计算机技术与发展2024，Vol.34Issue(7)：31-39,9.DOI:10.20165/j.cnki.ISSN1673-629X.2024.0088
多尺度注意力特征融合的单图像超分辨率研究
Research on Single Image Super-resolution Based on Multi-scale Attention Feature Fusion
摘要
Abstract
High resolution means that the image has a high pixel density,which can provide more details,which often play a key role in the application.Image super-resolution based on generative adversarial networks has attracted more and more attention in recent years due to its potential to generate rich details.Aiming at the problem that the existing network model ignores the learning of essential texture features from features and the limited receptive field,based on Real-ESRGAN and multi-scale attention feature fusion,the network is op-timized,and the residual-in-residual dense block is replaced by a large kernel decomposition and multi-scale learning.The method of combining the module with the dual branch structure of the global learning and down-sampling module proposes a single image super-resolution reconstruction algorithm based on multi-scale attention fusion,which enhances the interaction between each local and global token pair to form a richer and more informative representation.Super-resolution reconstruction experiments of 2,3,4 times were carried out on the data set.The reconstruction results were evaluated by peak signal-to-noise ratio(PSNR)and structural similarity(SSIM),and compared with SRCNN,SRGAN,EDSR,RDN,RCAN,HAN,ENLCA,MAN and Real-ESRGAN methods.The results show that the proposed algorithm is better than other models,and has better visual effect.关键词
生成对抗网络/图像超分辨率/多尺度注意力特征融合/大核分解/全局学习与下采样/令牌Key words
generative adversarial network/image super-resolution/multi-scale attention feature fusion/large kernel decomposition/global learning and down-sampling/token分类
信息技术与安全科学引用本文复制引用
沈学利,翟宇琦,关刘美,苏婷..多尺度注意力特征融合的单图像超分辨率研究[J].计算机技术与发展,2024,34(7):31-39,9.基金项目
国家自然科学基金资助项目(62173171) （62173171）
