A deep learning-based terahertz three-dimensional high-resolution imaging method

Abstract

This invention discloses a deep learning-based terahertz three-dimensional high-resolution imaging method, comprising the following steps: simulating terahertz signals using the transfer matrix method, and forming a terahertz dataset based on the impulse response sequence of the simulated signal and the simulated terahertz time-domain spectral signal; constructing a UNet-BiLSTM network architecture model, and training the UNet-BiLSTM network architecture using the terahertz dataset formed in step one; acquiring terahertz three-dimensional data using a terahertz time-domain spectral system, and inputting the acquired UNet-BiLSTM network architecture from step two to obtain impulse response sequences of terahertz time series at different locations; reconstructing the terahertz three-dimensional data using the impulse response sequences of the terahertz time series at different locations, and performing three-dimensional imaging on the reconstructed terahertz three-dimensional data using a tomographic imaging method. This invention can achieve high-precision feature extraction and three-dimensional imaging of terahertz signals.

Description

Technical Field

[0001] This invention belongs to the field of terahertz nondestructive testing, specifically relating to a deep learning-based terahertz three-dimensional high-resolution imaging method. Background Technology

[0002] Terahertz waves (THz, frequency range 0.1THz–10THz) overlap with millimeter waves in the long-wavelength band and with infrared light in the short-wavelength band. They possess fingerprint-like spectral characteristics, penetrability, and low energy, and are widely used in materials analysis, aerospace, medicine, and atmospheric remote sensing. Terahertz imaging first uses a terahertz transmitter to generate terahertz waves, then illuminates the object under test with these waves. The object will produce different responses to the terahertz waves. The spatial distribution of the material's complex permittivity is obtained based on the intensity and phase information contained in these responses. Visualizing terahertz data through three-dimensional visualization technology can effectively lower the technical threshold of terahertz nondestructive testing, which is crucial for the field.

[0003] However, current terahertz 3D imaging methods mostly focus on improving the uneven light intensity distribution and low 3D imaging resolution by addressing the issues of terahertz devices and imaging algorithms (CN 107631995 B, CN 114910439 A). Fundamentally, the low resolution of terahertz images is mainly due to the difficulty in separating terahertz time-domain signal echoes, echo signal aliasing, and noise masking caused by multiple reflections of terahertz waves at interfaces, absorption of waves by the medium, and noise. Therefore, CN 116223428 A and CN 111427046 A propose a deep learning-based sparse solution method for terahertz deconvolution and a terahertz pulse echo localization method to improve detection accuracy, respectively, achieving signal resolution enhancement through sparse signal representation. However, sparse representation methods typically require hundreds of iterations to converge the least squares criterion. While optimizing the sparse solver using deep learning methods can improve efficiency, it also involves multiple iterations, which is extremely time-consuming for large-scale terahertz imaging. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a deep learning-based terahertz three-dimensional high-resolution imaging method to achieve high-precision feature extraction and three-dimensional imaging of terahertz signals. This method constructs a network architecture suitable for terahertz time-domain signals based on a U-Net structure network and a Bidirectional Long Short-Term Memory (BiLSTM) network. This network architecture enables rapid extraction of features from terahertz time-domain signals and achieves high-precision extraction of terahertz signal peaks at different interfaces without prior knowledge, thereby reconstructing terahertz three-dimensional data. Based on the peak extraction results, terahertz three-dimensional imaging can be achieved using tomographic imaging methods.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A deep learning-based terahertz three-dimensional high-resolution imaging method includes the following steps:

[0007] Step 1: Simulate terahertz signals using the transfer matrix method, and form a terahertz dataset based on the impulse response sequence of the simulated signals and the simulated terahertz time-domain spectral signals.

[0008] Step 2: Construct the UNet-BiLSTM network architecture model and train the UNet-BiLSTM network architecture using the terahertz dataset formed in Step 1.

[0009] Step 3: Acquire terahertz three-dimensional data using a terahertz time-domain spectroscopy system, input the UNet-BiLSTM network architecture trained in Step 2, and obtain impulse response sequences of terahertz time series at different locations;

[0010] Step 4: Reconstruct terahertz three-dimensional data from the impulse response sequences of terahertz time series at different locations, and perform three-dimensional imaging on the reconstructed terahertz three-dimensional data using tomographic imaging.

