Terahertz spectrum-based method for identifying dangerous goods in postal parcel

Abstract

The invention discloses a terahertz spectrum-based method for identifying dangerous goods in a postal parcel, which comprises the following steps of: firstly, collecting terahertz time-domain spectrum signals of dangerous goods samples in the parcel, calculating the absorptivity and the refractive index of each dangerous goods sample, inputting the actually measured absorptivity # imgabs0 # of a sample j into a trained generative adversarial network, and calculating the absorptivity # imgabs0 # of the sample j; expanding and enhancing actually measured absorptivity data to obtain M groups of absorptivity data of the sample j, and superposing any group of absorptivity of the sample j obtained by the generative adversarial network with the refractive index of the sample j to obtain M groups of absorptivity-refractive index mixed spectrums, training a hazardous article classification model by using the absorptivity-refractive index mixed spectrum data of the N * M groups of hazardous article samples and the absorptivity-refractive index mixed spectrum data of the N groups of samples, and finally identifying the to-be-detected wrapped sample based on the trained hazardous article classification model. According to the method, the package does not need to be opened, and rapid and accurate detection of dangerous goods in the package is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of terahertz spectrum recognition, and in particular relates to a method for identifying dangerous goods in postal parcels based on terahertz spectrum. Background Art

[0002] Dangerous goods refer to inflammable and explosive items, dangerous chemicals, radioactive items and other items that can endanger personal safety and property safety. Therefore, packages need to be inspected at stations, airports and other places. Due to its non-destructive, real-time and contactless characteristics, X-ray technology is widely used in package security inspection, but it has the problem of low recognition rate. Terahertz waves are electromagnetic waves with a wavelength between infrared and microwaves. Because they are in the transition region from photonics to electronics, they have many unique properties, such as fingerprint spectroscopy, low energy, special penetration, etc. As an example of special penetration, terahertz waves have stronger penetration into dielectric materials than visible light and infrared, so they have been used in non-destructive testing of many substances. At the same time, since the rotational energy levels and some vibration modes of many biological molecules are in the terahertz band, the pulse waves generated by the terahertz spectrometer are very sensitive to changes in molecular structure and the combination between molecules. Therefore, non-destructive detection of substances can be achieved through terahertz. When existing terahertz spectroscopy is used to identify substances, features are generally extracted from a single data set, such as absorption spectra or time domain spectra. This method does not fully integrate the amplitude and phase information of the pulsed terahertz wave, thus limiting the accuracy of the prediction. In addition, in addition to considering the impact of sample diversity, the detection of dangerous goods packages also needs to consider the package material, shape, etc. as well as the security inspection environment. There are many noises that affect the accuracy of identification, and the training data is limited. The directly established model cannot fully cover the above noises and is inaccurate. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and seek to design a method for identifying dangerous goods in postal parcels based on terahertz spectroscopy, which solves the problem of low recognition rate of existing methods for identifying dangerous goods in parcels. At the same time, it solves the problem that when using existing terahertz spectroscopy for material identification, the information extracted from a single data cannot fully reflect the characteristics of the material, the training data is limited and cannot cover the used usage scenarios, thus affecting the recognition accuracy.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for identifying dangerous goods in postal parcels based on terahertz spectroscopy comprises the following steps:

[0006] Step 1: Different dangerous goods samples are placed in different packages respectively, and the terahertz time domain spectrum signal of the dangerous goods samples in the packages is collected by a transmission terahertz time domain system. At the same time, the terahertz time domain spectrum signal data of the corresponding empty packages are collected as reference signals. The dangerous goods samples include different types of dangerous goods;

[0007] Step 2: Clean and filter the data based on the terahertz time-domain spectroscopy signal, calculate the absorptivity and refractive index of each dangerous goods sample, and obtain the original absorptivity and refractive index data of N dangerous goods samples. Among them, the measured absorption rate of sample j is for The measured refractive index is Y j 0 for j is a positive integer from 1 to N;

[0008] Step 3: The measured absorbance of sample j Input the trained adversarial generative network to expand and enhance the measured absorption rate data, and obtain M groups of absorption rate data of sample j, which are in, m is a positive integer from 1 to M;

