欢迎访问《兵工学报》官方网站,今天是

兵工学报 ›› 2025, Vol. 46 ›› Issue (7): 240797-.doi: 10.12382/bgxb.2024.0797

• • 上一篇    下一篇

复杂场景伪装小目标机载偏振遥感检测方法

沈英1, 张硕1, 王舒1, 苏云2, 薛芳2, 黄峰1,*()   

  1. 1 福州大学 机械工程及自动化学院, 福建 福州 350108
    2 北京空间机电研究所, 北京 100094
  • 收稿日期:2024-09-04 上线日期:2025-08-12
  • 通讯作者:
  • 基金资助:
    国家自然科学基金项目(62005049); 福建省自然科学基金项目(2020J01451); 福建省教育厅中青年教师教育科研项目(JAT190003)

A Method for Detecting the Camouflaged Small Target in Complex Scene Using Airborne Polarization Remote Sensing

SHEN Ying1, ZHANG Shuo1, WANG Shu1, SU Yun2, XUE Fang2, HUANG Feng1,*()   

  1. 1 College of Mechanical Engineering and Automation, Fuzhou University, Fuzhou 350108, Fujian, China
    2 Beijing Institute of Space Mechanics and Electricity, Beijing 100094, China
  • Received:2024-09-04 Online:2025-08-12

摘要:

无人机遥感探测在军事侦察领域发挥着重要作用,偏振探测利用偏振光与物体相互作用产生的偏振变化来提高目标对比度。然而在复杂场景下,伪装小目标与背景特征差异较小且空间信息不足,存在检测困难的问题。为此提出一种偏振伪装小目标检测算法(Polarization Camouflaged Small Object Detection-YOLO,PCSOD-YOLO),设计了高效层注意力模块-坐标注意力特征提取模块和空间金字塔池化跨阶段局部通道-3D权重注意力感受野模块,捕获目标的偏振特征和语义信息,增强上下文信息理解能力;设计了动态小目标检测头,通过动态卷积增强对小目标特征提取能力的同时,利用不同尺度的特征信息,联合多通道特征信息输出小目标检测结果。构建伪装小目标偏振图像数据集(Polarization Image of Camouflaged Small Objects,PICSO)。在PICSO数据集上的实验表明,所提出的方法可以有效检测伪装小目标,mAP0.5达到92.4%,mAP0.5:0.95达到47.8%,检测速率达到60.6帧/s,满足实时性要求。

关键词: 无人机, 小目标检测, 深度学习, 偏振成像, 动态卷积

Abstract:

Unmanned aerial vehicle (UAV) remote sensing detection plays an important role in military reconnaissance,and the polarization detection is to utilize the polarization changes generated by the interaction between polarized light and target to improve the target contrast.However,in complex scenes,the small targets are less distinguishable from the background due to their similar features and the insufficient spatial information,resulting in difficulties in detection.To this end,a polarization camouflaged small object detection (PCSOD)-YOLO algorithm is proposed,and an efficient layer attention module-coordinated attention (ELAM-CA) and a spatial pyramid pooling cross stage partial channel-3D weights attention (SPPCSPC-3DWA) module are designed to capture the polarization features and semantic information of target,enhancing the ability to understand the contextual information.A dynamic small target detection head is designed to enhance the ability to extract the features of small targets through dynamic convolution,and the detected results of small target are outputted using the feature information from different scales and the multi-channel feature information.A polarization image of camouflaged small objects (PICSO) dataset is constructed for the camouflaged small target polarization images.Experiments on the PICSO dataset show that the proposed method can effectively detect the camouflaged small targets,with mAP0.5 and mAP0.5:0.95 reaching 92.4% and 47.8%,respectively.The detection rate reaches 60.6 frames per second,meeting the real-time requirements.

Key words: unmanned aerial vehicle, small target detection, deep learning, polarization imaging, dynamic convolution