Terahertz time-domain spectral feature extraction and classification method and system for hidden articles in atmospheric environment

Abstract

The invention discloses a method for extracting and classifying terahertz time-domain spectral features of hidden articles in an atmospheric environment. Comprising the following steps: carrying out wavelet filtering on collected terahertz time-domain spectral data of the hidden article to obtain a denoised spectral data set; performing data enhancement on the denoised spectral data set by using a deep learning model to form an expanded data set; preprocessing the expanded data set to form a multi-channel graph feature data set; carrying out transfer learning on the multi-channel graph feature data set by utilizing a neural network model to obtain a hidden article real-time detection model; and detecting and classifying the hidden articles in real time by using the detection model. According to the invention, through adoption of advanced data enhancement and multi-channel processing technologies, the identification capability of hidden articles is obviously improved. Besides, transfer learning is carried out through the improved neural network model, key features in the terahertz spectrum are fully extracted and distinguished, and the feature extraction capability and classification performance of the model on terahertz time-domain spectrum data are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of concealed object detection, and more specifically, relates to a method and system for extracting and classifying terahertz time-domain spectral features of concealed objects in an atmospheric environment. Background Art

[0002] Passenger security inspection systems in public places such as railways and airports are important means to ensure the safety of public facilities and personnel, and are used to detect hidden objects and prevent terrorist threats. However, existing security inspection equipment has many limitations: metal detectors can only detect metal objects, and it is difficult to detect non-metallic objects and cannot distinguish between types of objects; optical imagers and infrared imagers cannot penetrate clothing and shielding materials, and it is difficult to detect hidden objects; X-rays are harmful to the human body due to their strong radiation and are not suitable for the detection of carry-on items; CT detectors are complex, costly, and slow to inspect. Therefore, existing security inspection technologies are difficult to meet the needs in terms of rapid and non-destructive detection of hidden objects.

[0003] With the development of terahertz spectroscopy technology, terahertz waves have shown obvious advantages in non-destructive detection of hidden objects due to their non-ionizing radiation, excellent penetration, material recognition ability and non-contact detection, and have become an important means of modern security inspection. However, in practical applications in public places such as railways and airports, hidden object detection based on terahertz technology still faces some challenges, especially how to improve the performance of detection algorithms and reduce false detection rates under different temperature and humidity backgrounds. This is because terahertz waves are very sensitive to the absorption of water molecules. In a high relative humidity environment, water vapor will absorb terahertz waves and weaken the intensity of signal transmission, thereby affecting the accuracy and reliability of the measurement. Existing terahertz data acquisition technology usually requires the detection of material spectral information under conditions of relative humidity below 5%, but this is difficult to achieve in practical applications.

[0004] The invention patent with the prior art publication number CN117726529A proposes a millimeter wave image enhancement method based on a self-attention adversarial generative network, including: using a terahertz millimeter wave sensor to collect images of logistics packages to obtain millimeter wave image datasets A and millimeter wave image datasets B; constructing an adversarial generative network, training the adversarial generative network through the millimeter wave image dataset A, and selecting through the millimeter wave image dataset B to obtain an image enhancement model; using the image enhancement model to enhance the image of the test set of the millimeter wave image dataset A. This scheme uses an adversarial generative network scheme, which has deficiencies in training stability, generated sample diversity, and the introduction of physical constraints. Summary of the invention

[0005] In order to overcome the problems of insufficient diversity and specificity of generated data in the prior art and low detection accuracy under different humidity conditions, the present invention provides a method and system for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment.

[0006] The primary purpose of the present invention is to solve the above technical problems. The technical solutions of the present invention are as follows:

[0007] The first aspect of the present invention provides a method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment, comprising the following steps:

[0008] Perform wavelet filtering on the collected terahertz time-domain spectrum data of hidden objects to obtain a denoised spectrum data set;

[0009] Use deep learning models to perform data enhancement on the denoised spectral dataset to form an expanded dataset;

[0010] Preprocessing the expanded data set to form a multi-channel image feature data set;

[0011] Use the neural network model to perform transfer learning on the multi-channel graph feature dataset, train the model, and use the loss function to optimize the model parameters to obtain a real-time hidden object detection model;

[0012] Use the trained detection model to perform real-time detection and classification of hidden objects.

