基于卷积神经网络的光学元件表面缺陷图像分类

侯劲尧; 刘卫国; 周顺; 高爱华; 葛少博; 肖相国

基于卷积神经网络的光学元件表面缺陷图像分类

1.
西安工业大学某某学院，陕西西安 710021
2.
西安应用光学研究所，陕西西安 710065

基金项目: 基金项目：陕西省科技厅重点研发项目资助（2019GY-063）

详细信息

作者简介:
侯劲尧（1995—），男，博士研究生，主要从事光电检测和图像处理方面的研究。E-mail：894608824@qq.com

通讯作者:
刘卫国（1964—），男，教授，博士生导师，主要从事光电子技术、电子材料技术研究。E-mail：wgliu@163.com

中图分类号: TP751
计量
- 文章访问数: 43
- HTML全文浏览量: 28
- PDF下载量: 9
- 被引次数: 0
出版历程
- 网络出版日期: 2022-11-17

Surface detection iamge of optical element surface defects based on convolutional neural network

1.
Xi'an Technological University, Xi'an 710021, China
2.
Xi'an Institute of Applied Optics, Xi'an 710065, China

摘要

摘要: 光学元件的表面疵病，即表面缺陷，其形状的大小会直接影响光学系统的性能，在对表面缺陷进行分类时，所面对的很多表面缺陷的形状都是不规则的，依靠普通的模式识别技术，分类很难达到预期的效果。为解决精密光学元件表面缺陷分类方法中精度低、耗时长的问题，提出了基于卷积神经网络的精密光学元件表面缺陷分类方法。采用散射法获取表面缺陷图像分析其成像特点，通过对图像进行旋转，镜像扩增了数据集，加强了网络的训练能力。使用AC训练网络模型，在不增加额外计算量的同时加强了网络的特征获取力。最后，通过Softmax分类器，将精密光学元件表面缺陷分为划痕、麻点及噪点3类。实验结果表明，所使用的模型对缺陷分类精度超过99.05%。
- 光学元件 /
- 表面缺陷 /
- 卷积神经网络 /
- 计算机视觉
Abstract: In the optical components, the surface defects are often caused by photoelectric action, photothermal action and plasma action, and the surface defects can directly affect the performance of the optical system. When classifying the surface defects, the shapes of many surface defects are irregular, so it is difficult to achieve the desired effect by relying on normal pattern recognition technology. To overvcome the low precision and long time consuming in traditional surface defect detection methods, a convolutional neural network based surface defect detection method is proposed in this paper. Firstly, the surface defect image is obtained by scattering method to analyze its imaging characteristics, and the training ability of the network is strengthened by rotating the image and mirroring the amplified data set. Furthermore, the AC training network model is used to strengthen the feature acquisition power of the network without increasing the extra calculation. Finally, the Softmax classifier is used to classify the surface defects into scratch, pitting and noise. Experimental results show that the defect detection accuracy of the model used is more than 99.05%.
- optical element /
- surface defect /
- convolutional neural network /
- computer vision

HTML全文

图 1 表面缺陷图像

Fig. 1 Surface defect image

下载: 全尺寸图片幻灯片

图 2 表面缺陷MER模型

Fig. 2 MER model of surface defects

下载: 全尺寸图片幻灯片

图 3 光学元件表面缺陷图像分类流程图

Fig. 3 Image classification flow chart of surface defects of optical components

下载: 全尺寸图片幻灯片

图 4 表面缺陷图像数据集的建立

Fig. 4 Establishment of surface defect image data set

下载: 全尺寸图片幻灯片

图 5 AlexNet网络模型

Fig. 5 AlexNet network model

下载: 全尺寸图片幻灯片

图 6 非对称卷积过程

Fig. 6 Asymmetric convolution process

下载: 全尺寸图片幻灯片

图 7 本文网络结构

Fig. 7 Network structure of this paper

下载: 全尺寸图片幻灯片

表 1 表面缺陷特征提取

Table 1 Extraction of surface defect features

表面缺陷	长轴/ pixel	短轴/ pixel	面积/ pixel²	外接矩形面积/pixel²	长宽比	矩形度P
1	1	1	4	1	1	4
2	1	1	4	1	1	4
11	3	1	8	3	3	2.67
12	2	2	8	4	1	2
13	3	2	10	6	1.5	1.67
16	8	6	48	50	1.33	0.96
17	9	8	56	72	1.13	0.78
18	11.26	7.73	69	87.07	1.46	0.79
19	10.29	9.84	78	101.20	1.05	0.77
20	32.65	31.75	88	104.28	1.03	0.84
26	87.74	28.87	1376	2 532.95	3.1	0.54

下载: 导出CSV

表 2 改进前后网络模型参数量的对比

Table 2 Comparison of the number of network model parameters before and after improvement

模型	模型总参数
Alexnet	58 271 811
Alexnet+AC	54 405 027

下载: 导出CSV

表 3 本文网络识别结果

Table 3 Network detection results in this paper %

表面缺陷	Scratch	Dig	Noisy
准确率	100	99.75	97.4

下载: 导出CSV

表 4 模型对比实验

Table 4 Model comparison experiments

Model	灵敏度/%	特异性/%	精确性/%	运行时间/s
VGG16	69.39	91.20	96.40	81.1
InceptionV3	71.74	92.29	97.12	79.35
ResNet50	73.87	94.17	97.32	78.56
AlexNet	77.20	92.34	97.20	18.54
Ours	89.97	96.73	99.05	18.46

