Abstract: The molecular vibration and rotation energy levels of explosives have unique fingerprint characteristics in the terahertz spectrum, and terahertz wave has strong permeability and low energy to non-polar substances and dielectric materials. Therefore, the use of terahertz spectrum can realize the nondestructive detection of dangerous goods in hidden environment. However, the standard library of terahertz absorption spectroscopy of materials is not perfect presently, the parameters of terahertz spectrometers on the market are different, and the detection standards are not uniform, resulting in the unreliable identification methods solely relying on absorption peaks. To address the problems, an identification technical route that no longer depends on the absorption peaks is proposed. In the method, firstly, terahertz absorption spectrum of substances with different frequency resolutions and different obstacle hidden conditions are extracted, the continuous wavelet transform of Marr is used to get a wavelet frequency domain scale map with unique characteristics, and then a data set is established. Secondly, combined with the transfer learning method, the transfer learning of Xception network is used to train and identify the data set. Experimental results show that this method is very effective in identifying explosive dangerous goods with different obstacle hidden conditions, and the recognition rates can reach 94%. It is indicated that the recognition accuracy of the proposed method can not be affected by system factors such as frequency resolution, which provides a new technical approach for non-destructive identification of dangerous goods hidden by obstacles such as package.

Key words: spectroscopy, terahertz spectroscopy, frequency resolution, continuous wavelet transform of Marr, transfer learning of Xception network