Analysis of terahertz spectral images of explosives and bio-agents using trained neural networks

A non-invasive means to detect and characterize concealed agents of mass destruction in near real-time with a wide field-of-view is under development. The method employs spatial interferometric imaging of the characteristic transmission or reflection frequency spectrum in the Terahertz range. However, the successful (i.e. low false alarm rate) analysis of such images will depend on correct distinction of the true agent from non-lethal background signals. Neural networks are being trained to successfully distinguish images of explosives and bioagents from images of harmless items. Artificial neural networks are mathematical devices for modeling complex, non-linear relationships. Both multilayer perception and radial basis function neural network architectures are used to analyze these spectral images. Positive identifications are generally made, though, neural network performance does deteriorate with reduction in frequency information. Internal tolerances within the identification process can affect the outcome.