复杂背景下的红外运动目标语义分割算法研究  

Research on Semantic Segmentation Algorithm for Infrared Moving Targets in Complex Backgrounds

作　　者：吉浩宇 孟卫华[1,2] 张新朝 段静菲[1,2] 张蒙 Ji Haoyu;Meng Weihua;Zhang Xinchao;Duan Jingfei;Zhang Meng(China Airborne Missile Academy,Luoyang 471009,China;National Key Laboratory of Air-based Information Perception and Fusion,Luoyang 471009,China)

机构地区：[1]中国空空导弹研究院,河南洛阳471009 [2]空基信息感知与融合全国重点实验室,河南洛阳471009

出　　处：《航空兵器》2025年第2期80-86,共7页Aero Weaponry

基　　金：国家自然科学基金项目(02564897)。

摘　　要：红外弱小目标的语义分割对于细节纹理特征的依赖更强,较深的网络结构对红外目标的语义分割不适用,难以将弱小目标从复杂背景中准确分割出来。本文针对复杂背景下红外运动目标的语义分割任务需求,在公开目标检测跟踪数据集的基础上,标注构建了红外图像语义分割数据集,基于STDC-Seg模型针对红外图像特点进行了优化,提出了一种红外目标语义分割算法STDC-Infrared。重新设计网络下采样结构,增加空间注意力模块和多尺度自适应融合模块。实验结果表明,本算法相比STDC-Seg,在红外图像语义分割数据集上平均交并比和平均像素精度分别提升了12.47%和12.55%,特别是空中飞机目标的交并比和像素精度分别提升了31.53%和35.82%,有效提升了复杂背景红外运动目标场景下语义分割准确性。The semantic segmentation of infrared weak targets relies more on detailed texture features,and deeper network architecture are not suitable for semantic segmentation of infrared targets,making it difficult to accurately segment weak targets from complex backgrounds.This article focuses on the semantic segmentation task requirements of infrared moving targets in complex backgrounds.Based on the publicly available target detection and tracking dataset,an infrared image semantic segmentation dataset is annotated and constructed.The STDC-Seg model is used to optimize the characteristics of infrared images,and an infrared target semantic segmentation algorithm STDC-Infrared is proposed.Redesign the network downsampling structure,adding spatial attention module and multi-scale adaptive fusion module.The experimental results show that compared with STDC-Seg,this algorithm has improved the average intersection to union ratio and average pixel accuracy by 12.47%and 12.55%,respectively,on the infrared image semantic segmentation dataset.In particular,the intersection to union ratio and pixel accuracy of infrared aircraft targets have improved by 31.53%and 35.82%,respectively,effectively improving the performance of semantic segmentation algorithms in complex background infrared moving target scenes.

关 键 词：语义分割 红外运动目标 复杂背景 STDC-Seg 多尺度自适应融合 

分 类 号：TJ760[兵器科学与技术—武器系统与运用工程] TN219[电子电信—物理电子学]