基于锚点集优化的硅片图像缺陷检测方法

Defect detection method of silicon wafer image based on anchor-set optimization

为有效解决机器视觉算法识别硅片图像缺陷时存在检测速度慢、准确率低和稳定性差等问题,提出一种锚点集的优化方法。将Faster R-CNN检测网络与ResNet相结合进行特征提取;针对硅片缺陷数据集的特点,从理论上通过期望的定位精度推导一种优化的锚点集;利用优化后的锚点集训练缺陷检测模型。实验结果表明,所提检测模型识别精度提升至94.1%,检测时间达到160 ms,满足工业检测的需求。

To solve the problems of low detection speed,low accuracy and poor stability in the recognition of silicon wafer image defects using machine vision algorithm effectively,an anchor-set optimization method was proposed.Feature extraction was carried out by combining Faster R-CNN detection network with ResNet.Focusing on the characteristics of silicon wafer defect data sets,an optimized anchor-set was deduced theoretically according to the expected location accuracy.The optimized anchor-set was used to train the defect detection model.Experimental results show that the accuracy of the proposed detection model is improved to 94.1%,while the detection time is 160 ms,which meets the requirements of industrial detection.