Few-shot warhead fragment group object detection based on feature reassembly and attention
和萌 1武江鹏 1梁超 1胡鹏宇 1任媛 1贺翾 1刘强辉1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 西安现代控制技术研究所,陕西西安 710065
折叠
摘要
战斗部破片群运动参数对弹药毁伤威力评估具有重要的意义.针对破片尺寸较小、背景信息复杂以及破片数据样本少导致的破片检测精度较低的问题,本文提出一种YOLOv5-FD的战斗部破片群目标检测方法.首先,在网络输出端增加微小目标检测层,将原始的三尺度改为四尺度,并在特征融合网络中引入内容感知特征重组(Content Aware ReAs-sembly of Features,CARAFE)上采样模块替换原有的最近邻插值上采样,减少小目标特征信息损失,提高弱小破片的提取能力.其次,在特征提取网络引入坐标注意力模块(Coordinate Attention,CA),加强对破片特征的提取,弱化背景信息,抑制复杂背景的干扰.最后,在模型训练过程中引入模型不可知元学习方法(Model Agnostic Meta Learning,MAML),达到仅用小样本破片数据集实现较高的检测性能.实验结果表明,YOLOv5-FD破片检测算法在自制破片数据集中,精确率达到了90.5％,召回率达到了85.4％,平均精度mAP_0.5达到了88.2％,与原始YOLOv5s算法相比分别提高了7.1％,7.9％和7.5％,有效提高了破片目标检测准确性.
Abstract
The motion parameters of warhead fragment group has important significance for evaluating the damage power of ammunition. Aiming at the problems of low fragment object detection precision caused by small fragment object size,complex background information,and few fragment samples,this paper proposed a YOLOv5-FD(You Only Look Once v5-Fragment Detection) method for warhead fragment group object detection. Firstly,a small object detection layer was added at the network output layer,changing the original three-scales to four-scales,and a lightweight upsampling module called CARAFE (Content Aware ReAssembly of FEatures) was introduced in the feature fusion network to replace the original nearest neighbor interpolation upsampling,reducing the loss of small object feature information and improving the ability to extract small fragments. Secondly,the CA (Coordinate Attention) module was introduced into the feature extraction network to enhance of fragment features,weaken background in-formation,and suppress interference from complex backgrounds. Finally,the MAML (Model Agnostic Meta Learning) algorithm was introduced during the model training process to achieve high detection per-formance using only few-shot fragment datasets. The experimental results show that YOLOv5-FD algo-rithm achieves the precision of 90.5％,the recall rate of 85.4％,and the average precision of 88.2％ in the self-made fragment datasets. Compared with the original YOLOv5s algorithm,it improved by 7.1％,7.9％,and 7.5％,respectively,effectively improving the precision of fragment object detection.
维普
万方数据