Synthetic visible-IR images pairs generation for multi-spectral NDT using flying spot thermography and deep learning
Génération de paires d'images synthétiques visible-IR pour le CND multi-spectral par thermographie Flying-Spot grâce à l'apprentissage profond
Résumé
Flying spot" laser infrared thermography (FST) is a non destructive testing technique able to detect small defects by scanning surfaces with a laser heat source. Defects, such as cracks on metallic parts, are revealed by the disturbance of heat propagation measured by an infrared camera. The association of this examination technique with inspection in the visible spectrum, giving access to surface textures and geometries difficult to observe in the IR spectrum, can increase both robustness and performance of the defect detection. However in a deep learning approach, the acquisition of large amounts of visible-IR pairs can be difficult and time-consuming. The present work proposes to explore visible-FST image pairs generation in the context of surface crack detection for metallic materials, using state-of-the-art deep generative models such as Stable Diffusion. Both accuracy of the generated samples and benefits for multi-spectral deep neural models training will be studied.