基于孪生网络和区域注意力机制的球形爆炸破片毁伤效应识别研究

doi:10.12382/bgxb.2023.1215

摘要/Abstract

摘要：

在信息化战争中,爆炸物破片毁伤效应评估对实现精准打击具有重要意义,然而在毁伤实验中毁伤区域的分布和几何信息主要由人工统计,获取效率低下、精度不可控。为此,提出基于孪生网络和区域注意力机制的轻量化图像分割模型,实现小样本下对小目标球形爆炸破片毁伤区域的高效、精准识别功能。通过引入孪生结构、区域注意力模块和多尺度卷积模块提高模型对爆炸破孔的感知能力;加入多约束条件的损失函数,并筛选最佳优化器,使模型优化时更加聚焦有效信息,加速模型收敛;提出连通域融合分水岭算法的毁伤区域量化检测方法,实现爆炸破孔重叠情况下的精确识别。实验结果表明,相比目前主流模型,所提方法实现了更高的效率和精度,对毁伤区域面积和直径预测结平均误差分别为4.78%和3.79%;研究工作为实现含破片爆炸物毁伤智能化评估提供了参考。

关键词: 孪生网络, 毁伤效应评估, 损伤识别, 区域注意力机制, 多尺度卷积

Abstract:

In the information warfare, the damage assessment of explosive fragments is of great significance to achieve accurate strike. However, the manual acquisition of distribution and geometric information of damage areas are inefficient in the damage experiment. To this end, a lightweight image segmentation model based on siamese networks and regional attention mechanisms is proposed, which achieves the efficient and accurate recognition of small-targeted spherical explosive fragmentation damage area under small samples. The model’s ability to perceive the explosion holes is improved by introducing the siamese structure, regional attention module and multi-scale convolution module. A loss function with multiple constraints is added and the best optimizer is screened so that the model optimization is more focused on the effective information for accelerating the model convergence. A quantitative detection method for the damaged area based on the connected-domain fusion watershed algorithm is proposed to achieve the accurate identification of the overlapping case of explosion broken holes. Experimental results show that the proposed method achieves higher efficiency and accuracy compared with the current mainstream models, and the average errors in predicting the area and diameter of damage region are 4.78% and 3.79%, respectively. The research work provides a reference for realizing the intelligent damage assessment of explosives containing fragments.

Key words: siamese network, damage effect assessment, damage recognition, regional attention mechanism, multi-scale convolution

中图分类号:

TJ06
TP29

李豪天, 崔欣雨, 刘梦真, 黄广炎, 吕中杰, 张宏. 基于孪生网络和区域注意力机制的球形爆炸破片毁伤效应识别研究[J]. 兵工学报, 2024, 45(12): 4259-4271.

LI Haotian, CUI Xinyu, LIU Mengzhen, HUANG Guangyan, LÜ Zhongjie, ZHANG hong. Research on the Identification of Spherical Explosive Fragmentation Damage Effect Based on Siamese Networks and Regional Attention Mechanisms[J]. Acta Armamentarii, 2024, 45(12): 4259-4271.

图/表 20

图1 含破片爆炸实验布置示意图

Fig.1 Schematic diagram of the explosive test with fragmentation

图2 数据增强方法

Fig.2 Data enhancement methods

图3 毁伤区域矩阵标注示例

Fig.3 Example of damage area matrix labeling

图4 毁伤区域标注示例

Fig.4 Annotation of damage area

图5 基于SALU-Net的毁伤区域量化算法结构

Fig.5 Structureof damage area quantization algorithm based SALU-Net

图6 孪生分类模块

Fig.6 Siamese classification module

图7 区域注意力机制

Fig.7 Regional attention mechanism

图8 区域注意力增强可视效果

Fig.8 Regional attention enhanced visualization

图9 不同扩张率下空洞卷积的感受野

Fig.9 Feeling fields fordilated convolution at different dilation rates

图10 基于空洞卷积的多尺度卷积模块

Fig.10 Multiscale convolution module based on dilated convolution

表1 各阶段网络参数

Table 1 Network parameters at each stage

阶段	学习率	损失	训练轮次
1	0.02	L_s+L_c	10
2	0.002	L_s+L_c+L_seg	50
3	0.0002	L_seg	10

表2 网络结构模块的消融实验结果

Table 2 Ablation experimental results of network structure modules

模型	查准率/%	召回率/%	F1值/ %	平均交并比/%	单帧耗时/%
LU-Net	59.1	94.7	72.8	78.3	14.5
LU-Net+MCL	81.7	90.4	85.8	87.5	14.5
LU-Net+MCL+RAM	82.5	94.1	87.9	89.2	15.1
LU-Net+MCL+ RAM+MCM	88.6	95.3	91.7	92.4	17.5