[0009] Step 4: Superimpose any set of absorbance of sample j obtained by the adversarial generative network with its refractive index to obtain M sets of absorbance-refractive index mixed spectra, which are in, The 1D-GAN adversarial generative network data is expanded M times, thereby obtaining N×M sets of absorbance-refractive index mixed spectral data;

[0010] Step 5: Divide the absorbance-refractive index mixed spectral data of the N×M groups of dangerous goods samples and the absorbance-refractive index mixed spectral data of the N groups of samples into a training set, a verification set and a test set respectively. The training set is used for dangerous goods classification model training, the verification set is used to correct the dangerous goods classification model parameters, and the test set is used to detect the generalization performance of the dangerous goods classification model to obtain a trained dangerous goods classification model;

[0011] Step 6: Collect the terahertz time-domain spectral signal of the package sample to be tested, and use the terahertz time-domain spectral signal of a blank optical path without a sample as a reference signal. Obtain the absorptivity and refractive index of the package sample to be tested through step 2. Input the absorptivity and refractive index into the dangerous goods classification network model trained in step 5, and the model outputs the category of the sample in the package.

[0012] Step 2 is specifically as follows: Step 201: Averaging multiple groups of terahertz time-domain spectral signals of the same dangerous goods sample collected and using the wavelet shrinkage denoising method to further reduce the residual noise, and then using the Fourier transform of the time domain spectrum E(t) (Formula 1) to extract the spectrum information carried by the sample, and obtain the amplitude and phase change information carried by the dangerous goods sample from the transfer function (Formula 2);

[0013]

[0014] In the formula, represents the Fourier transform of the time domain spectrum E(t), ω is the angular frequency, and φ(ω) are the frequency domain spectra The amplitude and phase of , i is a complex number, T(ω) is the transfer function, is the frequency domain spectrum of the sample, is the frequency domain spectrum of the reference signal, Δφ(ω) is the phase change caused by propagation in the sample, and ρ(ω) represents the amplitude ratio of the sample and reference signals;

[0015] Step 202: Using the transfer function and assuming that the extinction rate is much smaller than the refractive index in the terahertz band, the Fresnel law is used to derive the absorption rate and refractive index of the dangerous goods sample, as shown in Formula 3 and Formula 4;

[0016]

[0017] Where n(ω) represents the refractive index, α(ω) represents the absorptivity, d is the sample thickness, and c is the speed of light in a vacuum;

[0018] Step 203: The correlation between the refractive index and absorptivity of the dangerous goods sample (dangerous goods sample in the package) and the standard absorptivity and standard refractive index of the corresponding dangerous goods sample is measured according to the Pearson correlation coefficient. If the result is less than 0.9, it is retained, otherwise it is deleted. After screening, high-quality signals can be retained as much as possible to facilitate the next step of data enhancement.

[0019] In step 3, the adversarial generative network includes a generator and a discriminator. The measured absorption rate of the sample in step 2 is used as training data to train the adversarial generative network. Specifically, the measured absorption rate is input into the generator, the generated absorption rate is output, and then the generated absorption rate and the measured absorption rate are input into the discriminator, and the output is the probability that the generated absorption rate is true.

[0020] In step 5, three models are used to train the dangerous goods classification model, namely the support vector machine SVM model, the convolutional neural network CNN model and the Bayesian network Bayes model.

[0021] In step 5, the convolutional neural network includes an input block, an ECA network, and a classification block. The input block includes an input layer, a first convolutional layer, and a reshaping layer. The ECA network includes a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and an ECA module. The classification block includes a first fully connected layer, a second fully connected layer, and an output layer. The input block passes the absorbance-refractive index mixed spectrum to the first convolutional layer. The first convolutional layer and the reshaping layer reshape the absorbance-refractive index mixed spectrum into a feature map, which is input to the ECA network again. The second and third convolutional layers in the ECA network use 2D filters to capture the relationship between the filtered signals generated by the first convolutional layer, and then output 32 channels respectively. The fourth convolutional layer in the ECA network uses a 1×1 kernel to increase the number of channels to 64, thereby providing a feature map for the ECA network. Module A extracts more channel information; the ECA module consists of a global pooling layer, a one-dimensional convolution layer, and an activation layer that uses a sigmoid function to present nonlinear features to channel weights. Channel multiplication is performed on the output of layer Conv4 and the attention coefficient. In the classification block, two fully connected layers of sizes 256 and 128 are used to reduce the number of hyperparameters passed from the ECA network. After that, a dense layer based on the softmax activation function is used to output the probability of each category for classification. Batch normalization is applied after each convolution layer to standardize the layer output of each small batch. Finally, a pooling layer is used to reduce the output dimension.

