针对红外弱小目标的多尺度级联融合分割网络  

Multi-scale cascaded fusion network for infrared small target segmentation

作　　者：杨馨雨 杨晓梅[1] 方璇 YANG Xin-yu;YANG Xiao-mei;FANG Xuan(College of Electrical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电气工程学院,四川成都610065

出　　处：《激光与红外》2025年第3期436-443,共8页Laser & Infrared

摘　　要：红外弱小目标低信噪比、模糊形态等特点使得复杂背景下的针对此类目标的检测任务研究仍存在一定挑战。为弥补现有方法常见的目标信息丢失、与背景相互混淆等问题,提出了一种创新性的多尺度级联融合目标分割网络(MSCFNet)。具体而言,MSCFNet通过不同尺度特征间的多次信息交互,最大限度保留小目标。同时,设计了一种特征增强模块,从全局和局部两个层面有效提取并整合信息,增强目标表征能力并抑制背景杂波。实验结果证明,MSCFNet在多种复杂背景环境中能够有效地分割目标,并在两个公开红外弱小目标数据集上展现出更为优秀的分割性能。The low signal-to-noise ratio and fuzzy morphology of infrared(IR)small targets pose certain challenges in the research of such target segmentation tasks in complex backgrounds.To better separate small targets from clutter backgrounds,an innovative multi-scale cascaded fusion network(MSCFNet)is proposed.Specifically,MSCFNet preserves and utilizes small target information to the maximum extent through the multiple interaction between multi-scale features.At the same time,a feature enhancement module is designed to effectively extract and integrate target information from global semantic and local context,improving the discriminability of targets and complex backgrounds.The experimental results prove that MSCFNet can effectively segment IR small targets in various complex environments and exhibits better performance on two publicly available IR small target segmentation datasets.

关 键 词：红外弱小目标分割 红外图像 深度学习 特征增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TN219[自动化与计算机技术—计算机科学与技术]