Depth-Adaptive Object Identification Using Terahertz Time-Domain Spectroscopic Data

Abstract

This study focused on machine-learning-based object recognition using the terahertz (THz) band time-domain spectroscopy (TDS) measurements. The reflection responses from each object showed severe depth dependence due to the out-of-focus effect, which was hardly compensated in the analytical propagation model. Such depth dependence severely limits the application range of THz-TDS-based object recognition, where the depth adjustment is not easily achieved, such as in subsurface imaging or industrial product inspection in an assembly line. In this study, we developed a simple adaptive depth signal conversion scheme using a calibration object before neural-network-based supervised learning. Experimental validations using several mono- or disaccharides and their THz-TDS data demonstrated that our compensation scheme successfully enhanced the recognition ratio.