士兵目标的少样本深度学习检测方法

针对敌士兵数据集样本较少的问题，提出一种基于YOLOv3的少样本深度学习目标检测方法.利用数据增广提高少样本目标检测模型的鲁棒性，改进网络结构将浅层网络特征图跨层连接至深层网络，采用k-means聚类获取适合士兵目标特性的锚点框，利用预训练提高模型训练收敛速度.实验结果表明，本文方法对少样本敌士兵目标检测成功率mAP达到85.6%、检测精度IOU达到82.18%，且对小型和遮挡目标检测效果较好；部署在NVIDIA TITAN V GPU计算机和NVIDIA Xavier嵌入式计算平台上的检测速度分别达到54.6和26.8 fps，实时性好. 

A Deep Learning Detection Method for Soldier Target Based on Few Samples

A deep learning method detection for target with few samples based on YOLOv3 was proposed to solve the problem of small enemy soldiers’ datasets. Data augmentation was used to improve the robustness of the small-sample target detection model, and improve the network structure by connecting the shallow network feature map to the deep network across layers. k-means clustering was used to obtain anchor boxes suitable for soldier target characteristics, and pre-training was used to improve the convergence speed of model training. The results show that the method in this paper has a success rate (mAP) of 85.6% for target detection of enemy soldiers with small enemy soldiers’ datasets, a detection accuracy (IOU) of 82.18%, and a good detection effect for small and occluded targets. The detection speed deployed on NVIDIA TITAN V GPU computer and NVIDIA Xavier reaches 54.6 and 26.8 fps, which means a good real-time performance.