基于YOLOv8的晶圆背损伤缺陷自动识别方法
Automatic Identification of Wafer Back Damage Defects Based on YOLOv8

在半导体制造过程中，针对晶圆背损伤缺陷目标较小且粘连程度过高导致识别困难的问题，本研究在提取粘连目标轮廓包围盒的基础上，结合深度学习方法，提出了一种基于YOLOv8的晶圆背损伤缺陷自动识别方法。首先对采集图片进行图像预处理提取出所有感兴趣目标的连通域，结合连通域轮廓的包围盒信息将原始图片切片为局部图片，为缺陷目标准确计数奠定基础；其次采用迁移学习的策略，将局部图片数据集输入带有预训练权重的YOLOv8目标检测网络，采用输出的模型文件对局部图片中的目标识别；最终结合包围盒的信息对局部图片中的识别点位补偿，从而实现对于原始图片中的晶圆背损伤缺陷识别定位。实验结果表明，该方法识别准确率能够达到96.81%，与传统算法相比具有更好的识别精度。
In the semiconductor manufacturing process, for the wafer back damage defects target is small and the degree of adhesion is too high resulting in recognition difficulties, this study in the extraction of adhesion target contour enclosing box, combined with the deep learning method, proposed a YOLOv8 based wafer back damage defects automatic identification method. Firstly, image preprocessing is performed on the captured images to extract the connectivity domains of all the targets of interest, and the original images are sliced into localized images by combining the information of the enclosing boxes of the contours of the connectivity domains, which lays the foundation for the accurate counting of defective targets; secondly, the localized image dataset is inputted into the YOLOv8 target detection network with the pre-training weights by adopting the migration learning strategy, and the output model files are used to recognize the targets in the localized images. The output model file is used to identify the target in the local picture; finally, the information of the enclosing box is combined with the compensation of the identification points in the local picture, so as to realize the identification and localization of the wafer back damage defects in the original picture. Experimental results show that the recognition accuracy of this method can reach 96.81%, which is better than the traditional algorithm.