小样本驱动特征分段网络的防护材料折痕检测

doi:10.12382/bgxb.2022.0884

摘要/Abstract

摘要：

防刺服能在恐怖袭击、医闹伤害、违法犯罪等事件中有效保护生命安全,然而在生产制造及穿着使用中易产生机械折痕。立足于防护材料折痕缺陷的快速检测需求,创新性地在图像识别方法中提出特征分段网络结构,实现了小样本驱动下防护材料折痕的快速、精准检测功能。通过引入注意力机制和深度可分离卷积模块,并赋予损失函数与优化器两种典型参数,全面提高了特征分段网络模型的检测精度与效率;提出几何信息标注算法,搭建防护材料缺陷可视化检测平台,实现了机械折痕自动精准定位与几何信息输出。模型训练结果表明,特征分段网络模型的准确率可达96.19%,折痕缺陷几何信息标注误差在2%以内,优异的可视化检测功能可拓展到大型工程化自动检测领域。研究工作为下一步构建含有折痕缺陷的防刺装备防护性能预测模型奠定了基础。

关键词: 防护材料, 机械折痕检测, 特征分段神经网络, 几何信息标注

Abstract:

Stab-proof clothing can effectively protect life safety in terrorist attacks, medical problems, breaking the law and committing crimes and other incidents, but the mechanical creases of tab-proof clothing are easily produced in production and wearing. Based on the demand for rapid detection of crease defects of protective materials, a small sample-driven feature segmented neural network structure is proposed innovatively in the image recognition method, and the rapid and accurate detection of crease defects is realized. By introducing the attention mechanism and the depth-separable convolution module and giving the loss function and the optimizer two typical parameters, the detection accuracy and efficiency of the feature segmented neural network are improved comprehensively. A geometric information annotation algorithm is proposed and a visual detection platform is built for defect detection of protective materials, realizing the automatic and accurate location of mechanical creases and the output of geometric information. The results show that the accuracy of the model can reach 96.19%, and the annotation error of geometric information is less than 2%. The excellent visual detection function can be extended to the field of large-scale engineering automatic detection. The research work lays a foundation for constructing a protective performance prediction model of the stab-proof equipment with crease defects.

Key words: protective material, mechanical creases detection, feature segmented neural network, geometric information annotation

中图分类号:

刘梦真, 黄广炎, 张宏, 周宏元, 刘思宇. 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.

LIU Mengzhen, HUANG Guangyan, ZHANG Hong, ZHOU Hongyuan, LIU Siyu. Protective Material Crease Detection with Small Sample-driven Feature Segmented Neural Network[J]. Acta Armamentarii, 2024, 45(3): 963-974.

图/表 21

图1 图像采集系统

Fig.1 Image acquisition system

表1 系统硬件参数

Table 1 System hardware parameters

硬件	主要参数	厂商
工业相机	分辨率4022×3036	杭州海康威视数字技术股份有限公司
镜头	焦距8mm	杭州海康威视数字技术股份有限公司
工业光源	300mm×300mm	杭州海康威视数字技术股份有限公司
光源控制器	2通道,24V光源	大康控股集团有限公司
相机夹具	最大距离100mm	定制
可拼接镀铬精密管	ϕ25×400mm	定制
光源架	280mm×180mm	定制
透明板	280mm×180mm	定制
底板	300mm×300mm	定制

图2 使用FC370型防护材料的防刺服

Fig.2 Stab-proof clothing made of FC370 protective material

图3 无机械折痕图像(左)与压痕机(右)

Fig.3 Image without mechanical crease(left) and creasing machine(right)

图4 产生左下至右上(上)、左上至右下(中)以及竖直(下)机械折痕的防护材料图像

Fig.4 Images of protective material that produce bottom-left to top-right (upper), top-left to bottom-right (middle), and vertical (below) mechanical creases

图5 无机械折痕图像的标注

Fig.5 Annotation of image without mechanical crease

图6 有机械折痕图像的标注

Fig.6 Annotation of image with mechanical crease

图7 整体结构图

Fig.7 Overall structure diagram

表2 分割网络结构主要参数

Table 2 Major parameters of segmented network structure

层	网络	输入通道数	输出通道数	核尺寸	图像尺寸
	二维卷积层	3	32	5×5	1024×1024
1	二维卷积层	32	32	5×5	1024×1024
	池化层	32	32	2×2	512×512
	二维卷积层	32	64	5×5	512×512
2	二维卷积层	64	64	5×5	512×512
	二维卷积层	64	64	5×5	512×512
	池化层	64	64	2×2	256×256
	二维卷积层	64	64	5×5	256×256
	二维卷积层	64	64	5×5	256×256
3	二维卷积层	64	64	5×5	256×256
	二维卷积层	64	64	5×5	256×256
	池化层	64	64	2×2	128×128
注意力层	二维卷积层	64	4	1×1	128×128
	二维卷积层	4	64	1×1	128×128
4	二维卷积层	64	1024	15×15	128×128
5	二维卷积层	1024	1	1×1	128×128

