Publication: Localization and detection of solder mask peel off defects on printed circuit board using pre trained convolutional neural networks
Loading...
Date
2024-07
Authors
Seng, Yi Wen
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The Solder Mask Peel Off (SMPO) is categorized as a cosmetic defect found on the Printed Circuit Board (PCB) surface that might cause functional defects. This
research aims to improve and develop the SMPO defects detection and localization system by using You-Only-Look-Once (YOLO), a type of Convolutional Neural
Network (CNN). The significance of this research focuses on revolutionizing the PCB inspection systems by advancing the accuracy and efficiency of SMPO defect
detection. This ensures a more robust quality assurance process for PCB functionality check, consequently improves the overall product reliability. The defect and non defect images provided by Western Digital Sandisk Malaysia were cropped using the YOLO detection algorithms to extract the Region of Interest (RoI) and the defect area was then annotated in Roboflow. Each image was subsequently tiled into 16 sub images and augmented to enhance the diversity of the dataset. The dataset in Roboflow was exported to YOLO format versions to be used in YOLO models training in Google Colab. The performance of YOLOv5-segmentation and YOLOv8-segmentation were trained, validated, and compared. The best YOLO model variant and version will undergo hyperparameter tuning. As such, the trained YOLOv8 medium model had the best performance with 96.9% precision, 94.4% recall, 96.6% mAP50 for bounding box and 98.3% precision, 95.8% recall, 97.4% mAP50 for segmentation mask for SMPO defect detection with AdamW optimizer, 0.0001 learning rate, 16 batch size, 100 epochs, and SiLU activation function. The SMPO defect detection system developed has an average inspection speed of 1 second per image.