一种用于外场试验图像的引信炸点检测方法

doi:10.12382/bgxb.2022.0300

摘要/Abstract

摘要：

在无线电近炸引信抗干扰性能测试中,观测和识别炸点状态对引信的工作状态评估和性能改进至关重要。为此,提出一种基于深度神经网络的图像目标检测算法,用于引信炸点识别。在检测模型的结构设计及训练策略上实现以下新颖设计:模型基于高性能骨干网络ConvNeXt实现图像目标特征提取,使用密集连接跨阶段局部模块以及带通道注意力机制的多分支模块,以提升特征提取能力;使用任务解耦多检测头结构提升检测精度;在模型训练时,使用焦点损失函数作为分类和置信度的损失函数,使用完全交并比函数作为预测框回归的损失函数。研究结果表明:该检测算法在真实引信炸点图像数据集上实现平均精度为92.7%以及F1分数为87.4%的检测性能。实验结果表明,所提算法在引信炸点检测任务上性能优于现有典型检测模型。

关键词: 炸点观测, 目标检测, 深度学习, 卷积神经网络, 特征融合

Abstract:

In the anti-jamming performance test of radio proximity fuzes, the observation and recognition of the burst point state is essential for working state evaluation and performance improvement of the fuzes. Therefore, an image target detection algorithm based on the deep neural network is proposed for fuze burst point recognition. The algorithm achieves the following novel designs in the structure and training strategy of the model: The model realizes the target feature extraction based on the high-performance backbone ConvNeXt, and uses the cross stage partial structure based on dense connection and the multi-branch structure with channel attention to improve the feature extraction capability; it also applies a task-decoupled multi-detector structure to improve detection accuracy; focus loss functions are used as the loss functions of classification and confidence, and the complete intersection over union loss function is used as the loss function of prediction box regression in the model training. The proposed algorithm achieves an average precision of 92.7% and F1-score of 87.4% on the real fuze burst point image dataset. The results show the superiority of the proposed algorithm over the existing typical models in the fuze burst point detection task.

Key words: burst point observation, target detection, deep learning, convolutional neural network, feature

中图分类号:

TP391

周宇, 曹荣刚, 栗苹, 马啸. 一种用于外场试验图像的引信炸点检测方法[J]. 兵工学报, 2023, 44(8): 2453-2464.

ZHOU Yu, CAO Ronggang, LI Ping, MA Xiao. A Fuze Burst Point Detection Method for Outfield Test Images[J]. Acta Armamentarii, 2023, 44(8): 2453-2464.

图/表 16

图1 不同情况下的炸点形态

Fig.1 Burst point shape under different conditions

图2 观测设备总体布设关系示意图

Fig.2 Schematic diagram of overall layout relationship of observation equipment

图3 ConvNeXt网络结构

Fig.3 Structure of ConvNeXt network

图4 原多分支结构与带通道注意力机制的多分支模块

Fig.4 The original multi-branch structure and the multi-branch with channel attention module

图5 原始跨阶段局部模块与密集连接跨阶段局部模块

Fig.5 Original CSP Module and proposed DCSP Module

图6 解耦检测头

Fig.6 Decoupling detection head

图7 提出的检测算法模型的整体结构

Fig.7 The overall structure of the proposed detection algorithm model

表1 数据集目标标注真实框信息统计

Table 1 Information statistics of the annotated ground truth box of dataset target

目标类别	目标总数	面积占比/%
目标类别	目标总数	最大值	最小值	中位数	均值
1	2578	11.04	0.10	1.91	2.18
2	2180	6.94	1.27	3.19	3.59
3	2020	10.87	2.14	5.19	5.39

图8 数据增强示意图

Fig.8 Schematic diagram of the data augmentation

表2 骨干网络和检测头改进设计对模型性能的影响

Table 2 Effects of improved designs of backbone and head on detection model performance

实验组	骨干网络	检测头	准确率/%	召回率/%	平均精度/%	F1分数/%
模型1	DarkNet53	多检测头	74.6	55.0	74.7	59.8
模型2	ConvNeXt	单检测头	87.8	56.5	76.9	55.6
模型3	ConvNeXt	多检测头	85.0	57.2	79.8	65.7

图9 骨干网络和检测头改进设计消融实验中检测模型的PR曲线

Fig.9 The PR curves of detection models in the ablation experiment of improved designs of backbone and head

表3 颈部结构改进措施对模型性能的影响

Table 3 Effects of improvement measures of neck on detection model performance

实验组	DCSP 模块	MBCA 模块	准确率/%	召回率/ %	平均精度/%	F1分数/%
模型3			85.0	57.2	79.8	65.7
模型4	√		84.2	85.1	88.6	84.5
模型5		√	89.3	84.6	90.7	86.8
模型6	√	√	87.3	87.8	92.7	87.4

图10 颈部结构改进措施消融实验中检测模型的PR曲线

Fig.10 The PR curves of detection models in the ablation experiment of improvement measures for neck structure

表4 检测模型颈部结构改进前后的检测结果对比

Table 4 Comparison of the detection results before and after the improvement of the neck structure of the detection model

图11 颈部改进结构对特征图的强化效果

Fig.11 Enhancement effect of improved neck structure on feature map

表5 基于提出数据集的不同检测模型性能比较

Table 5 Performance comparison of different detection models based on the proposed dataset

检测模型	准确率/%	召回率/%	平均精度/%	F1分数/%
Faster RCNN 模型	63.2	85.4	81.3	70.2
SSD模型	78.4	62.3	74.5	61.4
YOLOv3模型	80.5	59.4	76.1	64.8
YOLOX模型	87.9	74.1	87.5	81.1
本文模型	87.3	87.8	92.7	87.4

