基于多尺度数学形态学的机械零件表面缺陷智能检测方法

摘要:

机械零件表面缺陷形态多样,,不同缺陷类型之间的特征差异较小,在检测中无法有效建立特征与缺陷类型之间的映射关系,导致检测很容易出现漏检和误检的问题。因此,研究一种基于多尺度数学形态学的机械零件表面缺陷智能检测方法。通过CCD机器视觉成像系统获取机械零件表面数字图像并开展几何校正、图像灰度变换、对比度拉伸处理。针对机械零件表面数字图像,考虑不同缺陷类型之间的特征差异较小,利用高斯金字塔进行多尺度分解,通过选取的多尺度结构元素对多尺度图像进行形态学的开运算和闭运算,得到多尺度数学形态学处理后的重构图像,以从不同尺度分析图像,提取多层次的特征信息。针对重构图像,进行Tamura纹理和几何参数计算,得到多维度特征参数,以准确地捕捉零件表面的特征。利用模拟退火算法优化径向基函数神经网络,提高非线性映射能力和分类能力,以此输入多维度特征参数,建立特征参数与缺陷类型之间的映射关系,得出缺陷类型的预测概率,选出最大概率值对应的类别作为最终的缺陷检测结果。结果表明:所研究方法的ROC曲线在图中更靠近左上角且AUC为0.868,更接近1,F1分数一直保持在0.8以上,说明所研究检测方法的误检和漏检情况较少,具有较高的准确性和可靠性。

Abstract:

The surface defects of mechanical parts have diverse forms, and the differences in features between different defect types are small. Therefore, it is difficult to effectively establish a mapping relationship between features and defect types in detection, which can easily lead to missed or false detections. Therefore, a multi-scale mathematical morphology based intelligent detection method for surface defects of mechanical parts is studied. Obtain digital images of mechanical parts surfaces through CCD machine vision imaging system and perform geometric correction, image grayscale transformation, and contrast stretching processing. For digital images of mechanical parts surfaces, considering the small differences in features between different defect types, a Gaussian pyramid is used for multi-scale decomposition. The selected multi-scale structural elements are subjected to morphological opening and closing operations on the multi-scale image to obtain a reconstructed image after multi-scale mathematical morphology processing. The image can be analyzed from different scales to extract multi-level feature information. For the reconstructed image, Tamura texture and geometric parameter calculations are performed to obtain multi-dimensional feature parameters, in order to accurately capture the surface features of the part. Using simulated annealing algorithm to optimize radial basis function neural network, improve nonlinear mapping ability and classification ability, input multi-dimensional feature parameters, establish the mapping relationship between feature parameters and defect types, obtain the prediction probability of defect types, and select the category corresponding to the maximum probability value as the final defect detection result. The results showed that the ROC curve of the studied method was closer to the upper left corner of the graph with an AUC of 0.868, which was closer to 1. The F1 score remained above 0.8, indicating that the studied detection method had fewer false positives and false negatives, and had high accuracy and reliability.