基于机器视觉的裸片表面缺陷在线检测研究  

Research on Online Detection of Die Surface Defects Based on Machine Vision

作　　者：林佳 王海明[1] 郭强生[1] 刘晓斌[1] 周丹 LIN Jia;WANG Haiming;GUO Qiangsheng;LIU Xiaobin;ZHOU Dan(The 45th Research Institute of CETC, Beijing 100176, China)

机构地区：[1]中国电子科技集团公司第四十五研究所,北京100176

出　　处：《电子工业专用设备》2018年第2期13-16,45,共5页Equipment for Electronic Products Manufacturing

摘　　要：针对准确和实时检测裸片表面缺陷的需求,提出了一种基于线性判别分析(Linear Discriminant Analysis,LDA)和支持向量机(Support Vector Machine,SVM)的在线检测算法。首先,采用高斯滤波方法滤除裸片表面图像中的噪声;然后,提取裸片表面缺陷的Hu不变矩和方向梯度直方图(Histogram of Oriented Gradients,HOG)特征,采用LDA方法对特征进行降维;接着,在离线建模阶段,学习裸片表面正常模式的高斯混合模型(Gaussian Mixture Model,GMM),并学习各种缺陷模式的SVMs;最后,在线检测阶段使用GMM判断是否存在缺陷,使用K最近邻(K Nearest Neighbor,KNN)算法分类缺陷的模式。提出算法在采集的裸片数据库中得到了88.11%的检测准确率,单幅图像的平均检测时间为44.7 ms。实验结果表明,提出算法具有较高的检测准确性与实时性,可以应用到实际生产中的裸片表面缺陷在线检测。For accurate and real-time detection of die surface defects,an online detection algorithm based on Linear Discriminant Analysis（LDA） and Support Vector Machine（SVM） is proposed.Firstly,the Gaussian filtering method is improved to filter the noise in the die surface image. Next,the Hu invariant moments and histogram of oriented gradients（HOG） features of die surface defects are extracted,and the LDA method is adopted to reduce the feature dimension. Then,in the off-line modeling phase,the GMM of the normal die surface pattern and the SVMs of various defect patterns are constructed. Finally,in the on-line detection phase,the GMM is used to judge whether there are defects,and the patterns of the defects are classified by K Nearest Neighbor（KNN） algorithm. The detection accuracy of the proposed algorithm is 88.1 1% in the captured die database. The average detection time of single image is 44.7 ms. The experimental results show that the proposed algorithm has high detection accuracy and is provided with real-time performance. It can be applied to the on-line detection of die surface defects in the actual production.

关 键 词：晶圆划片 裸片表面缺陷检测 表面特征 线性判别分析 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]