Research On Real-time Classification Of Hidden Dangerous Goods In Terahertz Time-domain Spectroscopy

The world is in a period of high incidence of public security incidents,and the economic losses caused by such incidents are immeasurable every year.Security inspection and monitoring of public places such as subways,stations,and airports is the most effective way to prevent sudden terrorist incidents.With the rapid development of security inspection technology,many methods for classification and identification by analyzing the spectrum of substances have appeared in the field of dangerous goods detection.Among them,the terahertz time-domain spectroscopy technology is widely used with the help of the specificity of the terahertz spectrum.However,traditional terahertz time-domain spectral classification methods often need to manually filter features based on empirical knowledge.The identification process is time-consuming and has poor generalization ability,and can only be applied to small sample scenarios,which is difficult to meet the real-time detection of dangerous goods in actual security inspections.,efficient demand.In the dangerous goods detection task,there are also problems such as insufficient training samples and limited types of identification.Aiming at the problems related to the real-time classification of terahertz time-domain spectra of hidden dangerous goods,this thesis takes flammable and explosive liquids as the research object,and focuses on the classification and identification of terahertz time-domain spectra under hidden structures.A time-domain spectral data identification method,which uses the terahertz time-domain spectral data of liquid dangerous goods under a concealed structure as the input of the model,and can realize the identification of dangerous goods without relying on artificial experience for feature extraction.The specific work content is as follows:(1)Aiming at the problem of low classification accuracy of spectral data caused by the changeable environment,item type and composition,and hidden structure in actual security inspection,the construction method of labeling data set under the condition of complex variable combination is studied.Seven kinds of flammable and explosive dangerous goods,such as alcohol,kerosene,edible oil,frankincense oil,turpentine oil,rosin oil,camphor oil,which are common in daily life,are selected as experimental samples,cotton,woolen wool,and leather are used as coverings,and a mineral water bottle is used for one time.Three concealed structures were designed as containers,and the reflective terahertz time-domain spectroscopy system developed by Japan's Advantest Company was used as a spectral measurement instrument to establish a terahertz time-domain spectral data set for concealed dangerous goods.(2)Aiming at the problem that the traditional terahertz time-domain spectral recognition method has low classification efficiency and accuracy when recognizing highdimensional composite data,a fusion residual network(Res Net)and Long-Short Term Memory(LSTM)network is proposed of terahertz time-domain spectroscopy for the identification of concealed dangerous goods.Firstly,the preprocessed terahertz timedomain spectral data set is extracted through the Res Net residual network to extract its spatial dimension features,and then the LSTM long and short-term memory recurrent neural network is used to mine deep time series features,and then the feature extraction of the fusion of Res Net and LSTM will be used.The method obtains the deep spectral features including spatial features and time series features and inputs them into the softmax classifier for classification and recognition.The experiment tested and analyzed the terahertz time-domain spectral data of 7 types of flammable and explosive liquids such as alcohol,kerosene,edible oil,frankincense oil,turpentine,rosin oil,and camphor oil under the hidden structure,and compared them with Res Net,CNN,Accuracy comparison of FCN and MLP.The experimental results show that using the terahertz time-domain spectral recognition method of the Res Net-LSTM network,the classification accuracy reaches98.33%,which is 4.84% higher than the recognition accuracy of the better CNN networks in Res Net,CNN,FCN and MLP.(3)Aiming at the problem that the lack of terahertz time-domain spectral data leads to the low accuracy of terahertz time-domain spectral recognition based on deep learning algorithms,a fully-connected assisted classification generative adversarial network based terahertz time-domain spectral concealment of dangerous goods samples enhancement is proposed.method.The generator and discriminator of FC-ACGAN are constructed using fully connected layers to improve the quality of the generated data and the ability to discriminate the authenticity of the data.In the experiment,the reflection-type terahertz time-domain spectrometer system was used to measure the terahertz time-domain spectral data of 7 types of flammable and explosive liquids such as alcohol,kerosene,edible oil,frankincense oil,turpentine,rosin oil,and camphor oil under the concealed structure.A total of 1407 Then,the enhanced data sets were injected into the trained classification models,and the recognition accuracy indicators were analyzed and tested,and compared with Mixup.Using FC-ACGAN to enhance and expand the original samples,the recognition accuracy of Res Net,CNN,FCN,and MLP classification models increased by4.8%,1.55%,5.76%,and 8.42%,respectively,while using Mixup to enhance the recognition accuracy of the expanded classification model.They have increased by 2.73%,1.48%,2.17%,and 1.73%,respectively,which is smaller than that of FC-ACGAN.(4)Combining the effectiveness of FC-ACGAN for the expansion of terahertz timedomain spectral data for concealed dangerous goods and the accuracy of Res Net-LSTM for the recognition of terahertz time-domain spectral data,a combination of FC-ACGAN and Res Net-LSTM for concealed dangerous goods is proposed.A method for identification of terahertz time-domain spectral data.First,the original terahertz time-domain spectral data set is input into the FC-ACGAN model to generate new data samples,and then the generated data samples are mixed with the original data samples to form an expanded data set,which is sent to the Res Net-LSTM model for distinction and recognition accuracy.It reaches 99.42%,which is 1.09% higher than the classification accuracy on the original dataset.Finally,by changing the number of generated samples in the training set,a comparative experiment is formed.The experimental results show that with the continuous increase of the number of generated samples in the training set,the classification accuracy is also continuously improved,which verifies the effectiveness of the samples generated by FC-ACGAN.