The shift towards high-speed, automated optical inspection in electronics manufacturing necessitates object detection models that optimally balance accuracy with inference speed. This study presents a controlled, like-for-like performance evaluation of two successive models from the YOLO for the detection of six common surface defects on printed circuit boards (PCBs) using the PKU synthetic dataset. Under identical training protocols, hyperparameters, and data splits, both models achieved strong overall accuracy. YOLOv10 attained a mean Average Precision (mAP@50) of 0. 893 and a stricter localization accuracy (mAP@50-95) of 0. 460, slightly outperforming YOLOv8 attained 0. 889 and 0. 451, respectively. This marginal gain in accuracy, particularly at higher IoU thresholds, is contextualized within the established performance trade-off framework for real-time detectors, where architectural efficiency gains must be weighed against potential gains in precision. A class-wise analysis revealed that "missingₕole" defects were detected with the highest reliability, while classes like "mousebite" and "spur" presented challenges, primarily in recall. The results underscore YOLOv10's architectural improvements for efficient, high-precision detection. The discussion extends to the implications of these findings for industrial deployment, where the choice between model versions hinges on the specific production line's tolerance for false positives versus the imperative for millisecond-level latency. Future work should focus on recall improvement for minority defect classes and explicit measurement of the accuracy-speed Pareto frontier on target hardware to fully inform model selection for in-line AOI systems.
Teke et al. (Mon,) studied this question.