Wire bonding object detection through deep learning architecture of oriented bounding boxes with YOLOv8

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Wire bonding inspection is a critical process in semiconductor packaging to ensure product quality and reliability. This study compares you only look once version 5 (YOLOv5) axis-aligned bounding box (AABB) and YOLOv8 oriented bounding box (OBB) for wire bond defect detection. The dataset comprises 1,063 images with 14,778 annotated bonding instances in both AABB and OBB formats. Experimental results are evaluated using precision, recall, F1-score, mAP@0.5, mAP@0.5:0.95, and inference time. The findings show that YOLOv8 OBB achieved superior recall (0.9904) and mAP@0.5:0.95 (0.6787) compared to YOLOv5 AABB, demonstrating improved localization of rotated and dense wire bonds. Customized YOLOv8 variants with feature pyramid network (FPN)+efficient multi-scale attention (EMCA) and spatial squeeze and excitation (SpatialSE)+EMCA were also tested, with FPN+EMCA achieving the best balance of accuracy and robustness. This work highlights the potential of OBB-based object detection in semiconductor manufacturing and establishes a foundation for future real-time inspection systems.