Dual-IRT-GAN: A defect-aware deep adversarial network to perform super-resolution tasks in infrared thermographic inspection
Liangliang Cheng, Mathias Kersemans
Abstract
InfraRed Thermography (IRT) is a valuable diagnostic tool for detecting defects in fiber-reinforced polymers in a non-destructive manner through the measurement of surface temperature distribution. Yet, thermal cameras typically have a low native spatial resolution resulting in a blurry and low-quality thermal image sequence. This study proposes a defect-aware Generative Adversarial Network (GAN) framework, termed Dual-IRT-GAN, in order to simultaneously perform Super-Resolution (SR) and defect detection tasks in infrared thermography. Furthermore, the visibility of defective regions in generated high-resolution images are enhanced by leveraging defect-aware attention maps from segmented defect images. Following a series of augmentation techniques and a second-order degradation process, the proposed Dual-IRT-GAN model is trained on an extensive numerically generated thermographic dataset of composite materials with various defect types, sizes and depts. The high inference performance of the virtually trained Dual-IRT-GAN is demonstrated on several experimental thermographic datasets which were obtained from composite coupon specimens with various defect types, sizes, and depths, as well as from aircraft stiffened composite panels having real (production) defects.