下载: 导出CSV

参考文献(18)

[1]	RAVIKUMAR S, RAMACHANDRAN K, SUGUMARAN V. Machine learning approach for automated visual inspection of machine component[J]. Expert Systems with Applications,2011,38(4):3260-3266. doi: 10.1016/j.eswa.2010.09.012
[2]	MEHRAN P, DEMIRLI K, SURGENOR B. Fuzzy machine vision based clip detection[J]. Expert Systems,2013,30(4):352-366. doi: 10.1111/j.1468-0394.2012.00641.x
[3]	KRIZHEVSKY A, SUTSKEVER I, HINTON G. ImageNet classification with deep convolutional neural networks[J]. Advances in Neural Information Processing Systems,2012,60(6):1097-1165.
[4]	LIU Dong, YAGN Yongying, WANG Lin, et al. Microscopic scattering imaging measurement and digital evaluation system of defects for fine optical surface[J]Optics Communications, 2007, 278(2): 240-246.
[5]	王贵林, 朱俊辉, 李嘉祥, 等. 大口径光学元件表面疵病在位检测与评价研究[J]. 应用光学,2019,40(6):1167-1173. doi: 10.5768/JAO201940.0605005 WANG Guilin, ZHU Junhui, LI Jiaxiang, et al. In situ detection and evaluation of surface defects of large aperture optical components[J]. Applied Optics,2019,40(6):1167-1173. doi: 10.5768/JAO201940.0605005
[6]	JIANG Hongzhen, LIU Yong, LIU Xu, et al. Quantitative measurement for the structure of defects on optical surface with lensless Fourier transform digital holographic microscopy[J]. Optik-International Journal for Light and Electron Optics,2016,127(15):5865-5869. doi: 10.1016/j.ijleo.2016.04.012
[7]	WANG Kungjeng , HAO Fanjiang, LEE Yaxuan, A multiple-stage defect detection model by convolutional neural network[J]. Computers & Industrial Engineering, 2022, 168: 108096.
[8]	范勇, 陈念年, 高玲玲, 等. 大口径精密表面疵病的数字化检测系统[J]. 强激光与粒子束,2009,21(7):5. FAN Yong, CHEN Niannian, GAO Lingling, et al. Digital detection system for large-caliber precision surface defects[J]. Intense Laser and Particle Beam,2009,21(7):5.
[9]	RWAGASORE E M, ZHANG X, GAO K, et al. Convolutional neural network for sapphire ingots defect detection and classification[J]. Optical Materials,2021,119(9):111292.
[10]	谢世斌, 杨甬英, 刘东, 等. 精密表面疵病检测美标数字化评价实现算法[J]. 应用光学,2015,36(2):266-271. doi: 10.5768/JAO201536.0204005 XIE Shibin, YANG Yongying, LIU Dong, et al. Digital evaluation algorithm for precision surface defects detection[J]. Journal of Applied Optics,2015,36(2):266-271. doi: 10.5768/JAO201536.0204005
[11]	李璐, 杨甬英, 曹频, 等. 大口径光学元件表面灰尘与麻点自动判别[J]. 强激光与粒子束,2014,26(1):109-114. LI Lu, YANG Yongying, CAO Pin, et al. Automatic identification of dust and pitting on the surface of large aperture optical element[J]. High Power Laser and Particle Beams,2014,26(1):109-114.
[12]	DUAN Y, LIU S, HU C, et al. Automated defect classification in infrared thermography based on a neural network[J]. NDT & E International,2019,107:102147.
[13]	赵博, 史迎馨. 卷积神经网络的高精密光学元件表面缺陷检测[J]. 激光杂志,2021,42(11):5. ZHAO Bo, SHI Yingxin. Surface defect Detection for high precision optical components based on convolutional neural network[J]. Laser Journal,2021,42(11):5.
[14]	HE Y, HU C, LI H, et al. A flexible image processing technique for measuring bubble parameters based on a neural network[J]. Chemical Engineering Journal,2022,429:132-138.
[15]	KONSTANTINOS M, KONSTANTINOS K , ANASTASIOS D , et al. Deep supervised learning for hyperspectral data classification through convolutional neural networks[C]//IEEE Interna- tional Geoscience and Remote Sensing Symposium, Milan, 2015: 4959-4962.
[16]	BANDHU A, ROY S S. Classifying multi-category images u- sing deep learning: a convolutional neural network model[C]//IEEE International Conference on Recent Trends in Electronics, Information & Communication Technology, Ban- galore, 2017: 915-919.
[17]	马建, 韩卫光. 基于卷积神经网络的工件表面缺陷检测方法[J]. 组合机床与自动化加工技术,2021(1):106-109. MA Jian, HAN Weiguang. Workpiece surface defect detection method based on convolutional neural network[J]. Modular Machine Tool & Automatic Manufacturing Technology,2021(1):106-109.
[18]	DING X, GUO Y, DING G, et al. ACNet: strengthening the kernel skeletons for powerful CNN via asymmetric convolution blocks[J]. International Conference on Computer Vision,2019:1911-1920.