IRT-GAN: A generative adversarial network with a multi-headed fusion strategy for automated defect detection in composites using infrared thermography

InfraRed Thermography (IRT) is a valuable diagnostic tool to non-destructively detect defects in fiber reinforced polymers. Often, a range of processing techniques are applied, e.g. principal component analysis, Fourier transformation, and thermographic signal reconstruction, in an attempt to enhance the defect detectability. Still, for the actual defect detection and evaluation, the interpretation by an expert operator is required which thus limits the (industrial) application potential of infrared thermography. This study proposes a Generative Adversarial Network (GAN) framework, termed IRT-GAN, to create a single unique thermal-image-to-segmentation translation of defects in composite materials. A large augmented numerical dataset has been simulated for a range of composite materials with different defects in order to train the IRT-GAN model. Integrated with the Spatial Group-wise Enhance layer, the IRT-GAN takes six pre-processed thermal images, thermographic signal reconstruction images in our case, as input and progressively fuses them via a multi-headed fusion strategy in the Generator. As such, this proposed IRT-GAN framework leads to the automated generation of a unique defect segmentation image. The high performance of the IRT-GAN, trained on the virtual dataset, is demonstrated on experimental data of both glass and carbon fiber reinforced polymers with various defect types, sizes, and depths. In addition, it is investigated how early, middle, and late-stage feature fusion in the GAN influences the segmentation performance.