IRT-Diffusion: a novel defect segmentation method based on a denoising diffusion probabilistic model and infrared thermography

Abstract

Infrared thermography (IRT) is widely used for non-destructive testing of composite structures owing to its capability to detect subsurface defects. However, practical interpretation of thermographic data still relies heavily on expert experience to distinguish defective regions from sound regions, thereby limiting the industrial applicability of IRT. To address this issue, a diffusion-based defect segmentation framework termed IRT-Diffusion is proposed. The framework combines thermal signal processing techniques with a denoising diffusion probabilistic model (DDPM) to progressively reconstruct defect structures from thermographic data. In addition, a stability sampling strategy is introduced to improve segmentation stability and reconstruction quality. The proposed IRT-Diffusion framework is trained on PVC specimens and achieves IoU, F1-score, Recall and Precision values of 88.4%, 93.4%, 94.9% and 92.7%, respectively. The model trained solely on PVC specimens was further evaluated on independently measured GFRP, CFRP and rubber-to-metal bonded specimens without retraining. These single-specimen transfer experiments provide preliminary evidence of the applicability of the proposed framework to thermographic data acquired from different material systems.