偏振可以提高无人机的自主侦察能力，但易受到探测角度和目标材质的影响，从而降低偏振检测的鲁棒性。为此，提出一种基于偏振图像的低空伪装目标实时检测算法YOLO-P，采用融合多偏振方向信息的编码图像作为输入，应用三维卷积模块提取不同偏振方向图像之间的联系特征;引入特征增强模块对多层次特征进行进一步增强;采用跨层级特征聚合网络，充分利用不同尺度的特征信息，完成特征的有效聚合，最终联合多通道特征信息输出检测结果。构建包含10类目标的低空伪装目标偏振图像数据集PICO(Polarization Image of Camouflaged Objects)。在PICO数据集上的实验结果表明，新方法可以有效检测伪装目标，mAP0。5:0。95达到52。0％，mAP0。5达到91。5％，检测速率达到55。0帧/s，满足实时性要求。
Real-time Detection of Low-altitude Camouflaged Targets Based on Polarization Encoded Images
Polarization can improve the autonomous reconnaissance capability of unmanned aerial vehicle, but it is easily interfered by the variation of detection angle and target materials, which affects the robustness of polarization detection. In this paper, a real-time low-altitude camouflaged target detection algorithm of YOLO-Polarization based on polarized images is proposed. The coded image fused with multi-polarization direction information is used as input, the 3D convolution module is applied to extract the connection features from the different polarization direction images, and a feature enhancement module ( FEM) is introduced to further enhance the multi-level features. In addition, the cross-level feature aggregation network is adopted to make full use of the feature information of different scales to complete the effective aggregation of features, and finally combined with multi-channel feature information output detection results. A dataset consisting of polarized images of low-altitude camouflaged targets ( PICO) which include 10 types of targets is constructed. The experimental results based on PICO dataset show that the proposed method can effectively detect the camouflaged targets, with mAP0. 5:0. 95 up to 52. 0％ and mAP0. 5 up to 91. 5％. The detection rate achieves 55. 0 frames/s, which meets the requirement of real-time detection.
万方数据
维普
