| In the current era of big data,the contradiction between the increasingly huge image data transmission and storage requirements and the lag in the update speed of existing networks and devices has gradually become prominent.Traditional solutions mostly use low-resolution compression to reduce the amount of image data and thus save image storage and transmission resources,but processing redundant information during image compression will cause loss of original data,resulting in severe distortion of compressed images.In addition,Compression artifacts can also occur after image compression,causing the image to lose its original sharpness.This thesis studies high-quality image compression and reconstruction algorithms without adding additional hardware resources.The main research work is as follows:(1)Aiming at the problem that some data is lost due to the processing of redundant information in the original image compression process,which leads to serious distortion of the compressed image,this thesis proposes an image compression model DSAGAN based on detail-enhanced self-attention generative adversarial network.While reducing the amount of original image data during the compression process,the loss of original data is minimized.The model inputs high-quality images,realizes image compression through encoder and quantization,and then uses the generation module and decoder to reconstruct the image to improve the quality of the compressed image;at the same time,in order to focus on global features,a self-attention module is added to the encoder,Effectively build relationships between regions.The DSAGAN model is trained and tested on the RESIDE exterior city image dataset.The experimental results show that,compared with the mainstream image compression models,the model proposed in this thesis can reduce the amount of image data while reducing image distortion and better image compression.(2)Aiming at the problem that the compressed image has edge artifacts and blocking effects that affect the image clarity,this thesis proposes an end-to-end edge edge-enhanced generative adversarial network,E~2GAN,to achieve high-quality reconstruction of compressed images.The model combines different features of the image,converts the feature vector of the compressed image into the feature vector of the generated image,uses a convolutional neural network for feature extraction,and uses a 9-layer Resnet module to ensure that the compressed image features are preserved during the conversion process.The product recovers low-level features from the feature vector;then edge images are added to the model to reconstruct image edges more accurately,and a new edge loss function is defined to optimize edge details.The E~2GAN model is trained and tested on Celeb A and ACDC medical datasets.The experimental results show that compared with the mainstream image reconstruction models PSNR and SSIM,the experimental results of this model are better and the image reconstruction quality is higher.(3)Based on the image compression and reconstruction algorithm studied in this thesis,an online image space management system for cloud storage is designed and built,which realizes the image compression and reconstruction with less distortion while compressing the amount of image data,and provides users with It provides convenient and quick image management and high-definition download functions,improves the efficiency of mobile users in storing,processing and transmitting images,and saves hardware and network resources.To sum up,this thesis constructs an image compression model DSAGAN based on detail-enhanced self-attention generative adversarial networks and a GAN-based end-to-end edge-enhanced image reconstruction model E~2GAN,and implements personal cloud storage based on the two models Image space management system.The actual application test results of the system show that the DSAGAN model can effectively compress the amount of image data on the premise of ensuring the image quality with less distortion,and the E~2GAN model can reconstruct JPEG compressed images into high-quality images. |
