结合结构重参数化方法与空间注意力机制的图像融合模型  

Image fusion model based on structure reparameterization method and spatial attention mechanism

作　　者：俞利新[1] 崔祺 车军[3] 许悦雷 张凡[2] 李帆[2] Yu Lixin;Cui Qi;Che Jun;Xu Yuelei;Zhang Fan;Li Fan(College of Air Traffic Control&Navigation,Air Force Engineering University,Xi’an 710051,China;Unmanned System Research Institute,Northwestern Polytechnical University,Xi’an 710072,China;National Key Laboratory of Science&Technology on Aircraft Control,AVIC Xi’an Flight Automatic Control Research Institute,Xi’an 710076,China)

机构地区：[1]空军工程大学空管领航学院,西安710051 [2]西北工业大学无人系统技术研究院,西安710072 [3]航空工业西安飞行自动控制研究所飞行器控制一体化技术国防科技重点实验室,西安710076

出　　处：《计算机应用研究》2022年第5期1573-1578,1600,共7页Application Research of Computers

基　　金：航空科学基金资助项目。

摘　　要：针对深度学习在红外与可见光图像融合时无法同时满足运算效率和融合效果的问题,提出了一种无监督端到端的红外与可见光图像融合模型。该模型的图像特征提取模块采用结构重参数化方法,有效提升了算法的运行效率;同时引入了注意力机制,减小了冗余信息对融合结果的干扰;损失函数基于结构相似度设计。对比实验结果表明,该模型保证了融合效果并提升了运行速度,相比于不使用结构重参数化的方法,运行速度提升了34%。Research on the problem that deep learning methods couldn’t meet the computing efficiency and fusion effect at the same time when infrared and visible images were fused,this paper proposed an unsupervised end-to-end infrared-visible image fusion model.The image feature extraction module adopted the method of structure reparameterization,which effectively improved the running speed of the algorithm.This model introduced attention mechanism to reduce redundancy in the source image.And it designed the loss function based on the structural similarity.The experimental results show that the model guarantees the fusion effect and improves the running speed.Compared with the method without using reparameterization of the structure,the running speed is increased by 34%.

关 键 词：结构重参数化 端到端 空间注意力机制 无监督学习 结构相似度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]