Deep learning-based defect detection in pulsed thermography

The current era emphasizes system productivity to meet global demand. The productivity index is closely tied to the reliability of the manufacturing process. To keep up with today’s manufacturing demands, inspection systems on the production line must prioritize both speed and quality. However, a key challenge in automated inspection is achieving a balance between high defect detection accuracy and minimizing false positives and false negatives. To address this, this study investigates the effectiveness of deep learning-based models for defect detection in Pulsed Thermography (PT) using a publicly available dataset of PVC specimens. Pulsed Phase Thermography (PPT) was applied to the raw thermograms to generate phase images and evaluate the performance of conventional methods. Two models were trained and evaluated for defect detection: a pre-trained YOLOv8 object detection model and a semantic segmentation model from Halcon. The YOLOv8 model demonstrated a high precision of 97.1%, but with a recall of 82%, indicating that it accurately detected defects but missed some. In contrast, the Halcon model achieved perfect recall (100%) but lower precision (78.2%), suggesting that it detected all defects but also introduced a significant number of false positives. The results highlight the trade-offs between precision and recall in these models, with YOLOv8 focusing on accuracy and Halcon on comprehensive defect detection. This study demonstrates the potential of deep learning techniques in enhancing defect detection performance in Pulsed Thermography applications.