基于坐标注意力机制的轻量化跨介质高速小目标检测方法

doi:10.12382/bgxb.2024.0380

摘要/Abstract

摘要：

在高速射弹测试领域,射弹的极高速度使探测技术面临显著挑战,实时检测与定位射弹的能力受限。由于射弹实验布设复杂、成本高昂,引起射弹检测可用数据集的稀缺。针对上述问题,搭建跨介质射弹入水测试系统,使用高速相机捕捉射弹,制作跨介质射弹数据集并采用数据增强的方法扩充现有数据集。针对射弹入水瞬间检测精度相较于射弹在空气和水中检测精度明显降低的问题,提出一种基于坐标注意力机制的跨介质小目标检测方法。该方法在高精度识别射弹小目标的基础上,替换传统的卷积为深度可分离卷积,在检测精度、速度和模型复杂度之间实现了良好的平衡。实验结果表明,新方法在射弹入水瞬间的检测精度提升了2.68%,召回率提升了7.31%,为跨介质高速小目标检测提供了解决方法。

关键词: 跨介质高速小目标, 目标检测, 注意力机制, 深度可分离卷积

Abstract:

In the field of high-speed projectile testing,the very high speed of projectiles poses a significant challenge to detection techniques,which in turn limits the ability to detect and localize projectiles in real time.Moreover,the available datasets for projectile detection are scarce due to the high cost and the complexity of projectile experimental setup.To address the above problems,a transmedia projectile water-entry test system is constructed.A transmedia projectile dataset is created from the projectile images captured by a high-speed camera, and expanded with data enhancement method.Aiming at the problem that the detection accuracy of projectile water-entry is significantly lower than that of projectile in air and water,a transmedia small target detection method based on a coordinate attention (CA) mechanism is proposed.The proposed method replaces the traditional convolution with depth-separable convolution on the basis of recognizing the projectile small targets with high accuracy,which achieves a good balance among detection accuracy,speed and model complexity.The experimental results show that the proposed method improves the detection accuracy by 2.68% and the recall ratio by 7.31% at the instant of projectile water-entry,which provides a solution for transmedia high-speed small target detection.

Key words: transmedia high-speed small target, target detection, attention mechanism, depth-separable convolution

中图分类号:

TP391

董毅, 孔筱芳, 罗红娥, 弯港, 夏言, 万敏杰. 基于坐标注意力机制的轻量化跨介质高速小目标检测方法[J]. 兵工学报, 2025, 46(6): 240380-.

DONG Yi, KONG Xiaofang, LUO Hong’e, WAN Gang, XIA Yan, WAN Minjie. Lightweight Transmedia High-speed Small Target Detection Method Based on Coordinate Attention Mechanism[J]. Acta Armamentarii, 2025, 46(6): 240380-.

图/表 24

图1 YOLOv7网络结构图

Fig.1 YOLOv7 network architecture diagram

图2 普通卷积示意图

Fig.2 Schematic diagram of ordinary convolution

图3 DSC示意图

Fig.3 Schematic diagram of depth-separable convolution

图4 DSC替换卷积层的训练效果图

Fig.4 The training effect of replacing the convolutional layer by DSC

图5 CA机制结构图

Fig.5 Structure of CA mechanism

图6 不同添加位置的mAP0.5对比图

Fig.6 Comparison of mAP0.5 with different additive positions

图7 射弹入水特征图

Fig.7 Characterization of projectile water-entry

图8 添加CA模块

Fig.8 Adding attention module

图9 总体实验布设

Fig.9 Overall experimental setup

表1 工况表

Table 1 Working conditions

工况编号	发射压力/MPa	波长/m	波陡/%	发射角度/(°)
1	4.0			30
2	4.0	1.9	3.0	30
3	4.0	1.9	4.5	30
4	4.0	1.9	6.0	30
5	4.5			30
6	4.5	1.9	3.0	30
7	4.5	1.9	4.5	30
8	4.5	1.9	6.0	30
9	5.0			30
10	5.0	1.9	3.0	30
11	5.0	1.9	4.5	30
12	5.0	1.9	6.0	30

表2 VEO 640L高速相机主要参数

Table 2 Main parameters of VEO 640L high-speed camera

参数	数值
传感器分辨率	1280×960
位深度/位	12
像素大小/μm	20
传感器尺寸/mm	25.6×16
帧频/(帧·s^-1)	4000
曝光时间/μs	20

表3 软硬件配置

Table 3 Hardware and software configuration

实验设备	配置参数
操作系统	Windows10
中央处理器CPU	12th Gen Intel(R) Core(TM) i7-12700 2.10GHz
图形处理器GPU	NVIDIA GeForce RTX 4070Ti
显存	12GB
深度学习框架	Pytorch
编程语言	Python 3.9
GPU加速器	CUDA 12.1

