基于图像处理和机器学习的纽扣电池外观缺陷分类方法  

A Method for the Classification of Appearance Defects in Button Batteries Based on Image Processing and Machine Learning

作　　者：林木泉 江月松 杜毅涛 LIN Muquan;JIANG Yuesong;DU Yitao(Key Laboratory of Industrial Automation Control Technology and Application,Minnan University of Science and Technology,ShiShi Fujian 362700,China;School of Mechatronics and Vehicle Engineering,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]闽南理工学院工业自动化控制技术与应用福建省高校重点实验室,福建石狮362700 [2]华东交通大学机电与车辆工程学院,南昌330013

出　　处：《德州学院学报》2024年第2期9-14,共6页Journal of Dezhou University

摘　　要：针对纽扣电池生产过程中出现的划痕、脏污、凹陷、飞边等外观缺陷,提出基于数字图像处理方法,结合机器学习的方法实现外观缺陷特征的分类。不同的缺陷特征提取难易度不同,针对负极缺陷样本,提取Hog特征数据后经SVM分类算法实现外观缺陷分类。使用卷积神经网络分类模型对正极缺陷进行特征的提取和分类,通过多个经典的分类模型进行对比实验。而缺陷类型中少样本的飞边缺陷则使用了求解极坐标变换后的斜率和孪生神经网络的方法实现识别,在不扩增数据的情况下,达到了较好的分类效果。Addressing the appearance defects such as scratches,dirt,indentations,and missing edges during the production of button batteries,this study proposes a method based on digital image processing techniques,combined with machine learning for the classification of appearance defects.The difficulty of extracting different appearance defect features varies,and for negative electrode appearance defect samples,the Hog feature data is extracted after SVM classification algorithm implementation to achieve negative electrode appearance defect classification.The convolutional neural network classification model is used to extract and classify the appearance defect features of the positive electrode,and a series of classic classification models are used for comparative experiments.Meanwhile,the lack of sample data in some types of appearance defects,such as protrusions,is solved by using the slope and twin neural network after solving the polar coordinate transformation,achieving a good classification effect without data expansion.

关 键 词：外观缺陷 机器学习 SVM 卷积神经网络 孪生神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]