IRT-diffusion: combined denoising diffusion probabilistic models with thermal signal processing methods for automated defect detection in composites using infrared thermography

Infrared thermography (IRT) is a reliable method for detecting defects in composites with advantages of full field, non-contact, easy operation and good visualization. Nevertheless, interpretation by experts is required to distinguish between defective and sound regions in the practical evaluation of defects, which limits the industrial applications of infrared thermography. In this study, a denoising diffusion probabilistic model (DDPM) framework named IRT-Diffusion is proposed to automatically segment defective regions in thermal images. IRT-Diffusion can reconstruct a defect segmentation image from a noisy image with standard Gaussian distribution by iteratively performing multiple denoising operations. Detection results from various traditional thermal signal processing methods are employed as input for the conditional noise predictor of IRT-Diffusion to generate more accurate defect segmentation results. The core innovation of this study is that the state-of-the-art generative model is first introduced and designed for defect identification in composites using infrared thermography. To assess the performance of IRT-Diffusion, experiments were conducted on several composites panels and compared with conditional variational autoencoder (CVAE) and conditional generative adversarial network (CGAN). The results demonstrate that the proposed method achieves superior quantitative metrics and effectively extracts defective regions.