空基下视多角度红外目标识别

Air-based downward-looking multi-angle infrared target recognition

为实现空基下视红外目标的快速高精度识别,提出了一种单阶段的空基下视多角度红外目标识别算法。首先使用Darknet-53结合SPP模块对红外目标进行特征提取,使局部特征与全局特征融合,提高特征图表达能力,最后借鉴RetinaNet中的Focal loss锁定目标的检测框,同时得出目标类型及检测精度。针对现有数据集多为平视,且视角单一的缺陷,使用复合翼无人机分别从不同高度和角度采集红外图像,构建多尺度下视红外目标数据集,在PyTorch架构上实现并进行性能验证实验,所提算法对分辨率为640×512的下视红外图像中目标识别的mAP达到91.74%,识别速度为33 f/s,满足空基平台前端的在线识别需求,且在公开红外船舶数据集上也具有较好的识别结果。实验表明该算法在保证精度的基础上满足实时性的要求,为后续用于复合翼无人机上的多尺度目标实时识别提供了理论技术。

In order to realize the rapid and high-precision recognition of air-based downward-looking infrared targets,a single-stage space-based down-looking multi-angle infrared target recognition algorithm is proposed.Firstly,use Darknet-53 combined with SPP module to perform feature extraction on infrared targets,to fuse local features and global features to improve the expression ability of feature maps,and finally use Focal loss in RetinaNet to lock the detection box of the target,and at the same time obtain the target type and detection accuracy.Aiming at the defect that the existing data sets are mostly head-ups and single viewing angles,a composite-wing drone was used to collect infrared images from different heights and angles,and a multi-scale down-view infrared target data set was constructed,which was implemented and verified on the PyTorch architecture.The proposed algorithm achieves 91.74%of the mAP of the downward-looking infrared target,the recognition speed is 33 FPS,and it also has a good recognition result on the public infrared ship data set.The experiments show that the algorithm meets the real-time requirements on the basis of ensuring accuracy,and provides theoretical technology for subsequent real-time recognition of multi-scale targets on compound-wing UAVs.