Robust identification of concealed dangerous substances using THz imaging spectroscopy

Nystad, Helle Emilia
Haakestad, Magnus Willum
Rheenen, Arthur Dirk van
Termset Emneord::Deteksjon
Nystad, Helle Emilia; Haakestad, Magnus Willum; Rheenen, Arthur Dirk van. Robust identification of concealed dangerous substances using THz imaging spectroscopy. Proceedings of SPIE, the International Society for Optical Engineering 2015 ;Volum 9483.
Size: 520k
False alarm rates must be kept sufficiently low if a method to detect and identify objects or substances is to be implemented in real life applications. This is also true when trying to detect and identify dangerous substances such as explosives and drugs that are concealed in packaging materials. THz technology may be suited to detect these substances, especially when imaging and spectroscopy are combined. To achieve reasonable throughput, the detection and identification process must be automated and this implies reliance on algorithms to perform this task, rather than human beings. The identification part of the algorithm must compare spectral features of the unknown substance with those in a library of features and determining the distance, in some sense, between these features. If the distance is less than some defined threshold a match is declared. In this paper we consider two types of spectral characteristic that are derived from measured time-domain signals measured in the THz regime: the absorbance and its derivative. Also, we consider two schemes to measure the distance between the unknown and library characteristics: Spectral Angle Mapping (SAM) and Principal Component Analysis (PCA). Finally, the effect of windowing of the measured time-domain signal on the performance of the algorithms is studied, by varying the Blackman-Harris (B-H) window width. Algorithm performance is quantified by studying the receiver-operating characteristics (ROC). For the data considered in this study we conclude that the best performance is obtained when the derivative of the absorbance is used in combination with a narrow B-H window and SAM. SAM is a more straight-forward method and requires no large training data sets and tweaking.
View Meta Data