[0022] Compared with the prior art, the present invention has the following beneficial effects: (1) by using a generative adversarial network to train data, not only the amount of data is expanded, but also the model training data is closer to the data obtained in the real environment, thereby improving the richness of the training data and thus improving the accuracy of identifying dangerous goods; (2) compared with traditional single data (such as absorption spectrum or time domain spectrum), the high-dimensional spectral information in the terahertz spectrum of dangerous goods is extracted by combining the absorption rate and the refractive index, and the amplitude and phase information of the pulsed terahertz wave are fully integrated, so that the characteristic information of the material can be fully extracted, thereby improving the accuracy of the prediction; (3) without opening the package, the dangerous goods in the package can be quickly and accurately detected. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is the absorption spectrum of the same sample obtained in different packages.

[0024] Figure 2 It is a 1D-GAN model.

[0025] Figure 3 Loss score curves for the generator and discriminator during training. DETAILED DESCRIPTION

[0026] The present invention will be further described below by means of specific examples.

[0027] Example 1

[0028] A method for identifying dangerous goods in postal parcels based on terahertz spectroscopy comprises the following steps:

[0029] Step 1: different dangerous goods samples are loaded into different packages respectively, and the terahertz time domain spectrum signal of the dangerous goods samples in the packages is collected by a transmission terahertz time domain system. At the same time, the terahertz time domain spectrum signal data of the corresponding empty packages are collected as a reference signal. The dangerous goods samples collected in this embodiment include different types of dangerous goods, such as 20 methamphetamine samples, 25 ketamine samples, 30 morphine samples, 28 marijuana samples, etc.

[0030] Step 2: Clean and filter the data based on the terahertz time-domain spectroscopy signal, and calculate the absorption rate and refractive index of each dangerous goods sample. Specifically include:

[0031] Step 201: Averaging multiple groups of terahertz time-domain spectrum signals of the same dangerous goods sample and using a wavelet shrinkage denoising method to further reduce residual noise, then using the Fourier transform of the time-domain spectrum E(t) (Formula 1) to extract the spectrum information carried by the sample, and obtaining the amplitude and phase change information carried by the dangerous goods sample from the transfer function (Formula 2);

[0032]

[0033] In the formula, represents the Fourier transform of the time domain spectrum E(t), ω is the angular frequency, and φ(ω) are the frequency domain spectra The amplitude and phase of , i is a complex number, T(ω) is the transfer function, is the frequency domain spectrum of the sample, is the frequency domain spectrum of the reference signal, Δφ(ω) is the phase change caused by propagation in the sample, and ρ(ω) represents the amplitude ratio of the sample and reference signals;

[0034] Step 202: Using the transfer function and assuming that the extinction rate is much smaller than the refractive index in the terahertz band, the Fresnel law is used to derive the absorption rate and refractive index of the dangerous goods sample, as shown in Formula 3 and Formula 4;

[0035]

[0036] Where n(ω) is the refractive index, α(ω) is the absorptivity, d is the sample thickness, and c is the speed of light in a vacuum. Figure 1The absorption rate (absorption spectrum) of the same sample obtained in different packages. It can be seen that although the absorption rate of the same substance has certain commonalities, the form and material of the package will cause differences in the absorption rate of the same sample, which will affect the identification result. The purpose of this embodiment is to use the data for training, which can not only extract the commonalities in the terahertz time-domain spectrum data of the same substance sample, but also retain the noise in the environment;

[0037] Step 203: The correlation between the refractive index and absorptivity of the dangerous goods sample (dangerous goods sample in the package) and the standard absorptivity and standard refractive index of the corresponding dangerous goods sample is measured according to the Pearson correlation coefficient. If the result is less than 0.9, it is retained, otherwise it is deleted. After screening, high-quality signals can be retained as much as possible to facilitate the next step of data enhancement.