表3 决策网络结构主要参数

Table 3 Major parameters of decision network structure

层	网络	输入通道数	输出通道数	核尺寸	图像尺寸
深度可分离卷积层	二维卷积层	1 025	1 025	5×5	128×128
	二维卷积层	1 025	8	5×5	128×128
	池化层	8	8	2×2	64×64
	二维卷积层	8	8	5×5	64×64
	二维卷积层	8	16	5×5	64×64
	池化层	16	16	2×2	32×32
	二维卷积层	16	32	5×5	32×32
1	线性层	66	1	—	1×1

图8 深度可分离卷积原理图

Fig.8 Schematic diagram of depth-wise separable convolution

图9 几何信息标注算法流程图

Fig.9 Flowchart of geometric information labeling algorithm

表4 网络训练主要参数

Table 4 Major parameters of training neural network

网络	学习率	权重衰减系数	单次训练样本数
分割网络	0.00002	0.5	2
决策网络	0.002	0.5	2

图10 损失值随训练周期的变化图

Fig.10 Plot of loss value versus training epoch

图11 不同训练周期时的损失值

Fig.11 Loss values of different training epochs

图12 不同参数时的训练时间

Fig.12 Training time with different parameter combinations

表5 模型效果的结果对比

Table 5 Comparison of the results of the models

模型	准确率/%	查准率/%	召回率/%	F1 值	图像处理帧率/ (帧·s^-1)
Models_Original	67.62	92.86	55.71	69.64	5.96
Models_Depthwise	87.62	95.97	85.00	90.15	7.47
Models_Depthwise_SE	96.19	100.00	94.29	97.06	6.06
YOLOv5s	84.60	82.60	84.00	83.28	46.95

表6 Models_Depthwise_SE模型的预测结果

Table 6 Predicted results of Models_Depthwise_SE

表7 噪声干扰下模型检测效果

Table 7 Model detection results under noise interference

表8 几何信息标注结果

Table 8 Result of geometric information annotation

图13 防护材料缺陷检测平台示意图

Fig.13 Schematic diagram of the defect detection platform for protective material

