Automated crack detection on metallic materials with flying-spot thermography using deep learning and progressive training

In non-destructive testing for metallic materials, ‘Flying-spot’ thermography allows the detection of cracks thanks to the scanning of samples by a local laser heat source observed in the infrared spectrum. However, distinguishing a crack from other surface structures such as air ducts or non-planar shapes on the material surface can be challenging in an automation perspective. To address this, we propose to use deep learning techniques, which can exploit contextual information but require a significant amount of labelled data. This study presents a training method based on curriculum learning and recent denoising diffusion models to generate synthetic images. The protocol progressively increases the complexity of training images, using successively simulated data from a multi-physics finite-element software, synthetically generated data with diffusion process, and finally real data. Several detection scores are measured for various machine learning and deep learning architectures, demonstrating the benefits of the proposed approach for regular application cases and degraded experimental conditions, consisting of limited thermal enlightenment recordings.