A Feature Set for Enhanced Automatic Segmentation of Hyperspectral Terahertz Images

Terahertz time-domain spectroscopic imaging (THz-TDS imaging) producesimages with hundreds of channels. Automatic as well as manual imageanalysis is therefore difficult. We propose to use a feature set thatreduces the number of channels down to 21 and still preserves all importantinformation. Both spectral and time-domain features areincluded in this set, and thereby information aboutdifferent content is gained. We show the practicalapplicability of this approach, by using it on images from different areas of interest. Wefurthermore illustrate its advantages to the classical approach byperforming a clustering-based image segmentation on the full spectraldata and the proposed feature set. Using this reduced but representative information improves thesegmentation quality and makes THz-TDS imageprocessing and segmentation feasible and less prone to the``curse of dimensionality´´.