Terahertz (THz) reflectometry has been employed to quality control of glass fiber-reinforced polymer (GFRP) composites in a nondestructive and contactless fashion. Owing to the relatively long wavelength of THz electromagnetic waves (300 um at 1 THz), as well as diffraction and absorption dispersion during propagation, THz imaging fails to visualize subtle damage in GFRP composites. We design one lightweight deep-learning network based on multiscale residual attention mechanism to characterize damage at different depths in GFRP composites. It is experimentally demonstrated that the proposed framework significantly improves the spatial resolution of THz images, thus allowing for the high-precision localization and classification of hidden damages. Compared to classic object detection models, our method has proved to be beneficial to improve the accuracy of damages characterization within THznondestructive testing (NDT) scenarios.
Terahertz Localization and Classification for Damages in GFRP Composites Using Complex Network with Residual Attention Mechanism
