基于机器视觉的车身缺陷识别与分类方法

Vehicle Body Defect Detection and Classification Based on Machine Vision

针对目前国内外汽车车身涂装工艺后缺陷识别自动化程度低,难以进行非接触高精度检测等问题,提出了一种结合二维视觉、三维视觉进行缺陷识别,通过卷积神经网络进行分类的方法。首先基于最大类间方差法和特征提取算法实现缺陷二维坐标的确定,算法受光照影响较小;其次基于卷积神经网络,实现对黑胶、划痕、凸点等三种常见缺陷类型的分类;然后基于RANSC聚类算法及PCA主成分分析,实现对车身平面拟合及缺陷法向量信息的确定;最后搭建了实验系统,通过双机器人标定及三维手眼标定实现对基准坐标系的坐标转换,实现对缺陷几何中心坐标、法向量、类型等信息的确定,系统的平均误差远小于现有打磨设备的尺寸。实验结果表面本方案可为后续自动化打磨设备提供工艺优化及打磨处理的数据输入。

The insufficient automation in vehicle body defect detection after the coating process makes it difficult to carry out non-contact high-precision detection.To solve this problem,this paper establishes a defect detection and classification system,which employs the two-dimensional vision and three-dimensional vision for identification,and the convolution neural network for classification.Firstly,the two-dimensional coordinates of defects are determined based on the Otsu method and the feature extraction algorithm which is less affected by light conditions.Secondly,based on the convolution neural network,the classification of three common defect types,i.e.,black glue,scratches,and bumps,is achieved.Thirdly,through the RANSC clustering algorithm and the principal component analysis(PCA),the body plane is fitted and the normal vector of the defect is obtained.Finally,the experimental system is built.The coordinate transformation of the reference coordinate system is realized,and the geometric center coordinates,normal vectors and defect types are determined through the dual robot calibration and 3 D hand-eye calibration.The system achieves an average error far smaller than the size of the existing grinding equipment.The experimental results show that the proposed system can provide data input of process optimization and grinding treatment for subsequent automatic grinding equipment.