Automate Industrial Printed Circuit Board Defect Detection: A Two Stage Approach
Year: 2026
Type: conference-paper
Abstract: Printed circuit boards (PCBs) are the major workhorses of many electronic products. With the growing demand for large-scale electronics manufacturing, inspection systems must deliver high accuracy at low false-positive rates without sacrificing throughput. Previous approaches to this problem include one-stage image processing techniques and machine learning models. However, these methods facing limitations under various lighting conditions and requirements of huge ground truth labeled datasets. To tackle this issue, we study the PCB solder inspection task as a two-stage vision problem: localize each solder joint and then classify whether it is good or defective. Our aim is to identify a deployable model that is both accurate, robust, and stable in the industrial setting. In the bounding box detection task, we employ the YOLOv8 model to localize the solder joints on the PCBs. For the defect detection task, we compare (i) a classical image-processing pipeline, (ii) a shallow SVM baseline, and (iii) CNN architectures. We compare the models on real images obtained from manufacturing lines. Extensive experiments show that simple Convolutional Neural Network architectures can achieve up to 99.4% accuracy, indicating that classical Deep Learning models are capable of automating PCB inspection at an industrial scale.
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203664526