[0013] Furthermore, the deep learning model for data enhancement of the denoised spectral dataset is a target driving condition generation model, and the target driving condition generation model adds a target condition generation network on the basis of the generator and the discriminator.

[0014] Furthermore, the preprocessing of the data-enhanced spectral data set includes calculating a wavelet spectrum, a Mel spectrum and an absorption spectrum of the spectral data.

[0015] Furthermore, the specific steps of calculating the wavelet spectrum of the spectral data are:

[0016] Perform discrete wavelet transform on the input signal to obtain the approximation coefficient and detail coefficient of the first-level decomposition;

[0017] The coefficients are interpolated and resampled using an interpolation function to ensure that the data matches the target shape;

[0018] The resampled approximation coefficients and detail coefficients are combined into a two-dimensional matrix as the final wavelet spectrum.

[0019] Furthermore, the specific steps of calculating the Mel spectrum of the spectral data are:

[0020] Convert the time domain signal into tensor form;

[0021] Calculate and set the window length and jump length to split the time domain signal tensor into multiple frames;

[0022] Use the Mel spectrum conversion function to convert the framed signal tensor into Mel spectrum representation;

[0023] Calculate the logarithmic Mel spectrum and use the logarithmic Mel spectrum to resize the Mel spectrum.

[0024] Furthermore, the specific steps of calculating the absorption spectrum of the spectral data are:

[0025] The time domain signals of the sample and the background are converted into frequency domain signals using Fourier transform;

[0026] The ratio of the sample signal to the background signal was calculated as the transfer function;

[0027] The absorption coefficient is calculated using the transfer function, and finally the absorption spectrum is obtained.

[0028] Furthermore, the neural network model for transfer learning of the multi-channel graph feature data set is a DenseNet201 model.

[0029] Furthermore, the DenseNet201 model is pre-trained using the visual database ImageNet.

[0030] Furthermore, by adding a Transformer module after the neural network model, the self-attention mechanism of the Transformer module is used to enhance the model's attention to key features and improve the feature representation capability.

[0031] The second aspect of the present invention provides a terahertz time-domain spectrum feature extraction and classification system for hidden objects in an atmospheric environment, the system comprising: a memory, a processor, the memory comprising a terahertz time-domain spectrum feature extraction and classification method program for hidden objects in an atmospheric environment, the terahertz time-domain spectrum feature extraction and classification method program for hidden objects in an atmospheric environment being executed by the processor to implement the following steps:

[0032] Perform wavelet filtering on the collected terahertz time-domain spectrum data of hidden objects to obtain a denoised spectrum data set;

[0033] Use deep learning models to perform data enhancement on the denoised spectral dataset to form an expanded dataset;

[0034] Preprocessing the expanded data set to form a multi-channel image feature data set;

[0035] Use the neural network model to perform transfer learning on the multi-channel graph feature dataset, train the model, and use the loss function to optimize the model parameters to obtain a real-time hidden object detection model;

[0036] Use the trained detection model to perform real-time detection and classification of hidden objects.

