Journal of the Microelectronics and Packaging Society

As high-performance semiconductor packaging technologies such as 2.5D and 3D packaging along with the advancement of artificial intelligence (AI) have emerged with more complex packaging designs, the number of areas that need to be inspected in the package continues to increase. As a result, a rule-based defect inspection system that defines inspection areas and sets thresholds manually is time-consuming and error-prone. To solve this problem, this study proposes a method to automatically extract the region of interest (ROI) from actual quad flat no-lead (QFN) and ball grid array (BGA) package images using deep learning models. In this study, we analyzed the effect of the amount on the model performance using the YOLOv8, YOLOv9, YOLOv10, and YOLOv11 models, which are commonly used in real-time object detection, and showed that the performance of automatic detailed ROI extraction can improve with small datasets through data augmentation and preprocessing techniques. In addition, it was proved that the deep learning model can detect important elements in semiconductor packages well with high accuracy by considering various lighting change conditions in the industrial site. This study will be used as important basic data to improve the automation and efficiency of the semiconductor package inspection system.

Keywords Semiconductor package inspection, Defect detection, ROI extraction, Deep learning, Machine vision