参考文献 25

[1]	周慧. 靶场光电经纬仪最优布站方法研究[J]. 测控技术, 2018, 37(3):142-144,149.
	ZHOU H. Reseach on optimal station distribution method of photoelectric theodolite in range[J]. Measurement & Control Technology, 2018, 37(3):142-144,149. (in Chinese)
[2]	董明荣, 许学忠, 张彤, 等. 空中炸点三基阵声学定位技术研究[J]. 兵工学报, 2010, 31(3):343-349.
	DONG M R, XU X Z, ZHANG T, et al. Research on three-array acoustic localization technology for the aerial blast points[J]. Acta Armamentarii, 2010, 31(3):343-349. (in Chinese)
[3]	吕韶昱, 刘蔚, 孟庆慈. 差分GPS技术在靶场弹道测量中的应用研究[J]. 测控技术, 2003, 22(7):57-61.
	LÜ S Y, LIU W, MENG Q C. Study on the application of dgps to trajectory measure in the proving ground[J]. Measurement & Control Technology, 2003, 22(7):57-61. (in Chinese)
[4]	李翰山, 雷志勇, 王泽民, 等. 一种新型天幕交汇立靶精度测量系统[J]. 电光与控制, 2007, 14(5):153-155,159.
	LI H S, LEI Z Y, WANG Z M, et al. A new type of firing accuracy measurement system for cross sky screen[J]. Electronics Optics & Control, 2007, 14(5):153-155,159. (in Chinese)
[5]	李翰山, 雷志勇, 王泽民, 等. 多光幕法测量高空弹丸炸点坐标原理与分析[J]. 仪器仪表学报, 2009, 30(3):621-624.
	LI H S, LEI Z Y, WANG Z M, et al. Principle and analysis of high altitude projectile location measurement using multi-screen target method[J]. Chinese Journal of Scientific Instrument, 2009, 30(3):621-624. (in Chinese)
[6]	李静, 雷志勇, 王泽民, 等. 近炸引信对空炸点位置测试方法研究[J]. 计算机测量与控制, 2011, 19(7):1602-1604.
	LI J, LEI Z Y, WANG Z M, et al. Research on test method of anti-air proximity fuze burst location[J]. Computer Measurement & Control, 2011, 19(7):1602-1604. (in Chinese)
[7]	秦晓燕, 袁广林. 基于主分量寻踪与分析的炮弹炸点检测[J]. 兵工学报, 2013, 34(10):1266-1272. doi: 10. 3969/ j. issn. 1000-1093. 2013. 10. 011
	QING X Y, YUAN G L. Detection of artillery blast point based on principal component pursuit and analysis[J]. Acta Armamentarii, 2013, 34(10):1266-1272. (in Chinese)
[8]	张冠武, 连豪, 谢敏. 战场侦察雷达对炸点目标的识别方法[J]. 雷达科学与技术, 2012, 10(4):396-399.
	ZHANG G W, LIAN H, XIE M. Classification method of howitzer impact point for a battlefield surveillance radar[J]. Radar Science and Technology, 2012, 10(4):396-399. (in Chinese)
[9]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 39(6):1137-1149. doi: 10.1109/TPAMI.2016.2577031 URL
[10]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, US:IEEE, 2016: 779-788.
[11]	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, US:IEEE, 2017: 6517-6525.
[12]	REDMON J, FARHADI A. YOLOv3: an incremental improvement: arXiv: 1804.02767v1[R/OL]. Ithaca, NY, US: Cornell University, 2018(2018-04-08) [2019-09-15]. https://arxiv.org/abs/1804.02767v1
[13]	LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, US:IEEE, 2017:2117-2125.
[14]	LIU Z, MAO H Z, WU C Y, et al. A ConvNet for the 2020s: arXiv:2201.03545v2[R/OL]. Ithaca, NY, US: Cornell University, 2022(2022-03-02) [2022-04-19]. https://arxiv.org/abs/2201.03545v2
[15]	DAN H, KEVIN G. Gaussian error linear units (GELUs): arXiv: 1606.08415v4[R/OL]. Ithaca, NY, US: Cornell University, 2020(2020-07-08) [2022-04-19]. https://arxiv.org/abs/1606.08415v4
[16]	TOUVRON H, CORD M, SABLAYROLLES A, et al. Going deeper with image transformers[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, QC, Canada:IEEE, 2021: 32-42.
[17]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, US:IEEE, 2016: 770-778.
[18]	LIU Z, LIN Y T, CAO Y, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal, QC, Canada:IEEE, 2021: 10012-10022.
[19]	SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston, MA, US:IEEE, 2015: 1-9.
[20]	ELFWING S, UCHIBE E, DOYA K. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning[J]. Neural Networks, 2018, 107: 3-11. doi: S0893-6080(17)30297-6 pmid: 29395652
[21]	WANG C Y, LIAO H Y M, WU Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Seattle, WA, US:IEEE, 2020: 390-391.
[22]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, US:IEEE, 2017: 4700-4708.
[23]	LIU S, QI L, QIN H F, et al. Path aggregation network for instance segmentation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, US:IEEE, 2018: 8759-8768.
[24]	NING C, ZHOU H, SONG Y, et al. Inception single shot multibox detector for object detection[C]// Proceedings of 2017 IEEE International Conference on Multimedia and Expo Workshops. Hong Kong, China: IEEE, 2017: 549-554.
[25]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]// IProceedings of the IEEE International Conference on Computer Vision. Venice, taly:IEEE, 2017: 2980-2988.