基于卷积神经网络的PCB板通用分拣算法

doi:10.16180/j.cnki.issn1007-7820.2020.02.011

摘要/Abstract

摘要：

传统PCB板分拣算法特征点由人工提取,位姿计算方法单一,且存在预处理步骤繁琐、样本要求高、位姿精度低、适用性低等缺点。为解决上述问题,文中提出一种基于卷积神经网络的PCB板通用分拣算法。改进的CaffeNet网络模型自动学习大量图片数据中PCB板的深层特征,并完成PCB板自动识别分类。运用改良RANSACS算法提高ORB算法特征点匹配正确率,采用最小二乘法计算模板图片与待匹配图片间角度差,以实现各类PCB板快速定位。实验证明,算法分拣正确率达到99.35%,具有良好的准确率和分拣效率。

关键词: 卷积神经网络, PCB板, 特征提取, CaffeNet, ORB算法, RANSACS算法, 最小二乘法

Abstract:

The feature points of the traditional PCB board sorting algorithm are manually extracted, and the pose calculation method is single, which has the disadvantages of complex pretreatment steps, high sample requirements, low pose accuracy and low applicability. To solve the above problems, a general sorting algorithm for PCB board based on convolution neural network was proposed in this paper. An improved CaffeNet network model was constructed to automatically learn the deep features of PCB boards in a large number of image data, completed the automatic recognition and classification of PCB boards. The modified RANSACS algorithm was used to improve the accuracy of feature point matching of ORB algorithm, and the least square method was conducted to calculate the angle difference between the image of the board and the image to be matched, so as to realize the rapid positioning of all kinds of PCB boards. The result showed that the sorting accuracy of the algorithm reached 99.35%, which had good accuracy and sorting efficiency.

Key words: CNN, PCBboard, featureextraction, CaffeNet, ORBalgorithm, RANSACSalgorithm, least squares

中图分类号:

TP312

王正军,姚一鸣,陈龙. 基于卷积神经网络的PCB板通用分拣算法[J]. 电子科技, 2020, 33(2): 60-65.

WANG Zhengjun,YAO Yiming,CHEN Long. PCB Board Common Sorting Algorithm Based on Convolutional Neural Network[J]. Electronic Science and Technology, 2020, 33(2): 60-65.

图/表 12

图1

表1

图2

图3

图4

图5

表2

图6

图7

图8

图9.

表3

参考文献 17

[1]	张桂梅, 张松, 储珺 . 一种新的基于局部轮廓特征的目标检测方法[J]. 自动化学报, 2014,40(1):2346-2355.
	Zhang Guimei, Zhang Song, Chu Jun . A new object detection algorithm using local contour features[J]. Acta Automatica Sinica, 2014,40(1):2346-2355.
[2]	王民, 王静, 张立材 , 等. 基于混合色彩空间的分块颜色特征提取算法[J]. 激光与光电子学进展, 2018,55(1):258-264.
	Wang Min, Wang Jing, Zhang Licai , et al. Block color feature extraction algorithm based on mixed color space[J]. Laser & Optoelectronics Progress, 2018,55(1):258-264.
[3]	陈洪, 李进强 . 改进的自适应SUSAN角点特征提取方法[J]. 测绘科学, 2017,42(12):118-121,126.
	Chen Hong, Li Jinqiang . An improved self-adapting corner feature extraction method[J]. Science of Surveying and Mapping, 2017,42(12):118-121,126.
[4]	David G Lowe . Distinctive image features from scale-invariant keypoints[J]. International Journal of Computer Vsion, 2004,60(2):91-110.
[5]	吴晓军, 邹广华 . 基于边缘几何特征的高性能模板匹配算法[J]. 仪器仪表学报, 2013(7):1462-1469.
	Wu Xiaojun, Zou Guanghua . High performance template matching algorithm based on edge geometric features[J]. Chinese Journal of Scientific Instrument, 2013(7):1462-1469.
[6]	秦海报 . 基于模板匹配的快速少纹理目标识别算法[J]. 电子科技, 2019,32(1):47-51.
	Qin Haibao . Texture-less object detection algorithm with speed-up based on template matching[J]. Electronic Science and Technology, 2019,32(1):47-51.
[7]	贾旭, 薛定宇, 崔建江 . 基于尺度、距离、旋转测度的角点匹配算法[J]. 东北大学学报(自然科学版), 2010(12):1689-1692.
	Jia Xu, Xue Dingyu, Cui Jianjiang . Corners matching algorithm based on measures of scale, distance and rotation[J]. Journal of Northeastern University(Natural Science Edition), 2010(12):1689-1692.
[8]	王丽君, 于莲芝 . 基于卷积神经网络的位置识别[J]. 电子科技, 2017,30(1):104-106.
	Wang Lijun, Yu Lianzhi . Location recognition based on convolution neural network[J]. Electronic Science and Technology, 2017,30(1):104-106.
[9]	Li D, Dong Y . Deep learning methods and applications[J]. Foundat0ns & Trends in Signal Processing, 2013,7(3):197-387.
[10]	Sun Y, Chen Y, Wang X . Deep learning face representation by join identification-verification[J]. International Conference on Neural Information Processing, 2014,27(17):1988-1996.
[11]	Le Q, Naiam J, Coates A , et al. On optimization methods for deep learning[J]. International Conference on Machine Learning, 2011,7(3):67-105.
[12]	Schmidhuber J . Deep learning in neural networks:an overview[J]. Neural Networks, 2014,61(12):85-117.
[13]	Jia Yangqing, Shelhamer E, Donahue J , et al. Caffe: convolutional architecturefor fast feature embedding[C]. Orlando:Proceedings of the ACM Conference on Multimedia, 2014.
[14]	Rublee E, Rabaud V, Konolige K , et al. ORB:An efficient alternative to SIFT or SURF[C]. Barcelona:IEEE International Conference on Computer Vision, 2012.
[15]	孙灏, 高俊强, 许苏苏 . 基于SURF算法和改进RANSAC算法的无人机影像匹配[J]. 测绘工程, 2017,26(11):55-59,64.
	Sun Hao, Gao Junqiang, Xu Susu . UAV image matching based on SURF algorithm and improved RANSAC algorithm[J]. Engineering of Surveying and Mapping, 2017,26(11):55-59,64.
[16]	Lin Hongwei, Cao Qi, Zhang Xiaoting . The convergence of least-squares progressive iterative approximation for singular least-squares fitting system[J]. Journal of Systems Science & Complexity, 2018,31(6):1618-1632.
[17]	Cao Yongxiu, Yu Jichang . Partial least squares method for treatment effect in observational studies with censored outcomes[J]. Wuhan University Journal of Natural Sciences, 2018,23(6):487-492.

网络层	经典模型	改进的模型
FC₆	4 096	512
FC₇	4 096	512
FC₈	1 000	4~6

角度计算参数值	图5(a)	图5(b)
M	-718 600	293 557.333
N	-1 309 137.5	252 570.667
tanα	0.548 911 02	1.162 278
α	208.76°	49.29°

编号	改进ORB的算法	最小矩形框算法	差值
1	49.29°	49.24°	0.05°
2	89.22°	89.26°	-0.04°
3	13.06°	13.00°	0.06°
4	94.49°	94.34°	0.15°
5	123.02°	123.13°	-0.11°
6	55.84°	55.77°	0.07°
7	245.43°	245.64°	-0.19°
8	263.31°	263.11°	0.20°
9	298.17°	297.99°	0.18°
10	341.63°	341.45°	0.19°