Detection Of Liquid Contraband Based On Terahertz Spectrum And Neural Network

Terahertz is now achieving increasing attention in security inspection owning to its non-destructiveness and deep penetrability of most packaging materials,such as leather,wood and wrapper.However,two major obstacles remain in spectral classification of liquid contraband:(1)The components of some contraband are complex and the position and intensity of its terahertz peak are difficult to obtain,which make extracting manual features accurately and efficiently become incapable.(2)The spectral overlapping effect in similar types of contraband makes it difficult to classify the spectrum.In order to solve the problem above,in this paper,we take the terahertz spectrum of liquid contraband as the research object,constructing the terahertz spectral dataset and analyzing the trend and characteristics of terahertz spectra.In addition,after preprocessing the terahertz spectral dataset,we use deep learning methods to extract and classify the feature of the dataset.The main research work are as follows:1.We construct two terahertz spectral datasets using the terahertz time-domain spectroscopy system,including a dataset with five different kinds of liquid contraband(pure ethanol,pure water,toner,sesame oil,and rapeseed oil)and a dataset of ethanol aqueous solutions with different concentrations(mixtures of pure ethanol and pure water in volume ratios of 3:0,2:1,1:1,1:2,0.5:2,0.5:4 and 0:3).2.We propose a real-time multi-class and multi-concentration liquid contraband spectral classification framework based on a Convolutional Neural Network(CNN).This framework can not only identify liquid contraband with complex composition,but also classify different concentrations of contraband.At the same time,we evaluate the robustness of the framework under different signal-to-noise ratio(SNR).3.Based on the liquid contraband spectral classification framework,we develop a graphical user interface for analysis and detection of liquid contraband,which realizes the upload and rapid identification of liquid contraband spectra.Experimental results demonstrate that our proposed CNN-based framework achieves the best performance compared with other deep learning and machine learning algorithms.In the case of low SNR,the classification accuracy of different types of liquid contraband and different concentrations of ethanol aqueous solutions can reach 98.00% and 97.14%,respectively.Meanwhile,the liquid contraband analysis and detection user interface designed by us can provide convenience for security personnel and accelerate the promotion of terahertz products in security inspections.