Identifying a concealed substance by terahertz spectroscopy using DenseFormer combined with MultiXtract

Terahertz fingerprints can be utilized for detecting hazardous materials. However, when used for human body security inspection, real-world conditions such as textile coverings and environmental humidity can interfere with terahertz wave propagation, leading to peak distortions that significantly impact peak matching accuracy and attenuate the ability to identify substances. To address these issues, terahertz time-domain spectroscopy (THz-TDS) data have been collected for common substances, such as flammable liquids and drugs, covered by various clothing material under different humidity conditions. Data pre-processing methods including wavelet denoising, peak segment extraction, and target-driven conditional generative network (TCGN) augmentation are used to improve data quality. Then, through the multi-feature extraction process, namely MultiXtract, unique three-channel images for these spectroscopy data have been obtained. DenseFormer network is used as the base model to classify these three-channel images, achieving a testing accuracy of 98.9%. The results demonstrate that combining DenseFormer with MultiXtract can significantly enhance the identification of concealed substances by terahertz spectroscopy.