Defect-aware Super-resolution Thermography by Adversarial Learning

Infrared thermography is a valuable non-destructive tool for inspection of materials. It measures the surface temperature evolution, from which hidden defects may be detected. Yet, thermal cameras typically have a low native spatial resolution resulting in a blurry and low-quality thermal image sequence and videos. In this study, a novel adversarial deep learning framework, called Dual-IRT-GAN, is proposed for performing super-resolution tasks. The proposed Dual-IRT-GAN attempts to achieve the objective of improving local texture details, as well as highlighting defective regions. The generated high-resolution images are then delivered to the discriminator for adversarial training using GAN's framework. The proposed Dual-IRT-GAN model, which is trained on an exclusive virtual dataset, is demonstrated on experimental thermographic data obtained from fibre reinforced polymers having a variety of defect types, sizes, and depths. The obtained results show its high performance in maintaining background colour consistency and removing undesired noise, and in highlighting defect zones with finer detailed textures in highresolution.