[0037] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0038] The present invention significantly improves the ability to identify hidden objects by adopting advanced data enhancement and multi-channel processing technology, increases the diversity and pertinence of generated data, and enhances the stability of the training process, thus showing significant advantages in processing complex tasks and data sets, especially in different humidity environments. The detection accuracy has been significantly improved. In addition, through transfer learning with an improved neural network model, the key features in the terahertz spectrum are fully extracted and distinguished, which significantly improves the model's feature extraction ability and classification performance on terahertz time-domain spectral data, and shows excellent performance under the training conditions of a limited data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to make the purpose and technical solution of the present invention clearer, the present invention provides the following drawings and descriptions:

[0040] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0041] Figure 2 A TCGN model diagram provided by an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of a multi-channel image feature data set construction process provided by an embodiment of the present invention;

[0043] Figure 4 A structural diagram of a transfer learning module provided in an embodiment of the present invention;

[0044] Figure 5 A schematic diagram of a TransformerBlock module provided in an embodiment of the present invention;

[0045] Figure 6 This is a structural diagram of the DenseNet201-Transformer module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0048] The present invention provides a method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment. Figure 1 The figure shows a flow chart of a method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment. The specific steps are as follows:

[0049] S1: Perform wavelet filtering on the collected terahertz time-domain spectrum data of hidden objects to obtain a denoised spectrum data set;

[0050] In a specific example, different labels are assigned to the collected items according to their categories, which are replaced by numbers. For example, the one-dimensional terahertz time-domain spectral data of [alcohol, kerosene, cooking oil, frankincense oil, turpentine, ...] is combined with the label [0, 1, 2, 3, 4, ...] after wavelet filtering and denoising to form the labeled terahertz time-domain data, which is convenient for training.

[0051] S2: Use the deep learning model to perform data enhancement on the denoised spectral dataset to form an expanded dataset;

[0052] More specifically, the deep learning model for data enhancement of spectral data is the TCGN target-driven conditional generation model, which adds a target conditional generation network to the generator and discriminator, and can better control the specific properties of the generated data. This design enables TCGN to generate synthetic data that better meets the needs, improving the diversity and specificity of the generated data. It can also better adjust the training process of the generator and discriminator through the feedback mechanism of the target extractor, reducing the uncontrollability in training. This feedback mechanism helps the generator learn to generate realistic data faster, thereby improving training efficiency. The TCGN model is shown in the figure Figure 2 shown.

[0053] S3: preprocessing the expanded data set to form a multi-channel image feature data set;

[0054] More specifically, the preprocessing of the collected spectral data includes calculating a wavelet spectrum, a Mel spectrum and an absorption spectrum of the spectral data.

[0055] More specifically, the specific steps of calculating the wavelet spectrum of spectral data are:

[0056] For the input signal x(t), discrete wavelet transform (SWT) is used for decomposition. The basic formula of wavelet transform is:

[0057]

[0058] Among them, ψ j,k (t) is the scaling function and wavelet function, j and k are the summation indices, j represents the level, and k represents the translation. j,k is the wavelet coefficient.

[0059] In the first-level decomposition, the approximation coefficient cA and detail coefficient cD are obtained, and the formula is:

[0060] cA=SWT approx (x(t))

[0061] cD=SWT detail (x(t))

[0062] Among them, SWT approx () represents the discrete wavelet transform operation used to obtain approximate coefficients, SWT detail () represents the discrete wavelet transform operation used to obtain detail coefficients.

[0063] The coefficients are interpolated and resampled using an interpolation function to ensure that the data matches the target shape. The formula of the interpolation function is:

[0064]

[0065] Among them, x is the x coordinate of the interpolation target point, i and j are the indexes of the original data points, and x i and x j is the x-coordinate value of the known data point, y j is the y-coordinate value of the known data point, and the interpolation result f interp (x) is used for resampling.

[0066] The approximation coefficients and detail coefficients are resampled using an interpolation function to ensure that the data matches the target shape, as given by:

[0067]

[0068] Among them, new_x is a sequence of equidistant points of the target shape.

[0069] The resampled approximation coefficients and detail coefficients are combined into a two-dimensional matrix as the final wavelet spectrum. The formula is:

[0070]

[0071] More specifically, the steps of calculating the Mel spectrum of the spectral data are as follows:

[0072] Convert the time domain signal into tensor form, the formula is:

[0073] clip = torch.Tensor(x(t))

[0074] Calculate and set the window length and jump length to divide the time domain signal tensor into multiple frames. The formula is:

[0075]

[0076] Wherein, sampling_rate is the sampling rate, which is set to 1900000000 in this embodiment, scale is a scaling factor used to adjust the window size and hop size according to the sampling rate, which is set to 100000000, and hop_size is the hop size, which is set to 10.