图11 4种优化器在训练过程中损失随训练周期的变化

Fig.11 Changes of loss with epoch for four optimizers during training

表3 4种优化器的实验结果

Table 3 Experimental results of four optimizers

优化器	F1值/%	MIoU/%
SGD	91.785	92.357
Adagrad	91.250	91.900
Adam	94.669	94.907
RMSprop	95.241	95.428

表4 不同分割模型在本文数据集上的评价指标

Table 4 Evaluation indexes of the different segmentation models on the datasets in this paper

模型	F1值/%	MIoU/%	单帧耗时/ms
Deeplabv3^[34]	83.553	85.760	40.6
U-Net^[24]	94.296	94.570	47.3
MALUNet^[35]	79.997	83.206	12.6
SALU-Net	95.241	95.428	17.5

图12 毁伤区域量化结果

Fig.12 Quantitative results of damage area

图13 破孔面积和直径的误差率分布

Fig.13 Error rate distribution of broken hole area and diameter

图14 破孔误差率样本统计

Fig.14 Histogram of sample statistics of broken hole error rate

表5 SALU-Net模型毁伤区域识别效果

Table 5 The damage identification effect of SALU-Net model

图15 破片毁伤区域可视化检测平台

Fig.15 Platform for visualization and detection of fragmentation damage areas

参考文献 36

[1]	李峰, 石全, 孙正. 目标毁伤效果评估技术研究综述[J]. 兵器装备工程学报, 2018, 39(9): 69-72.
	LI F, SHI Q, SUN Z. Summary of technical research on the evaluation of target mutilation[J]. Journal of Ordnance Equipment Engineering, 2018, 39(9): 69-72. (in Chinese)
[2]	LI H S, ZHANG X Q, ZHANG X W. Calculation model and method of target damage efficiency assessment based on warhead fragment dispersion[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1-8.
[3]	JENA B, SAXENA S, NAYAK G K, et al. Artificial intelligence-based hybrid deep learning models for image classification: the first narrative review[J]. Computers in Biology and Medicine, 2021, 137: 104803.
[4]	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
[5]	THOMA M. A survey of semantic segmentation: arXiv:1602.06541[R/OL]. Ithaca, NY, US: Cornell University, 2016(2016-05-11). https://arxiv.org/abs/1602.06541.
[6]	BRODLEY C E, UTGOFF P E. Multivariate decision trees[J]. Machine Learning, 1995, 19(1): 45-77.
[7]	SHAWE-TAYLOR J, SUN S. A review of optimization methodologies in support vector machines[J]. Neurocomputing, 2011, 74(17): 3609-3618.
[8]	de AMORIM R C. A survey on feature weighting based K-Means algorithms: arXiv:1602.06541[R/OL]. Ithaca, NY, US: Cornell University, 2016(2016-05-11). https://arxiv.org/abs/1602.06541.
[9]	LECUN Y, BOTTOU L, BENGIO Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.
[10]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. ImageNet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
[11]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation: arXiv:1311.2524[R/OL]. Ithaca, NY, US: Cornell University, 2014(2014-10-22). https://arxiv.org/abs/1311.2524.
[12]	WEI F J, YAO G, YANG Y, et al. Instance-level recognition and quantification for concrete surface bughole based on deep learning[J]. Automation in Construction, 2019, 107: 102920.
[13]	雷江波, 王泽民, 李静. 基于Faster R-CNN的破片群图像目标检测研究[J]. 国外电子测量技术, 2021, 40(1): 70-74.
	LEI J B, WANG Z M, LI J. Detection of objects in fragments based on Faster R-CNN[J]. Foreign Electronic Measurement Technology, 2021, 40(1): 70-74. (in Chinese)
[14]	魏琦, 何子清, 王亚林, 等. 基于DenseNet和注意力机制的静爆场破片识别方法研究[J]. 兵器装备工程学报, 2023, 44(2): 259-265.
	WEI Q, HE Z Q, WANG Y L, et al. Research on fragment identification of a static explosion field based on DenseNet and attention mechanism[J]. Journal of Ordnance Equipment Engineering, 2023, 44(2): 259-265. (in Chinese)
[15]	LIU D B, HE Z N, CHEN D D, et al. A network framework for small-sample learning[J]. IEEE Transactions on Neural Networks and Learning Systems, 2020, 31(10): 4049-4062.
[16]	WANG Y Q, YAO Q M, KWOK J, et al. Generalizing from a few examples: a survey on few-shot learning: arXiv:1904.05046[R/OL]. Ithaca, NY, US: Cornell University, 2020(2020-03-29). https://arxiv.org/abs/1904.05046.
