Advances in Knowledge-Based Technologies

In pharmaceutical quality control, technologies are necessary that identify different chemical compounds. Thisidentification should be performed in a non invasive way. The Terahertz (THz) technology has proven to be a usefultool in that aim. In these wavelengths chemical compounds have characteristic absorption spectra while at the sametime most packaging materials such as carton, plastics, and ceramics are not absorbing [1].There are databases that contain the characteristic spectral expression of many chemical compounds in the infraredrange [2]. As an emerging technique most THz spectra are not comparably characterized yet. There are databases suchas [3] but the quality and method of acquisition declaredly differ. For most specific applications this is a problem.Particularly to address this problem of comparability, we propose a procedure to detect peaks in THz measurementsof solids acquired by time-domain spectroscopy. On a set of six hyperspectral imaging measurements of chemicalcompounds, this procedure will be presented here.We especially propose a method to determine the Dynamic Range (DR) of spectra based on standard peak detection.Furthermore, we propose a method for baseline correction. In spite of the normalization with a reference measurement,most transmittance spectra do not have a constant baseline which makes the classification of peaks difficult. Wesimulate the basic shape of the spectra and propose to use this to create such a constant baseline.We illustrate this procedure with six measurements of chemical compounds, 36 spectra each. We use unsupervisedlearning, more particularly hierarchical clustering, to find a robust representation for these compounds by their peaks.This shall illustrate how such a procedure can be used to build comparable THz databases.