Speaker
Description
In non-destructive testing (NDT), Flying-spot thermography has established itself as a reliable method for detecting surface-breaking cracks through local laser heating and infrared scanning. While effective, the automated analysis of these thermal sequences remains a significant challenge. The thermal signature of a crack—typically a high-frequency discontinuity in the thermal field—can be visually ambiguous, often resembling surface artifacts such as emissivity variations, non-planar geometries, or sensor noise. Consequently, the state of the art in automated thermographic analysis has shifted toward deep learning techniques, and predominantly Convolutional Neural Networks (CNNs) architecture, which are typically trained to detect and localize anomalies using large, task-specific annotated datasets.
However, standard CNN-based methods have some critical limitations in industrial NDT scenario. In a difficult task like segmentation, where the goal is to classify every pixel from an image, they are prone to overfit to local texture statistics rather than learning the semantic structure of a crack. This tendency results in poor generalization when applied to new experimental setups (e.g., different cameras or material surfaces). Furthermore, these models suffer from “catastrophic forgetting”: when fine-tuned on a new domain, they rapidly lose the ability to detect defects in the original domain, rendering them unsuitable for changing industrial environments, such as changes in sensing hardware, inspected materials, or inspection modalities.
In this work, we investigate the use of Foundation Models (FMs) for image segmentation across multiple thermal imaging scenarios. Foundation Models are deep neural networks pretrained on very large scale and highly diverse datasets, and designed to extract rich, high-level representations from images. Unlike conventional CNNs, which are typically pretrained on comparatively much smaller datasets (e.g., ImageNet) and which rely on local texture and contrast cues, FMs focus on global shape and structural information, making them inherently more robust to the high-frequency noise present in thermal imagery. We propose a segmentation framework in which a pretrained FM is used as the encoder and coupled with a lightweight decoder to generate segmentation masks. This approach is systematically compared to a standard CNN-based segmentation baseline across four evaluation settings: (i) in-domain performance, (ii) zero-shot generalization to unseen domains, (iii) transfer learning to new domains, and (iv) robustness to catastrophic forgetting after domain adaptation.
The proposed framework is evaluated on two flying-spot thermography datasets using Intersection-over-Union (IoU) and clDice, a topology-aware metric suited to thin crack structures. The results demonstrate three key advantages of our approach over a U-Net baseline. First, we achieve slightly higher segmentation accuracy in the source domain, effectively separating crack patterns from thermal artifacts. Secondly, during sequential domain adaptation, FMs exhibit strong robustness to catastrophic forgetting, retaining more than 80% of their original performance, whereas the CNN baseline collapses to near-zero accuracy. Finally, while zero-shot generalization remains challenging, FMs show improved robustness in unseen domains, detecting crack-like structures without the severe hallucinations observed in standard CNNs. These results indicate that FMs offer a promising approach for robust and label-efficient automated crack detection in industrial thermographic inspection.