图10 高速相机捕获的射弹图片

Fig.10 Projectile images captured by high-speed camera

图11 数据增强后的图片

Fig.11 Images after data enhancement

图12 标记示意图

Fig.12 Labeled graph

图13 识别示意图

Fig.13 Schematic diagram of identification

表4 射弹识别置信度

Table 4 Projectile identification confidence levels

工况及平均置信度	发射压力/MPa	波陡/%	发射角度/(°)	检测置信度
工况及平均置信度	发射压力/MPa	波陡/%	发射角度/(°)	射弹在空气中	射弹入水瞬间	射弹在水中
工况1	4.0	3.0	30	0.85	0.83	0.91
工况2	4.0	4.5	30	0.88	0.79	0.94
工况3	4.0	6.0	30	0.93	0.86	0.95
工况4	4.5	3.0	30	0.94	0.56	0.94
工况5	4.5	4.5	30	0.88	0.89	0.94
工况6	4.5	6.0	30	0.96	0.80	0.94
平均置信度		0.91	0.79	0.94

图14 图像增强对比图

Fig.14 Image enhancement comparison

表5 不同模型的评价指标数值

Table 5 Values of evaluation indicators for different models

网络模型	mAP0.5/ %	参数量/ 百万	FLOPS/ Giga	帧频/ Hz
Faster R-CNN(ResNet)	97.12	41.36	269.00	14.36
Faster R-CNN(MobileNet)	86.74	82.36	137.48	21.36
YOLOv3SPP	94.18	62.58	117.12	41.73
YOLOv7	95.43	37.62	106.47	73.88

图15 各模型FPS柱状图

Fig.15 Histogram of FPS for each model

图16 添加CA机制前后热力图对比

Fig.16 Comparison of thermodynamic diagram before and after adding the CA mechanism

表6 数据增强对泛化能力的影响

Table 6 Effect of data enhancement on generalization ability

图像数据增强方法	P/%	R/%	mAP0.5/%
原始数据	81.4	85.37	92.83
图像扩充	90.87	81.39	94.33
图像扩充+CLAHE	91.89	82.93	95.43

表7 不同改进方法对模型性能的影响

Table 7 Effects of different improvement methods on model performance

网络模型	CA	DSC	Params/ M	FLOPS/ G	mAP0.5/ %	P/ %	R/ %
YOLOv7			37.62	106.47	95.43	91.89	82.93
		√	30.15	97.77	93.00	87.50	85.37
	√	√	30.41	97.79	95.52	94.87	90.24

