This study aimed to apply deep learning, using the YOLOv5 object detection method, to automate defect detection in internal threads of fasteners, replacing traditional manual inspection methods that are inadequate for meeting the high-quality standards in Taiwan's fastener industry.
The study used the YOLOv5 model to detect and classify internal thread defects, focusing on the impact of annotation methods, data augmentation and transfer learning. It compared three YOLOv5 models (YOLOv5s, YOLOv5m and YOLOv5l) with varying parameter sizes to identify the best model for defect detection.
The study found that burrs and granules shared similar features, causing confusion and that bounding box labeling and defect segmentation methods significantly affected model performance. YOLOv5m outperformed other models in terms of mAP and data augmentation improved model accuracy. Pre-trained weights helped the model converge faster, but freezing initial layers slightly reduced performance.
The study faced challenges in distinguishing similar defects, particularly burrs and granules and in generalizing the model across data from multiple manufacturers. Additionally, the bounding box annotation method and fine segmentation caused convergence issues.
This research provides practical solutions for automating internal thread defect detection in fastener manufacturing, offering insights into model selection, data augmentation and annotation strategies for more accurate and efficient defect detection.
The study introduces a novel application of YOLOv5 for fastener defect detection, offering new insights into data augmentation, transfer learning, and annotation methods and provides a comparative analysis of model performance in an industrial context.