参考文献 29

[1]	邱日祥, 韩启龙. 防刺服的现状与发展[J]. 装备发展, 2020(5):77-81.
	QIU R X, HAN Q L. Current situation and development of stab resistant clothing[J]. Police Technology, 2020(5):7-81. (in Chinese)
[2]	冯浩, 汪泽幸, 何斌, 等. 防刺服的研究与发展现状[J]. 产业用纺织品, 2021, 39(2):8-13,17.
	FENG H, WANG Z X, HE B, et al. Research and existing situation in stab resistance clothing[J]. Technical Textiles, 2021, 39(2):8-13,17. (in Chinese)
[3]	BULNES F G, USAMENTIAGA R, GARCIA D F, et al. An efficient method for defect detection during the manufacturing of web materials[J]. Journal of Intelligent Manufacturing, 2016, 27(2):431-445. doi: 10.1007/s10845-014-0876-9 URL
[4]	ZHANG K B, YAN Y D, LI P F, et al. Fabric defect detection using salience metric for color dissimilarity and positional aggregation[J]. IEEE Access, 2018, 6:49170-49181. doi: 10.1109/ACCESS.2018.2868059 URL
[5]	ZHU H J, WANG Y C, FAN J W. IA-Mask R-CNN: improved anchor design mask R-CNN for surface defect detection of automotive engine parts[J]. Applied Sciences, 2022, 12(13): 6633. doi: 10.3390/app12136633 URL
[6]	WU C H, WANG D Z, IP A, et al. A particle swarm optimization approach for components placement inspection on printed circuit boards[J]. Journal of Intelligent Manufacturing, 2009, 20(5):535-549. doi: 10.1007/s10845-008-0140-2 URL
[7]	廖达海, 殷明帅, 罗宏斌, 等. 基于耦合去噪算法的航空发动机中Si₃N₄圆柱滚子表面缺陷的检测方法[J]. 兵工学报, 2022, 43(1):191-198.
	LIAO D H, YIN M S, LUO H B, et al. Detection and analysis of surface defects of Si₃N₄ cylindrical roller in aero-engine based on coupled denoising algorithm[J]. Acta Armamentarii, 2022, 43(1):191-198. (in Chinese)
[8]	张良安, 陈洋, 谢胜龙, 等. 基于机器视觉与深度学习的飞机防护栅裂纹检测系统[J]. 兵工学报, 2023, 44(2):507-516.
	ZHANG L A, CHEN Y, XIE S L, et al. The detection system for cracks in aircraft protective grid based on the technology of machine vision and deep learning[J]. Acta Armamentarii, 2023, 44(2):507-516. (in Chinese)
[9]	MITCHELL T M, MITCHELL T M. Machine learning[M]. New York,NY,US: McGraw-Hill, 1997.
[10]	YE D S, FUH Y, ZHANG Y, et al. Defects recognition in selective laser melting with acoustic signals by SVM based on feature reduction[C]//Proceedings of the IOP Conference Series:Materials Science and Engineering. 2018. DOI:10.1088/1757-899X/436/1/012020.
[11]	GARCIA-MORENO A I, ALVARADO-OROZCO J M, IBARRA-MEDINA J, et al. Image-based porosity classification in Al-alloys by laser metal deposition using random forests[J]. The International Journal of Advanced Manufacturing Technology, 2020, 110(9):2827-2845. doi: 10.1007/s00170-020-05887-6
[12]	AMINZADEH M, KURFESS T R. Online quality inspection using Bayesian classification in powder-bed additive manufacturing from high-resolution visual camera images[J]. Journal of Intelligent Manufacturing, 2019, 30(6):2505-2523. doi: 10.1007/s10845-018-1412-0
[13]	CHEN L Q, YAO X, XU P, et al. Rapid surface defect identification for additive manufacturing with in-situ point cloud processing and machine learning[J]. Virtual and Physical Prototyping, 2021, 16(1):50-67. doi: 10.1080/17452759.2020.1832695 URL
[14]	HINTON G E, SALAKHUTDINOV R. Reducing the dimensionality of data with neural networks[J]. Science, 2006, 313(5786):504-507. doi: 10.1126/science.1127647 pmid: 16873662
[15]	KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6):84-90. doi: 10.1145/3065386 URL
[16]	MASCI J, MEIER U, DAN C, et al. Steel defect classification with Max-Pooling Convolutional Neural Networks[C]//Proceedings of the 2012 International Joint Conference on Neural Networks.Brisbane, QLD,Australia:IEEE, 2012.
[17]	张博尧, 冷雁冰. 基于YOLOv4网络模型的金属表面划痕检测[J]. 兵工学报, 2022, 43(增刊1):214-221.
	ZHANG B Y, LENG Y B. Detection of metal surface scratch based on YOLOv4 network model[J]. Acta Armamentarii, 2022, 43(S1): 214-221. (in Chinese) doi: 10.12382/bgxb.2022.A011
[18]	CHEN P H, HO S S. Is overfeat useful for image-based surface defect classification tasks?[C]//Proceedings of the IEEE International Conference on Image Processing. Phoenix, AZ,US:IEEE, 2016.
[19]	WEIMER D, SCHOLZ-REITER B, SHPITALNI M. Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection[J]. CIRP Annals, 2016, 65(1):417-420. doi: 10.1016/j.cirp.2016.04.072 URL
[20]	CHU L Y, HU X, HU J H, et al. Exact and consistent interpretation for piecewise linear neural networks: a closed form solution[C]//Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York, NY,US:ACM, 2018.
[21]	RUSSELL B C, TORRALBA A, MURPHY K P, et al. LabelMe:a database and web-based tool for image annotation[J]. International Journal of Computer Vision, 2008, 77(1): 157-173. doi: 10.1007/s11263-007-0090-8 URL
[22]	TORRALBA A, RUSSELL B C, YUEN J. LabelMe:online image annotation and applications[J]. Proceedings of the IEEE, 2010, 98(8):1467-1684. doi: 10.1109/JPROC.2010.2050290 URL
[23]	HU J, SHEN L, SUN G, et al. Squeeze-and-excitation networks[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT,US:IEEE, 2017.
[24]	CHOLLET F. Xception:deep learning with depthwise separable convolutions[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI,US:IEEE, 2017.
[25]	BAO P, ZHANG L, WU X L. Canny edge detection enhancement by scale multiplication[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(9):1485-1490. pmid: 16173190
[26]	CHUNG J Y, GULCEHRE C, CHO K H, et al. Empirical evaluation of gated recurrent neural networks on sequence modeling:arXiv:1412.3555[R]. Ithaca,NY,US: Cornell University, 2014:14123555.
[27]	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. doi: 10.1002/(ISSN)1099-131X URL
[28]	DICKSON M C, BOSMAN A S, MALAN K M. Hybridized loss functions for improved neural network generalization[C]//Proceedings of the Pan-African Artificial Intelligence and Smart Systems Conference.Berlin,Germany:Springer, 2021.
[29]	ZHANG Z J. Improved adam optimizer for deep neural networks[C]//Proceedings of the 2018 IEEE/ACM 26th International Symposium on Quality of Service. New York, NY,US:ACM, 2018.