Instance segmentation for PCB defect detection with Detectron2

Abstract

Printed circuit boards (PCBs) are essential in modern electronics, where even minor defects can lead to failures. Traditional inspection methods struggle with complex PCB designs, necessitating automated deep learning techniques. Object detection models like Faster R-CNN and YOLO rely on bounding boxes for defect localization but face overlap issues, limiting precise defect isolation. This paper presents a segmentation-based PCB defect detection model using Detectron2’s Mask R-CNN. By leveraging instance segmentation, the model enables pixel-level defect localization and classification, addressing challenges such as shape variations, complex structures, and occlusions. Trained on a dataset of 690 COCO-annotated images, the model underwent rigorous experimentation and parameter tuning. Evaluation metrics, including loss functions and mean average precision (mAP), assessed performance. Results showed a steady decline in loss values and high precision for defects like mouse bites and missing holes. However, performance was lower for complex defects like spurs and spurious copper. This study highlights the effectiveness of instance segmentation in PCB defect detection, contributing to improved quality control and manufacturing automation.