[0077] Use the Mel spectrum conversion function to convert the framed signal tensor into Mel spectrum representation. The formula is:

[0078] spec=MelSpectrogram(sample_rat=sampling_rate,

[0079] n fft =n fft ,

[0080] win length =window length ,

[0081] hop length =hop length ,

[0082] n_mels = n_mels)(clip)

[0083] Among them, MelSpectrogram() is a function that converts a signal into a Mel spectrum. fft is the window size to use for Fourier transform, set to 8192. n_mels refers to the number of mel frequency bands to generate, set to 128.

[0084] Calculate the logarithmic Mel spectrum, the formula is:

[0085] spec_log=log(spec+∈)

[0086] Here, ∈ is a small value set to avoid zero problems in logarithmic calculations, which is set to 1e-6.

[0087] The logarithmic Mel spectrum is used to adjust the size of the Mel spectrum. The formula is:

[0088] spec_resized=Resize((target_height,target_width))(spec_log)

[0089] Among them, target_height and target_width are set to 128 and 250 respectively. These parameters are used to adjust the data to a size suitable for transfer learning model training.

[0090] More specifically, the specific steps of calculating the absorption spectrum of spectral data are:

[0091] Fourier transform is used to convert the time domain signals of the sample and background into frequency domain signals. The formula is:

[0092]

[0093] Where i represents the imaginary unit and w is the angular frequency, which depends on the frequency of the signal. s (t) is the time domain signal of the sample, E r (t) is the time domain signal of the background. Through Fourier transform, we get their corresponding frequency domain signal E s (t) and E r (t).

[0094] The ratio of the sample signal to the background signal is calculated as the transfer function, and the formula is:

[0095]

[0096] The absorption coefficient is calculated using the transfer function, and the absorption spectrum is finally obtained. The formula is:

[0097]

[0098] The thickness of the sample is d, which is set to 0.01.

[0099] The extended dataset after TCGN data enhancement is used to construct a multi-channel image feature dataset with unique features through signal preprocessing. The construction process is as follows: Figure 3 As shown, the first picture on the left represents the terahertz time-domain spectrum data; the second picture represents the wavelet spectrum, Mel spectrum, and absorption spectrum obtained by signal processing; the third picture represents the final characteristic uniqueness multi-channel image feature data set.

[0100] S4: Use the neural network model to perform transfer learning on the multi-channel graph feature dataset, train the model, and use the loss function to optimize the model parameters to obtain the hidden object real-time detection model;

[0101] Transfer learning can provide faster training speed and better performance under the condition of limited data sets. Specifically, the neural network model used is DenseNet201, which is pre-trained based on the visual database ImageNet. Since terahertz time-domain spectroscopy data is usually scarce, direct training from scratch can easily lead to overfitting, while transfer learning can also perform well with less data, reducing the dependence on large-scale labeled data. The pre-trained DenseNet201 has learned effective feature representations, and transfer learning only needs to fine-tune the model, thereby reducing training time and computing resources and accelerating the convergence of the model. The pre-trained model has been optimized on a large-scale data set and has good robustness and generalization capabilities. Transfer learning uses these weights as a starting point and significantly improves the performance of the model in terahertz time-domain spectroscopy data processing. DenseNet201 enhances feature propagation and reuse through dense connections. This structural advantage helps to more effectively capture and utilize the complex features in terahertz time-domain spectroscopy data, further improving model performance. The module structure of transfer learning is as follows: Figure 4 shown.