[0011] Further, step one includes:

[0012] S11. Simulate terahertz time-domain spectral signals under different conditions using the transfer matrix method:

[0013] Using the electromagnetic wave transmission matrix, an initial THz propagation simulation model is established in the frequency domain. The initial THz propagation simulation model is as follows:

[0014]

[0015] In the formula, P i The phase change of the incident terahertz signal is represented by the following formula: ω is the angular frequency, and c is the speed of light. D is the complex refractive index of the i-th layer;i,i+1 This describes the behavior of the terahertz signal at each interface between two media, including the corresponding Fresnel coefficients, calculated using the following formula: r i,i+1 and t i,i+1 These are the Fresnel reflection and transmission coefficients, respectively;

[0016] The reflective transfer function R(ω) is expressed as:

[0017]

[0018] The simulated terahertz time-domain signal can be obtained using the following formula:

[0019] E(t)=F -1 (R(ω)F(E ref (t)))

[0020] In the formula, F(E) ref (t) is the Fourier transform of the incident THz pulse, F -1 This is the inverse Fourier transform;

[0021] S12. Terahertz time-domain spectral signals under different conditions were simulated through step S11, and terahertz impulse response sequences under different conditions were obtained as feature information based on the simulation data; the terahertz impulse response sequences were used as prediction results and the simulation data were used as input to form a terahertz training dataset.

[0022] Furthermore, step two includes:

[0023] S21. Constructing the UNet-BiLSTM network architecture model: The UNet-BiLSTM network architecture model includes the UNet convolutional neural network and the BiLSTM network; the UNet convolutional neural network downsamples the input terahertz pulse signal and continuously extracts signal features, and then reconstructs the signal using the extracted features through upsampling and feature mapping; after obtaining the reconstructed signal, the BiLSTM network captures the contextual information in the sequence data;

[0024] S22. Input the terahertz dataset obtained in step one into the constructed UNet-BiLSTM network architecture for training, using the SGDM training method.

[0025] Furthermore, in step S21, the UNet-BiLSTM network architecture model includes a UNet convolutional neural network and a BiLSTM network. UNet is an end-to-end convolutional neural network that includes an encoding module and a decoding module. The encoding module obtains higher-level signal features from the input terahertz pulse signal by downsampling and continuously extracting signal features. Max pooling layers are added to the encoding module to identify extreme signal features. The decoding module reconstructs the signal using the features extracted by the encoding module through upsampling and feature mapping.

[0026] Furthermore, both the encoder and decoder are composed of convolutional neural networks.

[0027] Furthermore, step three includes: acquiring terahertz three-dimensional data IM using a terahertz time-domain spectroscopy system. m×n×t Where m, n, and t represent the terahertz image length, terahertz image width, and terahertz time window length, respectively; the three-dimensional data is input into the UNet-BiLSTM network architecture according to the terahertz time-domain spectral signals at different locations to obtain the impulse response sequences IM1 of the time series at different locations. m×n×t This enables high-precision extraction of peak values ​​at the sample interface.

[0028] The present invention has the following beneficial effects:

[0029] This invention simulates the propagation characteristics of terahertz waves in the measured material using the transfer matrix method. It obtains the terahertz signal peaks at different interfaces using different simulated time-domain spectral signals, and constructs a terahertz dataset based on the interface peak information and the simulated terahertz time-domain spectral signals. This enhances the interpretability, quality, and reliability of the dataset, providing a simpler method for constructing datasets.

[0030] This invention employs a UNet-BiLSTM network and a constructed dataset. The UNet network is used to extract spatial features of the data, while the BiLSTM network is used to extract temporal features. In this way, the invention can accurately identify and separate different terahertz echo signals and precisely locate the peak interfaces of terahertz signals. This provides high-precision peak information for the three-dimensional reconstruction of terahertz data. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of a deep learning-based terahertz three-dimensional high-resolution imaging method according to the present invention.

[0032] Figure 2(a) shows a terahertz analog signal at 200 μm;

[0033] Figure 2(b) shows a terahertz analog signal at 500 μm;

[0034] Figure 2(c) shows a terahertz analog signal at 700 μm;

[0035] Figure 3 This is a block diagram of the UNet-BiLSTM network module according to an embodiment of the present invention;

[0036] Figure 4 The peak value of the time-domain signal characteristic interface obtained through the network in this embodiment of the invention;

[0037] Figure 5 This is a tomographic image of reconstructed terahertz three-dimensional data in an embodiment of the present invention. Detailed Implementation

[0038] To make the purpose, technical solution, and advantages of the invention clearer, the invention will be described in further detail.

