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 remote sensing detection plays an important role in military reconnaissance, and polarization detection utilizes the polarization changes generated by the interaction between polarized light and objects to improve target contrast. However, in complex scenes, the difference between disguised small targets and background features is small and the spatial information is insufficient, which poses a problem of difficult detection. To this end, this article proposes a polarization camouflage small object detection algorithm PCSOD-YOLO (Polarization Camouflaged Small Object Detection YOLO), and designs ELAM-CA (Efficient Layer Attention Module Coordinated Attention) and SPPCSPC-3DWA (Spatial Pyramid Pooling Cross Stage Partial Channel-3D Weights Attention) modules to capture the polarization features and semantic information of the target, enhancing the ability to understand contextual information; A dynamic small target detection head was designed to enhance the feature extraction ability of small targets through dynamic convolution, while utilizing feature information from different scales and combining multi-channel feature information to output small target detection results. Construct the Polarization Image of Camouflaged Small Objects (PICSO) dataset for camouflage small target polarization images. Experiments on the PICSO dataset show that the proposed method can effectively detect disguised small targets, with mAP0.5 reaching 92.4% and mAP0.5:0.95 reaching 47.8%. The detection rate reaches 60.6 frames per second, meeting real-time requirements.

Key words: UAV, small object detection, deep learning, polarization imaging, dynamic convolution