[0038] Specifically, the dangerous goods sample is ground and pressed into a tablet to make a solid sample tablet, and then a transmission terahertz time-domain system is used to collect the wrapped terahertz time-domain spectral signal. At the same time, the sample-free optical path is collected to obtain reference time-domain spectral signal data, and the standard absorption rate and standard refractive index of the dangerous goods sample are obtained based on the terahertz time-domain spectral signal.

[0039] After processing in step 2, the original absorption rate and refractive index data of N dangerous goods are obtained. Among them, the measured absorption rate of sample j is for The measured refractive index is Y j 0 for j is a positive integer from 1 to N.

[0040] Step 3: Since the dangerous goods samples are in packages, the data collected from the same sample in different packages are quite different, and there are a lot of interference signals in the collection environment, such as mechanical vibration of the collection equipment. In addition, the amount of samples collected in steps 1-2 is limited in the actual process, and the amount of information contained in the collected data is small, which is far lower than the training data required by the neural network. Therefore, the measured absorption rate of sample j is The trained generative adversarial network (1D-GAN) is input to expand and enhance the measured absorption rate data, and M groups of absorption rate data of sample j are obtained, which are in, m is a positive integer from 1 to M;

[0041] like Figure 2As shown, the adversarial generative network (1D-GAN) includes a generator and a discriminator. The measured absorption rate of the sample in step 2 is used as training data to train the adversarial generative network (1D-GAN). Specifically, the measured absorption rate is input into the generator, and the generated absorption rate is output. Then, the generated absorption rate and the measured absorption rate are input into the discriminator, and the output is the probability that the generated absorption rate is true. Figure 3 These are the loss score curves of the generator and discriminator networks during training. The closer the two curves are, the better the generation effect of the generator is, and the generated absorption rate is approximately close to the true absorption rate.

[0042] The similarity between the absorption coefficient spectral data generated by three dangerous goods (wherein, dangerous goods 1 is methamphetamine, dangerous goods 2 is ketamine, and dangerous goods 3 is morphine) and the real absorption coefficient spectrum is evaluated, and the root mean square error RMSE of the measured absorbance and the generated absorbance in Table 1 is obtained. It can be seen from Table 1 that when epoch increases from 0 to 100, the RMSE of the generated data and the real data decreases rapidly. Although when epoch is 500, it can be seen that the general outline of the generated spectrum is similar to the real spectrum, but there is a large noise. RMSE is not the smaller the better. In order to better simulate real-time data acquisition, the main features of the original terahertz absorptivity of the sample are retained. The absorptivity doped with unstable noise is the data we hope for. Therefore, the data obtained by the adversarial generation network (1D-GAN) trained in this embodiment is close to the real terahertz absorptivity.

[0043] Table 1

[0044]

[0045] Step 4: Superimpose any set of absorbances of sample j obtained by the 1D-GAN with its refractive index to obtain M sets of absorbance-refractive index mixed spectra, which are in, The 1D-GAN adversarial generative network data is expanded M times, thereby obtaining N×M sets of absorbance-refractive index mixed spectral data;

[0046] The refractive index of the sample can also be amplified by a generative adversarial network (1D-GAN), but the characteristics of the generated refractive index data are not much different from the original refractive index data, so it is chosen not to be processed.

[0047] Step 5: The absorbance-refractive index mixed spectral data of N×M groups of dangerous goods samples (the data gan dataset enhanced by the adversarial generative network in step 4) and the absorbance-refractive index mixed spectral data of N groups of samples (the data raw dataset obtained after cleaning and marking in step 2) are respectively divided into a training set, a verification set and a test set. The training set is used for dangerous goods classification model training, the verification set is used to correct the parameters of the dangerous goods classification model, and the test set is used to detect the generalization performance of the dangerous goods classification model to obtain a trained dangerous goods classification model. This embodiment uses three models to train the dangerous goods classification model, namely the support vector machine SVM model, the convolutional neural network CNN model and the Bayesian network Bayes model. Table 2 shows the classification and recognition accuracy of different models for two types of dangerous goods data. It can be seen that after being processed by the adversarial generative network, the accuracy of the three models is improved, the CNN model has the smallest improvement, and the Bayes model has the largest improvement.