[0039] like Figure 1 As shown, this invention is a terahertz three-dimensional high-resolution imaging method based on deep learning, comprising the following steps:

[0040] Step 1: Simulate terahertz signals using the transfer matrix method, and generate a terahertz dataset based on the impulse response sequence of the simulated signals and the simulated terahertz time-domain spectral signal.

[0041] S11. Simulate terahertz time-domain spectral signals under different conditions using the transfer matrix method:

[0042] Using the electromagnetic transmission matrix (TMM), an initial THz propagation simulation model is established in the frequency domain, as shown in equation (1):

[0043]

[0044] In the formula, P i The phase change of the incident terahertz signal is represented by the following formula: ω is the angular frequency, and c is the speed of light. D is the complex refractive index of the i-th layer; i,i+1 This describes the behavior of the terahertz signal at each interface between two media, including the corresponding Fresnel coefficients, calculated using the following formula: r i,i+1 and t i,i+1 These are the Fresnel reflection and transmission coefficients, respectively.

[0045] Therefore, the reflective transfer function R(ω) can be expressed as:

[0046]

[0047] Finally, the simulated terahertz time-domain signal can be obtained using equation (3):

[0048] E(t)=F -1 (R(ω)F(E ref (t))) (3)

[0049] In the formula, F(E) ref (t) is the Fourier transform of the incident THz pulse, F -1 This is the inverse Fourier transform.

[0050] S12. Terahertz time-domain spectral signals under different conditions were simulated through step S11, and terahertz impulse response sequences under different conditions were obtained as feature information based on the simulation data; the terahertz impulse response sequences were used as prediction results and the simulation data were used as input to form a terahertz training dataset.

[0051] Step 2: Construct the UNet-BiLSTM network architecture model and train the UNet-BiLSTM network architecture using a terahertz dataset:

[0052] S21. Construct the UNet-BiLSTM network architecture model, such as Figure 3 As shown, the UNet-BiLSTM network architecture model includes a UNet convolutional neural network and a BiLSTM network. UNet is an end-to-end convolutional neural network comprising an encoder and a decoder. The encoder downsamples the input terahertz pulse signal and extracts higher-level signal features through continuous feature extraction. A max-pooling layer (MaxPoo) is added to the encoder to identify extreme signal features. The decoder reconstructs the signal using the features extracted by the encoder through upsampling and feature mapping. Both the encoder and decoder are composed of convolutional neural networks, and different network structures and parameters can be selected depending on the task and data. After obtaining the reconstructed signal through the UNet convolutional neural network, the BiLSTM network better captures the contextual information in the sequence data, improving the model's expressive power and predictive performance.

[0053] S22. Input the terahertz training dataset obtained in step one into the constructed UNet-BiLSTM network architecture for training. Set the training rounds to 200 rounds, the initial learning rate to 0.001, use the SGDM training method, and set the minimum number of samples to 64.

[0054] Step 3: Acquire terahertz three-dimensional data using a terahertz time-domain spectroscopy system, input the data into the UNet-BiLSTM network architecture from Step 2, and obtain the impulse response sequences of terahertz time series at different locations to achieve high-precision extraction of peak values ​​at the sample interface.

[0055] Terahertz three-dimensional data (IM) acquired using a terahertz time-domain spectroscopy systemm×n×t Where m, n, and t represent the terahertz image length (in mm), terahertz image width (in mm), and terahertz time window length (in ps), respectively. The three-dimensional data, with terahertz time-domain spectral signals at different locations, are input into the UNet-BiLSTM network architecture to obtain the impulse response sequences IM1 for time series at different locations. m×n×t This enables high-precision extraction of peak values ​​at the sample interface.

[0056] Step 4: Reconstruct terahertz three-dimensional data from the impulse response sequences of terahertz time series at different locations, and perform three-dimensional imaging on the reconstructed terahertz three-dimensional data using tomographic imaging.

[0057] Example:

[0058] Taking a sample with debonding defects as an example, the implementation process of the technical solution of the present invention is introduced:

[0059] Step 1: Use the transfer matrix method to simulate the reflection signal of the terahertz wave after propagation in the sample. Here, simulated signals of debonding defects with different thicknesses are simulated. Three of them are selected as shown in Figure 2(a), Figure 2(b) and Figure 2(c). A terahertz dataset is formed based on the impulse response sequence of the simulated signal and the simulated terahertz time-domain spectral signal.

