现代电子技术2025，Vol.48Issue(11)：42-50,9.DOI:10.16652/j.issn.1004-373x.2025.11.007
基于CNN-Transformer融合网络的红外图像超分辨率方法
Infrared image super-resolution method based on CNN-Transformer fusion network
摘要
Abstract
For the problem of infrared image super-resolution,both CNN-based methods and Transformer-based methods have achieved competitive results in recent years.To a certain extent,CNN and Transformer are complementary in feature representation.In order to combine the advantages of the both,a CNN-Transformer feature fusion network(CTF-Net)containing two branches is proposed.The CTF-Net fuses the local features of CNN and the global information of Transformer effectively.Specifically,the CNN branch strives to take the residual-in-residual dense block(RRDB)as the feature extraction backbone,so as to extract enhanced local features.The Transformer branch strives to combine self-attention and channel attention,and capture complete global information in both spatial and channel dimensions.In addition,in view of the relatively insufficient high-frequency information of infrared images,effective contrast loss is introduced.By moving away from blurred negative samples and closer to sharpened positive samples,the utilization and recovery of high-frequency features are enhanced while improving the lower limit of super-resolution results.A large number of experiments show that the proposed CTF-Net achieves optimal performance indicators,and the edges and textures of the super-resolution images generated by the CTF-Net are clearer.To sum up,this method further promotes the high-quality application of infrared imaging technology.关键词
超分辨率/红外图像/CNN/Transformer/特征融合/对比损失Key words
super-resolution/infrared image/CNN/Transformer/feature fusion/contrast loss分类
信息技术与安全科学引用本文复制引用
杨海航,万显荣,周文洪,吴津..基于CNN-Transformer融合网络的红外图像超分辨率方法[J].现代电子技术,2025,48(11):42-50,9.基金项目
国家自然科学基金资助项目(61931015) （61931015）