图17 不同改进方法的mAP0.5曲线

Fig.17 mAP0.5 curves for different improvement methods

参考文献 22

[1]	WU J H, LIU J. Velocity measurement of a highspeed underwater projectile with laser barriers[J]. IEEE Access, 2021, 9:61193-61199.
[2]	孙嘉伟, 弯港, 顾金良, 等. 基于线阵相机的水下射弹动态参数及超空泡演化过程测量方法研究[J]. 兵工学报, 2024, 45(8):2564-2572. doi: 10.12382/bgxb.2023.0408
	SUN J W, WAN G, GU J L, et al. Research on the measurement method of underwater projectile dynamic parameters and super vacuole evolution process based on line array camera[J]. Acta Armamentarii, 2024, 45(8):2564-2572. (in Chinese)
[3]	侯宇, 黄振贵, 郭则庆, 等. 超空泡射弹小入水角高速斜入水试验研究[J]. 兵工学报, 2020, 41(2):332-341. doi: 10.3969/j.issn.1000-1093.2020.02.015
	HOU Y, HUANG Z G, GUO Z Q, et al. Experimental study on high-speed oblique water entry of super-vacuum projectile with small water entry angle[J]. Acta Armamentarii, 2020, 41(2):332-341. (in Chinese) doi: 10.3969/j.issn.1000-1093.2020.02.015
[4]	祁晓斌, 施瑶, 刘喜燕, 等. 阶梯式圆柱射弹小角度入水弹道特性研究[J]. 力学学报, 2023, 55(11):2468-2479.
	QI X B, SHI Y, LIU X Y, et al. Ballistic characterization of stepped cylindrical projectiles with small-angle water entry[J]. Journal of Mechanics, 2023, 55(11):2468-2479. (in Chinese)
[5]	路丽睿, 魏英杰, 王聪, 等. 不同头型射弹低速倾斜入水空泡及弹道特性试验研究[J]. 兵工学报, 2018, 39(7):1364-1371. doi: 10.3969/j.issn.1000-1093.2018.07.014
	LU L R, WEI Y J, WANG C, et al. Experimental study on low-speed tilted water entry vacuoles and ballistic characteristics of different head-type projectiles[J]. Acta Armamentarii, 2018, 39(7):1364-1371. (in Chinese)
[6]	EVERINGHAM M, VAN GOOL L, WILLIAMS C K I, et al. The pascal visual object classes (VOC) challenge[J]. International Journal of Computer Vision, 2010, 88:303-338.
[7]	JIAO L C, ZHANG F, LIU F, et al. A survey of deep learning-based object detection[J]. IEEE Access, 2019, 7:128837-128868. doi: 10.1109/ACCESS.2019.2939201
[8]	张侠. 基于深度学习的高速目标检测与跟踪研究[D]. 西安: 西安工业大学, 2022.
	ZHANG X. Research on high-speed target detection and tracking based on deep learning[D]. Xi’an: Xi’an University of Technology, 2022. (in Chinese)
[9]	谢富, 朱定局. 深度学习目标检测方法综述[J]. 计算机系统应用, 2022, 31(2):1-12
	XIE F, ZHU D J. A review of deep learning target detection methods[J]. Computer System Applications, 2022, 31(2):1-12. (in Chinese)
[10]	ZHAO Z Q, ZHENG P, XU S, et al. Object detection with deep learning:a review[J]. IEEE Transactions on Neural Networks and Learning Systems, 2019, 30(11):3212-3232.
[11]	杜芸彦, 杨锦辉, 李鸿, 等. 基于改进Faster RCNN的少样本目标检测算法[J]. 电光与控制, 2023, 30(5):44-51.
	DU Y Y, YANG J H, LI H, et al. Sample less target detection algorithm based on improved Faster RCNN[J]. Electro-Optics and Control, 2023, 30(5):44-51. (in Chinese)
[12]	LIU W, ANGUELOV D, ERHAN D, et al. S:Single shot multibox detector[C]// Proceedings of the Computer Vision-ECCV 2016:14th European Conference. Amsterdam,the Netherlands: Springer, 2016:21-37.
[13]	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.Washington,D.C.,US:IEEE, 2016:779-788.
[14]	REDMON J, FARHADI A. YOLOv3:an incremental improvement:arXiv:1804.02767[R]. Ithaca,NY,US:Cornell University, 2018:1804.02767.
[15]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv 4:optimal speed and accuracy of object detection:arXiv:2004.10934[R]. Ithaca,NY,US:Cornell University, 2020:2004.1093.
[16]	张旭, 陈慈发, 董方敏. 基于改进YOLOv7的PCB缺陷检测算法[J]. 计算机工程, 2024, 50(12):318-328. doi: 10.19678/j.issn.1000-3428.0068588
	ZHANG X, CHEN C F, DONG F M. PCB defect detection algorithm based on improved YOLOv7[J]. Computer Engineering, 2024, 50(12):318-328. (in Chinese) doi: 10.19678/j.issn.1000-3428.0068588
[17]	SHANG Z F, WANG C P. A detection algorithm for projectile target[J]. Advanced Materials Research, 2014, 1044:972-975.
[18]	唐云祁, 李同, 丁建伟, 等. 弹种自动识别算法研究与实现[J]. 中国人民公安大学学报(自然科学版), 2016, 22(1):24-26.
	TANG Y Q, LI T, DING J W, et al. Research and implementation of automatic identification algorithm of bullet species[J]. Journal of Chinese People’s Public Security University (Natural Science Edition), 2016, 22(1):24-26. (in Chinese)
[19]	QIU Q X, LI P F. Research on bullet recognition technology based on deep learning[C]// Proceedings of International Conference on Energy Storage and Intelligent Vehicles.Singapore:Springer, 2022:678-687.
[20]	WANG C Y, BOCHKOVSKIY A, LIAO H Y M. YOLOv7:trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington,D.C.,US:IEEE, 2023:7464-7475.
[21]	HOU Q B, ZHOU D Q, FENG J S. Coordinate attention for efficient mobile network design[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Washington,D.C.,US:IEEE, 2021:13713-13722.
[22]	杨环宇, 王军, 吴祥, 等. 一种坐标通道注意力深度学习网络的军用飞机识别方法[J]. 兵工学报, 2024, 45(7):2128-2143. doi: 10.12382/bgxb.2023.0526
	YANG H Y, WANG J, WU X, et al. A coordinate channel attention deep learning network for military aircraft recognition[J]. Acta Armamentarii, 2024, 45(7):2128-2143. (in Chinese)