[0048] Table 2

[0049]

[0050] The convolutional neural network includes an input block, an ECA network, and a classification block. The input block includes an input layer, a first convolution layer, and a reshaping layer. The ECA network includes a second convolution layer, a third convolution layer, a fourth convolution layer, and an ECA module. The classification block includes a first fully connected layer, a second fully connected layer, and an output layer. The input block passes the absorbance-refractive index mixed spectrum to the first convolution layer. The first convolution layer and the reshaping layer reshape the absorbance-refractive index mixed spectrum into a feature map, which is input to the ECA network again. The second and third convolution layers in the ECA network use 2D filters to capture the relationship between the filtered signals generated by the first convolution layer, and then output 32 channels respectively. The fourth convolution layer in the ECA network uses a 1×1 kernel to increase the number of channels to 64, thereby extracting more channel information for the ECA module.

[0051] The ECA module consists of a global pooling layer, a one-dimensional convolution layer, and an activation layer that uses a sigmoid function to present nonlinear features to channel weights. Channel multiplication is performed on the output of layer Conv4 and the attention coefficient. In the classification block, two fully connected layers of size 256 and 128 are used to reduce the number of hyperparameters passed from the ECA network. After that, a dense layer based on the softmax activation function is used to output the probability of each category for classification. Batch normalization is applied after each convolution layer to standardize the layer output of each mini-batch, and finally a pooling layer is used to reduce the output dimension.

[0052] The experimental results show that the combination of absorptivity and refractive index can extract high-dimensional spectral information from terahertz spectra and effectively identify dangerous goods in packages. For convolutional neural networks (CNNs), these high-dimensional spectral information are abstracted into feature maps and passed to convolutional neural networks (CNNs) equipped with effective channel attention (ECA) mechanisms to calibrate the interdependencies between data channels.

[0053] Step 6: Collect the terahertz time-domain spectral signal of the package sample to be tested, and use the terahertz time-domain spectral signal of a blank optical path without a sample as a reference signal. Obtain the absorptivity and refractive index of the package sample to be tested through step 2. Input the absorptivity and refractive index into the dangerous goods classification network model trained in step 5, and the model outputs the category of the sample in the package.

Claims

1. A method for identifying dangerous goods in postal parcels based on terahertz spectroscopy, characterized in that: The following steps are involved: Step 1: Different dangerous goods samples are placed in different packages respectively, and the terahertz time domain spectrum signal of the dangerous goods samples in the packages is collected by a transmission terahertz time domain system. At the same time, the terahertz time domain spectrum signal data of the corresponding empty packages are collected as reference signals. The dangerous goods samples include different types of dangerous goods; Step 2: Clean and filter the data based on the terahertz time-domain spectroscopy signal, calculate the absorptivity and refractive index of each dangerous goods sample, and obtain the original absorptivity and refractive index data of N dangerous goods samples. Among them, the measured absorption rate of sample j is for The measured refractive index is Y j 0 for j is a positive integer from 1 to N; Step 3: The measured absorbance of sample j Input the trained adversarial generative network to expand and enhance the measured absorption rate data, and obtain M groups of absorption rate data of sample j, which are in, m is a positive integer from 1 to M; Step 4: Superimpose any set of absorbance of sample j obtained by the adversarial generative network with its refractive index to obtain M sets of absorbance-refractive index mixed spectra, which are in, The 1D-GAN adversarial generative network data is expanded M times, thereby obtaining N×M sets of absorbance-refractive index mixed spectral data; Step 5: Divide the absorbance-refractive index mixed spectral data of the N×M groups of dangerous goods samples and the absorbance-refractive index mixed spectral data of the N groups of samples into a training set, a verification set and a test set respectively. The training set is used for dangerous goods classification model training, the verification set is used for correcting dangerous goods classification model parameters, and the test set is used for testing the generalization performance of the dangerous goods classification model to obtain a trained dangerous goods classification model. Three models are used to train the dangerous goods classification model, namely, a support vector machine SVM model, a convolutional neural network CNN model or a Bayesian network Bayes model; Step 6: Collect the terahertz time-domain spectral signal of the package sample to be tested, and use the terahertz time-domain spectral signal of a blank optical path without a sample as a reference signal. Obtain the absorptivity and refractive index of the package sample to be tested through step 2. Input the absorptivity and refractive index into the dangerous goods classification network model trained in step 5, and the model outputs the category of the sample in the package.

