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

Area of Science:

• Spectroscopy
• Materials Science
• Imaging Technology

Background:

• Terahertz (THz) spectroscopy is promising for detecting hazardous materials.
• Real-world conditions like clothing and humidity distort THz signals, hindering accurate substance identification.
• Existing methods struggle with peak distortions and signal attenuation.

Purpose of the Study:

• To improve the accuracy of hazardous substance identification using THz spectroscopy under realistic conditions.
• To develop a robust method for analyzing THz time-domain spectroscopy (THz-TDS) data affected by environmental factors.
• To enhance the detection capabilities for concealed materials in human body security screening.

Main Methods:

• Collected THz-TDS data for various substances concealed by clothing under different humidity levels.
• Applied data pre-processing techniques: wavelet denoising, peak segment extraction, and target-driven conditional generative network (TCGN) augmentation.
• Utilized MultiXtract for multi-feature extraction to create unique three-channel images from spectroscopy data.
• Employed a DenseFormer network for classifying the generated three-channel images.

Main Results:

• Achieved a high testing accuracy of 98.9% in substance identification.
• Demonstrated significant improvement in data quality and peak matching accuracy after pre-processing and augmentation.
• Successfully generated unique three-channel images representing complex spectral features.

Conclusions:

• The combination of DenseFormer network and MultiXtract significantly enhances the identification of concealed substances via THz spectroscopy.
• The developed method effectively overcomes challenges posed by textile coverings and humidity.
• This approach shows great potential for advanced human body security inspection systems.