一种提高晶粒表面缺陷检测准确性的方法研究  

Research on a Method to Improve the Detection Accuracy of Wafer Die Surface Defects

作　　者：赵春东[1] 朱奎锋 苏艳杰 陈俭永 张东洋 陈晓艳[1] ZHAO Chundong;ZHU Kuifeng;SU Yanjie;CHEN Jianyong;ZHANG Dongyang;CHEN Xiaoyan(College of Information and Automation,Tianjin University of Science and Technology,Tianjin 300222,China;Tianjin FLY Tech Company,Ltd.,Tianjin 300350,China)

机构地区：[1]天津科技大学电子信息与自动化学院,天津300222 [2]天津福莱迪科技发展有限公司,天津300350

出　　处：《传感技术学报》2020年第7期1013-1019,共7页Chinese Journal of Sensors and Actuators

基　　金：天津市重点研发计划科技支撑重点项目(18YFZCGX00360)。

摘　　要：晶粒质量影响半导体芯片性能,针对晶粒表面缺陷检测的问题,本文提出了一种新的缺陷检测方法,能有效抑制图像噪声,提高缺陷检测准确性。首先,采用形态学闭合重建方法去除噪声干扰。然后,将重建图像灰度级别作为聚类样本,构建一个基于离差阵的K-means分割算法准则函数,不断更新聚类中心直至准则函数收敛,完成图像分割。最后,提取分割晶粒轮廓信息,根据轮廓间的差异性,利用轮廓相似度识别缺陷晶粒。对2937张晶粒图像进行处理并测试,结果表明:该方法能有效滤除了图像噪声,更准确更快速地进行图像分割,对良品检测准确率为100%,对三种典型表面缺陷的平均检测准确率为99.92%。The quality of the wafer die affects the performance of the semiconductor chip.In view of the problem of defect detection on the surface of the die,this paper proposes a new defect detection method that can effectively filter the image noises and improve the detection accuracy of defects.Firstly,a morphological closed reconstruction method is used to filter noises.Secondly,taking the gray levels of the reconstructed image as clustering objects,a clustering criterion of the K-means segmentation algorithm based on dispersion matrix is constructed,and the clustering center is continuously updated until it converges to complete the image segmentation.Finally,the contour of the segmented die is extracted to calculate the contour similarity index depends on the differences of the contours,which is used to identify the defective grains.Through 2937 die images processing and testing,it shows that the proposed method is of strong robustness to different noises,the segmentation is faster with high accuracy,and the accuracy of good die reaches 100%while the defect detection 99.92%.

关 键 词：计算机视觉 形态学闭合重建 聚类 轮廓信息 缺陷检测 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]