2. The method for identifying dangerous goods in postal parcels based on terahertz spectroscopy according to claim 1, characterized in that: Step 2 is specifically as follows: Step 201: averaging multiple groups of terahertz time-domain spectrum signals of the same dangerous goods sample collected and using the wavelet shrinkage denoising method to further reduce the residual noise, and then using the Fourier transform of the time domain spectrum E(t) to extract the spectrum information carried by the sample, and obtain the amplitude and phase change information carried by the dangerous goods sample from the transfer function; In the formula, represents the Fourier transform of the time domain spectrum E(t), ω is the angular frequency, and φ(ω) are the frequency domain spectra The amplitude and phase of , i is a complex number, T(ω) is the transfer function, is the frequency domain spectrum of the sample, is the frequency domain spectrum of the reference signal, Δφ(ω) is the phase change caused by propagation in the sample, and ρ(ω) represents the amplitude ratio of the sample and reference signals; Step 202: using the transfer function and assuming that the extinction rate is much smaller than the refractive index in the terahertz band, the absorption rate and refractive index of the dangerous goods sample are derived using the Fresnel law; Where n(ω) represents the refractive index, α(ω) represents the absorptivity, d is the sample thickness, and c is the speed of light in a vacuum; Step 203: The correlation between the refractive index and the absorbance of the dangerous goods sample and the standard absorbance and the standard refractive index of the corresponding dangerous goods sample is measured according to the Pearson correlation coefficient. If the result is less than 0.9, it is retained, otherwise it is deleted.

3. The method for identifying dangerous goods in postal parcels based on terahertz spectroscopy according to claim 1, characterized in that: In step 3, the adversarial generative network includes a generator and a discriminator. The measured absorption rate of the sample in step 2 is used as training data to train the adversarial generative network. Specifically, the measured absorption rate is input into the generator, the generated absorption rate is output, and then the generated absorption rate and the measured absorption rate are input into the discriminator, and the output is the probability that the generated absorption rate is true.

4. The method for identifying dangerous goods in postal parcels based on terahertz spectroscopy according to claim 1, characterized in that: In step 5, the convolutional neural network includes an input block, an ECA network and a classification block. The input block includes an input layer, a first convolutional layer and a reshaping layer. The ECA network includes a second convolutional layer, a third convolutional layer, a fourth convolutional layer and an ECA module. The classification block includes a first fully connected layer, a second fully connected layer and an output layer. The input block transfers the absorbance-refractive index mixed spectrum to the first convolutional layer. The first convolutional layer and the reshaping layer reshape the absorbance-refractive index mixed spectrum into a feature map, which is input to the ECA network again. The second and third convolutional layers in the ECA network use 2D filters to capture the relationship between the filtered signals generated by the first convolutional layer, and then output 32 channels respectively. The fourth convolutional layer in the ECA network The layer uses a 1×1 kernel to increase the number of channels to 64, thereby extracting more channel information for the ECA module; the ECA module consists of a global pooling layer, a one-dimensional convolution layer, and an activation layer that uses a sigmoid function to present non-linear features to channel weights. Channel multiplication is performed on the output of layer Conv4 and the attention coefficient. In the classification block, two fully connected layers of sizes 256 and 128 are used to reduce the number of hyperparameters passed from the ECA network. After that, a dense layer based on the softmax activation function is used to output the probability of each category for classification. Batch normalization is applied after each convolution layer to standardize the layer output of each small batch, and finally a pooling layer is used to reduce the output dimension.

Images