基于可见光-红外特征级融合的低光照下伪装目标智能检测技术  

Intelligent detection of camouflage object based on visible-infrared feature-level fusion in low-light conditions

作　　者：公金成 孙殿星 彭锐晖 徐乐 张一泓 GONG Jincheng;SUN Dianxing;PENG Ruihui;XU Le;ZHANG Yihong(Qingdao Innovation and Development Base,Harbin Engineering University,Qingdao 266000,China;Institute of Information Fusion,Naval Aeronautical University,Yantai 264001,China)

机构地区：[1]哈尔滨工程大学青岛创新发展基地,山东青岛266000 [2]海军航空大学信息融合研究所,山东烟台264001

出　　处：《指挥控制与仿真》2025年第2期40-49,共10页Command Control & Simulation

基　　金：国防科技重点实验室基金(2023-JCJQ-LB-016)。

摘　　要：低光照环境下的伪装目标检测是揭伪领域的难题之一,尤其随着伪装技术的不断发展,目标与环境背景高度融合,若此时的光照条件较差,往往会导致常规单模态目标检测算法性能退化。针对该问题,构建了一种以目标检测任务为引导的特征级融合网络。首先,设计了一种残差密集连接,实现多个维度信息提取和堆叠,提升目标在原始信息中的显著程度,获得伪装目标融合特征;然后,将融合特征送入YOLOv7网络进行伪装目标检测,通过损失函数优化、空间-通道注意力机制综合,有效提升了低光照下伪装目标检测效果。另外,构建了一个低光照环境下的光学-红外伪装目标数据集,对所提方法进行实测数据验证,在该数据集上的mAP@0.5为87.38%,精确率P为85.45%,表明该算法在低光照条件下对伪装目标具有检测优势。Camouflaged targets detection in low-light environments is one of the challenges in the field of deception detection.Especially with the continuous advancement of camouflaged technology,targets are highly integrated with their environmental background.Poor lighting conditions can often lead to performance degradation in conventional single-modal detection algorithms.To address this issue,this paper proposes a feature-level fusion network guided by the object detection task.First,this paper designs a residual dense connection to extract and stack information from multiple dimensions,enhancing the prominence of the target within the original information to obtain fused features of camouflaged targets.Then,the fused features are fed into the YOLOv7 network for camouflaged target detection.By optimizing the loss function and integrating spatial-channel attention mechanisms,the detection performance of camouflaged targets under low-light conditions is effectively improved.Additionally,this paper constructs an optical-infrared camouflaged target dataset for low-light environments to validate the proposed method with empirical data.The dataset shows an mAP@0.5 of 87.38%and a precision(P)of 85.45%,indicating that the proposed algorithm has a detection advantage for camouflaged targets under low-light conditions.

关 键 词：伪装目标检测 特征级融合 损失函数 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]