[17]	FINN C, ABBEEL P, LEVINE S. Model-agnostic meta-learning for fast adaptation of deep networks: arXiv:1703.03400[R/OL]. Ithaca, NY, US: Cornell University, 2017(2017-07-18). https://arxiv.org/abs/1703.03400.
[18]	YI C C, CHEN Q R, XU B, et al. Steel strip defect sample generation method based on fusible feature GAN model under few samples[J]. Sensors, 2023, 23(6): 3216.
[19]	WANG H H, LI Z L, WANG H Q. Few-shot steel surface defect detection[J]. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 5003912.
[20]	XU H S, ZHU Y C. Real-time object tracking based on improved fully-convolutional siamese network[J]. Computers & Electrical Engineering, 2020, 86: 106755.
[21]	FENG K X, CHASPARI T. Few-shot learning in emotion recognition of spontaneous speech using a siamese neural network with adaptive sample pair formation[J]. IEEE Transactions on Affective Computing, 2023, 14(2): 1627-1633.
[22]	PEI M T, YAN B, HAO H L, et al. Person-specific face spoofing detection based on a siamese network[J]. Pattern Recognition, 2023, 135: 109148.
[23]	LUAN C H, JING Z B, ZUO J H. A defect-sensitive loss function based on siamese network to defect detection with imbalanced samples[J]. Journal of Physics: Conference Series, 2021, 1802(4): 042085.
[24]	ZHAO X P, MA M Y, SHAO F. Bearing fault diagnosis method based on improved Siamese neural network with small sample[J]. Journal of Cloud Computing, 2022, 11(1): 79.
[25]	RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015. Munich, Germany:Springer, 2015: 234-241.
[26]	NIU Z Y, ZHONG G Q, YU H. A review on the attention mechanism of deep learning[J]. Neurocomputing, 2021, 452: 48-62.
[27]	YU F, KOLTUN V. Multi-scale context aggregation by dilated convolutions: arXiv:1511.07122[R/OL]. Ithaca, NY, US: arXivCornell University, 2016(2016-04-30). https://arxiv.org/abs/1511.07122.
[28]	惠康华, 杨卫, 刘浩翰, 等. 基于YOLOv5的增强多尺度目标检测方法[J]. 兵工学报, 2023, 44(9): 2600-2610. doi: 10.12382/bgxb.2022.1147
	HUI K H, YANG W, LIU H H, et al. Enhanced multi-scale target detection method based on YOLOv5[J]. Acta Armamentarii, 2023, 44(9): 2600-2610. (in Chinese) doi: 10.12382/bgxb.2022.1147
[29]	SZEGEDY C, IOFFE S, VANHOUCKE V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning: arXiv:1602.07261[R/OL]. Ithaca, NY, US: Cornell University, 2016(2016-08-23). https://arxiv.org/abs/1602.07261.
[30]	ZHOU D Q, HOU Q B, CHEN Y P, et al. Rethinking bottleneck structure for efficient mobile network design[C]//Computer Vision-ECCV 2020. Glasgow, UK:Springer, 2020.
[31]	LI X Y, SUN X F, MENG Y X, et al. Dice loss for data-imbalanced NLP tasks[C]// Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Online: Association for Computational Linguistics, 2020: 465-476.
[32]	ADYA M, COLLOPY F. How effective are neural networks at forecasting and prediction? a review and evaluation[J]. Journal of Forecasting, 1998, 17(5/6): 481-495.
[33]	刘梦真, 黄广炎, 张宏, 等. 小样本驱动特征分段网络的防护材料折痕检测[J]. 兵工学报, 2024, 45(3): 963-974. doi: 10.12382/bgxb.2022.0884
	LIU M Z, HUANG G Y, ZHANG H, et al. Protective material crease detection with small sample-driven feature segmented neural network[J]. Acta Armamentarii, 2024, 45(3): 963-974. (in Chinese) doi: 10.12382/bgxb.2022.0884
[34]	RUAN J C, XIANG S C, XIE M Y, et al. MALUNet: a multi-attention and light-weight unet for skin lesion segmentation: arXiv:2211.01784[R/OL]. Ithaca, NY, US: Cornell University, 2022(2022-11-03). https://arxiv.org/abs/2211.01784.
[35]	CHEN L C, ZHU Y, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]// Computer Vision-ECCV 2018. Munich, Germany: Springer, 2018: 833-851.
[36]	PENG C, LIU Y K, GUI W H, et al. Bubble Image segmentation based on a novel watershed algorithm with an optimized mark and edge constraint[J]. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 5005110.