[0102] In a specific embodiment, the multi-channel graph feature dataset converted to obtain feature uniqueness is input into the DenseNet201 network for transfer learning. The first convolutional layer size is 3×3, which is used to extract initial features. Then comes the core part of DenseNet, Dense Block, which consists of multiple dense blocks. Each dense block contains a series of convolutional layers, and the inputs of these layers include the feature maps of all previous layers in the block. In DenseNet201, there are four dense blocks, containing 6, 12, 48 and 32 convolutional layers respectively. The connections between these layers are dense, so that the features can be propagated and reused within the block. Transition layers are inserted between dense blocks to reduce the size of feature maps. The transition layer usually includes a 1x1 convolutional layer and an average pooling layer. This helps control the number of parameters and speed up training. After the last dense block, global average pooling is applied to generate a fixed-size feature vector. Finally, the feature vector after global average pooling is input to the classification layer for image classification tasks. The category loss function uses multi-class cross entropy loss calculation, which combines the softmax activation function and the cross entropy loss function for multi-classification tasks, and can effectively calculate the gradient and perform back propagation. This helps the model converge faster during training, and the gradient will not disappear or explode.

[0103] The formula for multi-class cross entropy loss is as follows:

[0104]

[0105] Among them, L is the loss value, N is the number of samples, set to 19500, C is the number of categories, set to 13, y ijis the true label of the i-th sample in the j-th category, is the predicted probability of the i-th sample in the j-th class.

[0106] Furthermore, by adding a Transformer module after the neural network model, the self-attention mechanism of the Transformer module is used to enhance the model's attention to key features and improve the feature representation capability.

[0107] The Transformer Block module captures and enhances feature relationships through the self-attention mechanism Multi-Head Attention and the Feedforward Neural Network, enabling the model to more effectively extract and utilize time series features, especially weak signal features in terahertz spectral data. The schematic diagram of the Transformer Block module is shown in the figure. Figure 5 After adding the Transformer Block module to DenseNet201, the DenseNet201-Transformer model is obtained, as shown in Figure 6 As shown in the figure. This model uses two Transformer Block modules in succession after DenseNet201, giving full play to the advantages of Transformer in feature extraction and global dependency capture, enhancing the model's attention to and representation of key features, and improving the model's extraction and utilization efficiency of time series features through the sequence attention mechanism, thereby improving the recognition accuracy and stability of hidden objects. In the DenseNet201-Transformer model, the DenseNet201 part is mainly responsible for extracting the multi-scale spatial features of the input data, while the Transformer part further captures global dependencies and time series features on this basis. This combination can fully utilize the respective advantages of DenseNet201 and Transformer, thereby improving the performance of the model in complex tasks.

[0108] More specifically, an early stopping strategy is adopted in the process of training the model. The early stopping strategy is an effective method to improve the generalization ability of the model, which helps to balance the bias and variance of the model during the training process and prevent overfitting.

[0109] S5: Use the trained detection model to perform real-time detection and classification of hidden objects.

[0110] The following experiment compares multiple denoising methods to evaluate their impact on model performance. The evaluation methods include no denoising, normalization, standardization, mean filtering, median filtering, SNV filtering and wavelet filtering, all of which are existing technologies. These denoising methods are applied to three existing feature combinations Mel_Stft_Mfcc, Mel_Stft_Mfcc, Wavelet_Mel_Absorption and three models DenseNet201, ResNet, Inception, forming a total of nine combinations. Then, the results of these nine combinations are averaged as the performance indicators of each denoising method. The average performance indicators of each denoising method are shown in Table 1.

