Evaluation of Lightweight Object Detection and Instance Segmentation Techniques for Industrial Printed Circuit Board Solder Quality Assessment

Authors : Tuncay Soylu, Emel Soylu

Pages : 128-138

Doi:10.18466/cbayarfbe.1669378

View : 17 | Download : 67

Publication Date : 2025-12-29

Article Type : Research Paper

Abstract :The rapid progress of deep learning has transformed the detection of soldering defects in printed circuit boards (PCBs), outperforming traditional manual inspections and rule-based machine vision systems.This study evaluates the performance of state-of-the-art YOLO models—specifically YOLOv11 and YOLOv12—for both object detection and instance segmentation of solder defects in aerospace PCBs using the open-source SolDef_AI dataset. We compare multiple variants (nano, small, and medium) to assess their accuracy, efficiency, and suitability for industrial quality control. Our experiments reveal that YOLO-v11s-seg achieves the highest mean average precision (mAP50-95: 0.853 for detection, 0.822 for segmentation), demonstrating superior defect localization capabilities, particularly for challenging classes such as \\\"poor_solder\\\" and \\\"spike.\\\" While YOLOv12 models exhibit competitive detection performance, they show slightly lower segmentation accuracy, indicating potential areas for architectural refinement. Smaller models (YOLO-v11n, YOLO-v12n) offer a favorable balance between speed and precision, making them viable for real-time applications. The findings highlight the effectiveness of YOLO-based deep learning in automating solder defect inspection, with implications for improving manufacturing quality assurance in electronics production.
Keywords : Image Processing, Object Detection, Quality Control, Soldering Defect Detection

ORIGINAL ARTICLE URL