| Title |
Real-time TSV 3D Shape Defect Inspection System Using Deep Learning based Fast Object Detection |
| Authors |
(Kyeong Beom Park) ; (Jae Yeol Lee) ; (Harim Lee) |
| DOI |
https://doi.org/10.5573/JSTS.2025.25.6.645 |
| Keywords |
TSV; deep learning; object detection; YOLO; inspection system |
| Abstract |
Conventional inspection methods for Through-Silicon Via (TSV) 3D shape defect detection, such as Scanning Electron Microscopy (SEM) and X-ray inspection, have been widely used due to their precision in structural analysis. However, these methods suffer from significant limitations including high equipment cost, long inspection time, and the inability to operate in real-time. Moreover, SEM is inherently a destructive technique, while X-ray imaging lacks sufficient resolution to detect nanoscale shape anomalies or polymer residues. These drawbacks hinder the implementation of fast and scalable inspection systems, which are increasingly demanded in modern semiconductor manufacturing, especially for high-density 3D integration. Therefore, a new approach is urgently required?one that ensures high detection accuracy while also being non-destructive, fast, and suitable for realtime inspection in practical production environments. In this paper, we develop a real-time TSV 3D shape defect inspection system implemented with a deep learning-based object detection method. For the real-time operation of the proposed system, YOLOv8 and YOLOv10 are utilized because the YOLO family of networks can guarantee fast inference performance as well as excellent detection performance. The YOLOv8 and YOLOv10 have intrinsic differences in the network architecture, such as anchor-free detection structure and NMS-free training and a dualhead structure, resulting in different inference and detection performance. Therefore, based on the performance comparison of the two networks, the appropriate model should be selected according to the specific needs for either faster inference or higher detection accuracy. In addition, for more reliable training of object detection networks, we collect 3D point cloud data containing TSV normal and defective pattern data created on real 8-inch silicon wafers.
By obtaining the datasets from real silicon wafers, we can ensure the reliability and the practical applicability of the trained network performance. For the performance comparison, we utilize several performance metrics, which are processing time, precision, F1 score, and Fβ score. Finally, extensive evaluations confirm that the YOLOv8-l model achieves the highest precision (0.99997) and F1 score (0.99989), while the YOLOv10-n model exhibited the fastest processing time (0.18601 seconds) and the highest Fβ score (0.92565). |