[0111] Table 1

[0112] DenoisingMethods Avg_acc Ava_loss Avg_pre Avg_recall Avg_F1 nodenoising 85.41％ 0.075 86.12％ 86.34％ 86.8％ normalization 90.48％ 0.023 91.23％ 91.54％ 91.11％ meancentralization 91.76％ 0.019 92.12％ 92.33％ 91.82％ meanfiltering 91.44％ 0.022 91.68％ 91.87％ 91.43％ medianfiltering 92.59％ 0.015 92.88％ 93.1％ 92.7％ SNV filtering 91.49％ 0.021 91.72％ 91.92％ 91.53％ waveletfiltering 94.92％ 0.012 95.21％ 95.32％ 94.9％

[0113] Avg_acc, Avg_loss, Avg_pre, Avg_reacll, and Avg_F1 represent the average accuracy, average loss, average precision, average recall, and average F1 score, respectively. From the results, we can see that wavelet filtering performs best, so this method is selected as the denoising method in subsequent experiments.

[0114] Next, different feature combinations are evaluated to determine the optimal feature combination for the classification task. The tested feature combinations include time domain, frequency domain, Stft, Mfcc, Mel spec, Wavelet spec, Absorption spec, Mel_Stft_Mfcc, Wavelet_Mel_Mfcc, and Wavelet_Mel_Absorption. These features are used to train DenseNet201, ResNet, and Inception models, and the average performance indicators are shown in Table 2.

[0115] Table 2

[0116] Features Avg_acc Ava_loss Avg_pre Avg_recall Avg_F1 Timedomain 84.64％ 0.062 85.23％ 85.42％ 85.77％ Frequencydomain 84.82％ 0.052 85.45％ 85.71％ 86.1％ Stft 86.96％ 0.048 86.2％ 86.42％ 86.71％ Mfcc 89.35％ 0.032 90.8％ 91.12％ 91.33％ Melspec 92.21％ 0.025 92.52％ 92.71％ 92.93％ Waveletspec 93.47％ 0.022 93.82％ 94.11％ 94.34％ Absorption spec 94.46％ 0.015 94.71％ 95.02％ 95.4％ Mel_Stft_Mfcc 94.92％ 0.012 95.12％ 95.3％ 95.62％ Wavelet_MeL_Mfcc 95.76％ 0.008 96.2％ 96.52％ 96.81％ Wavelet_MeL_Absorption 96.8％ 0.005 97.31％ 97.5％ 97.83％

[0117] From the results, it can be seen that the feature combination Wavelet_Mel_Absorption of wavelet spectrum, Mel spectrum and absorption spectrum used in this application performs best and is adopted in subsequent model comparison experiments.

[0118] Finally, the performance of different models is compared using the wavelet filtering denoising method combined with the Wavelet_Mel_Absorption feature combination. The compared models include ResNet, ResNeXt, EfficientNet, Inception, DenseNet64, DenseNet201, DenseNet201_SE_CBAM, and DenseNet201_Transformer. The performance indicators of each model are shown in Table 3.

[0119] Table 3

[0120] Models Acc Loss Pre Recall F1 ResNet 94.62％ 0.002 95.23％ 95.42％ 95.77％ ResNeXt 94.81％ 0.002 95.45％ 95.71％ 96.1％ EfficientNet 92.93％ 0.038 93.2％ 93.42％ 93.71％ Inception 93.55％ 0.0028 93.8％ 94.12％ 94.33％ DenseNet64 95.26％ 0.017 95.52％ 95.71％ 95.93％ DenseNet201 97.67％ 0.01 97.82％ 98.11％ 98.34％ DenseNet201_SE_CBAM 97.44％ 0.015 97.71％ 98.02％ 98.4％ DenseNet201_Transformer 98.92％ 0.001 99.12％ 99.3％ 99.62％

[0121] From the results, it can be seen that the DenseNet201_Transformer model used in this application shows the highest performance, verifying its effectiveness in capturing local and global features in complex classification tasks.

