,the image resolution will inevitably be reduced.As a result,image super-resolution reconstruction technology has emerged.This technology refers to the reconstruction of low-resolution images with specific networks and algorithms to improve their resolution and obtain a specified highresolution image based on existing image devices.In recent years,there are many models based on deep learning for image super-resolution reconstruction tasks,among which generative adversarial networks have become a new and increasingly emerging approach to their outstanding performance in processing high-dimensional data distributions and feature learning.In this thesis,improvements are made to the super-resolution image reconstruction method based on generative adversarial networks,and the specific research work is as follows:(1)To address the problems of the difficult training process,smoothness of reconstructed image edges and distortion of details in traditional generative adversarial networks,a RRD-SRWGAN model is proposed based on the traditional SRGAN(Super Resolution Generative Adversarial Network).The model uses Wasserstein distance instead of Jensen-Shannon divergence,which is a measure of data distribution in generative adversarial networks,to optimize the training process of the network and solve the problem of training instability.This improves the structure of the generator and removes the Batch Normalization(BN)layer from the generator,which improves the feature capture capability of the network and eliminates the artifact problem.The model is experimentally shown to outperform SRGAN in terms of structural similarity and peak signal-to-noise ratio,and generates images with richer texture details.(2)The role of the attention mechanism is analyzed,and the Shuffle Attention(SA)mechanism is introduced to address the problem of high-frequency information loss in reconstructed images.Channel Shuffle operation is used to efficiently combine spatial attention and channel attention mechanism,and it is lightly processed and incorporated into the discriminator,further improving the structure of the discriminator and improving the feature extraction capability without significantly increasing the computational load of the network.Through specific experiments,it is shown that the model incorporating the attention mechanism in this thesis not only improves objective evaluation indexes compared with the original network model,but also improves the subjective visual perception of the reconstructed images to a certain extent,and achieves better reconstruction results.