A physics-informed framework for feature extraction and defect segmentation in pulsed infrared thermography

This paper presents a robust and interpretable methodology for defect detection in active infrared thermography data applied to polyvinyl chloride (PVC) specimens. Our approach integrates a physics-based cooling model to describe the transient thermal response of each pixel, from which five primary temporal features are extracted via least-squares fitting. These features are then enriched with local spatial statistics through neighborhood-based computations, resulting in a 15-dimensional descriptor per pixel. The resulting feature set is used to train a random forest classifier, which achieves high overall accuracy (99.3%), competitive intersection-over-union (0.705), and an outstanding ROC AUC (0.998). In contrast to deep encoder–decoder networks that require extensive computational resources and large annotated datasets, the proposed pipeline offers enhanced interpretability and significantly reduced computational overhead. Comparative analysis with state-of-the-art deep learning models, such as those reported in Wei et al., (2023), demonstrates that our approach achieves similar performance while providing a transparent insight into the contribution of each feature. The proposed method is especially suitable for engineering failure analysis where model transparency, rapid evaluation, and integration into existing inspection protocols are critical. Future work will extend the framework to accommodate a broader range of defect types and material systems, aiming to further enhance industrial applicability and diagnostic reliability. • Physics-based cooling model extracts five temporal thermographic features. • Spatial augmentation creates a 15-dimensional pixel-level descriptor. • Random forest classifier achieves near-perfect ROC AUC of 0.998 accuracy. • Enhanced interpretability with lower overhead than deep learning models. • Ideal for transparent, rapid, cost-effective engineering defect analysis.