[0122] The second aspect of the present invention provides a terahertz time-domain spectrum feature extraction and classification system for hidden objects in an atmospheric environment, the system comprising: a memory, a processor, the memory comprising a terahertz time-domain spectrum feature extraction and classification method program for hidden objects in an atmospheric environment, the terahertz time-domain spectrum feature extraction and classification method program for hidden objects in an atmospheric environment being executed by the processor to implement the following steps:

[0123] Perform wavelet filtering on the collected terahertz time-domain spectrum data of hidden objects to obtain a denoised spectrum data set;

[0124] Use deep learning models to perform data enhancement on the denoised spectral dataset to form an expanded dataset;

[0125] Preprocessing the expanded data set to form a multi-channel image feature data set;

[0126] Use the neural network model to perform transfer learning on the multi-channel graph feature dataset, train the model, and use the loss function to optimize the model parameters to obtain a real-time hidden object detection model;

[0127] Use the trained detection model to perform real-time detection and classification of hidden objects.

[0128] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment, characterized in that: The steps include: Perform wavelet filtering on the collected terahertz time-domain spectrum data of hidden objects to obtain a denoised spectrum data set; Use deep learning models to perform data enhancement on the denoised spectral dataset to form an expanded dataset; Preprocessing the expanded data set to form a multi-channel image feature data set; Use the neural network model to perform transfer learning on the multi-channel graph feature dataset, train the model, and use the loss function to optimize the model parameters to obtain a real-time hidden object detection model; Use the trained detection model to perform real-time detection and classification of hidden objects.

2. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 1, characterized in that: The deep learning model for data enhancement of the denoised spectral data set is a target driving condition generation model, and the target driving condition generation model adds a target condition generation network on the basis of the generator and the discriminator.

3. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 1, characterized in that: The preprocessing of the data-enhanced spectral data set includes calculating a wavelet spectrum, a Mel spectrum and an absorption spectrum of the spectral data.

4. The method for extracting and classifying terahertz time-domain spectroscopy features of hidden objects in an atmospheric environment according to claim 3, characterized in that: The specific steps of calculating the wavelet spectrogram of spectral data are: Perform discrete wavelet transform on the input signal to obtain the approximation coefficient and detail coefficient of the first-level decomposition; The coefficients are interpolated and resampled using an interpolation function to ensure that the data matches the target shape; The resampled approximation coefficients and detail coefficients are combined into a two-dimensional matrix as the final wavelet spectrum.

5. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 3, characterized in that: The specific steps of calculating the Mel spectrum of spectral data are as follows: Convert the time domain signal into tensor form; Calculate and set the window length and jump length to split the time domain signal tensor into multiple frames; Use the Mel spectrum conversion function to convert the framed signal tensor into Mel spectrum representation; Calculate the logarithmic Mel spectrum and use the logarithmic Mel spectrum to resize the Mel spectrum.

6. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 3, characterized in that: The specific steps of calculating the absorption spectrum of the spectral data are: The time domain signals of the sample and the background are converted into frequency domain signals using Fourier transform; The ratio of the sample signal to the background signal was calculated as the transfer function; The absorption coefficient is calculated using the transfer function, and finally the absorption spectrum is obtained.

7. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 1, characterized in that: The neural network model for transfer learning of the multi-channel graph feature data set is the DenseNet201 model.

8. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 7, characterized in that: The DenseNet201 model is pre-trained using the visual database ImageNet.

9. The method for extracting and classifying terahertz time-domain spectral features of hidden objects in an atmospheric environment according to claim 1, characterized in that: By adding a Transformer module after the neural network model, the self-attention mechanism of the Transformer module is used to enhance the model's attention to key features and improve the feature representation capability.

10. A terahertz time-domain spectrum feature extraction and classification system for hidden objects in an atmospheric environment, characterized in that: The system includes: a memory and a processor, wherein the memory includes a program for extracting and classifying the terahertz time-domain spectrum features of hidden objects in an atmospheric environment. When the program for extracting and classifying the terahertz time-domain spectrum features of hidden objects in an atmospheric environment is executed by the processor, the steps of the method for extracting and classifying the terahertz time-domain spectrum features of hidden objects in an atmospheric environment as described in any one of claims 1 to 9 are implemented.

Images