[0060] Step 2: Construct the UNet-BiLSTM network architecture model, such as... Figure 3 As shown, the UNet-BiLSTM network architecture is trained using a terahertz dataset;

[0061] Step 3: Acquire terahertz three-dimensional data using a terahertz time-domain spectroscopy system, input the data into the aforementioned network architecture to obtain impulse response sequences of terahertz time series at different locations, such as... Figure 4 The image shows the terahertz time-domain signal and impulse response sequence at a certain location, where Layer1 and Layer2 are the upper and lower interfaces of the defect to be identified, respectively, to achieve high-precision extraction of the peak value of the sample defect interface.

[0062] Step 4: Reconstruct terahertz 3D data from the impulse response sequences of terahertz time series at different locations. Use tomographic imaging to perform 3D imaging on the reconstructed terahertz 3D data to obtain, for example... Figure 5 The three-dimensional imaging results are shown.

[0063] The above description is only a preferred embodiment of the present invention. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the present invention. The content of this specification should not be construed as a limitation of the present invention.

Claims

1. A terahertz three-dimensional high-resolution imaging method based on deep learning, characterized in that, Includes the following steps: Step 1: Simulate terahertz signals using the transfer matrix method, and form a terahertz dataset based on the impulse response sequence of the simulated signals and the simulated terahertz time-domain spectral signal; Step 1 includes: S11. Simulate terahertz time-domain spectral signals under different conditions using the transfer matrix method: Using the electromagnetic wave transmission matrix, an initial THz propagation simulation model is established in the frequency domain. The initial THz propagation simulation model is as follows: ； In the formula, The phase change of the incident terahertz signal is represented by the following formula: , Angular frequency, At the speed of light, For the first Complex refractive index of the layer; This describes the behavior of the terahertz signal at each interface between two media, including the corresponding Fresnel coefficients, calculated using the following formula: , and These are the Fresnel reflection and transmission coefficients, respectively; Reflective transfer function Represented as: ； The simulated terahertz time-domain signal can be obtained using the following formula: ； In the formula, For the Fourier transform of the incident THz pulse, This is the inverse Fourier transform; S12. Terahertz time-domain spectral signals under different conditions were simulated through step S11, and terahertz impulse response sequences under different conditions were obtained as feature information based on the simulation data; the terahertz impulse response sequences were used as prediction results and the simulation data were used as input to form a terahertz training dataset. Step 2: Construct the UNet-BiLSTM network architecture model and train the UNet-BiLSTM network architecture using the terahertz dataset obtained in Step 1; Step 2 includes: S21. Constructing the UNet-BiLSTM network architecture model: The UNet-BiLSTM network architecture model includes the UNet convolutional neural network and the BiLSTM network; the UNet convolutional neural network downsamples the input terahertz pulse signal and continuously extracts signal features, and then reconstructs the signal using the extracted features through upsampling and feature mapping; after obtaining the reconstructed signal, the BiLSTM network captures the contextual information in the sequence data; S22. Input the terahertz dataset obtained in step one into the constructed UNet-BiLSTM network architecture for training, using the SGDM training method; Step 3: Acquire terahertz three-dimensional data using a terahertz time-domain spectroscopy system, input the UNet-BiLSTM network architecture trained in Step 2, and obtain impulse response sequences of terahertz time series at different locations; Step 4: Reconstruct terahertz three-dimensional data from the impulse response sequences of terahertz time series at different locations, and perform three-dimensional imaging on the reconstructed terahertz three-dimensional data using tomographic imaging.

2. The terahertz three-dimensional high-resolution imaging method based on deep learning as described in claim 1, characterized in that, In step S21, the UNet-BiLSTM network architecture model includes a UNet convolutional neural network and a BiLSTM network. UNet is an end-to-end convolutional neural network, which includes an encoding module and a decoding module. The encoding module obtains higher-level signal features from the input terahertz pulse signal by downsampling and continuously extracting signal features. Max pooling layers are added to the encoding module to identify extreme signal features. The decoding module reconstructs the signal using the features extracted by the encoding module through upsampling and feature mapping.

3. The terahertz three-dimensional high-resolution imaging method based on deep learning as described in claim 2, characterized in that, Both the encoding and decoding modules are composed of convolutional neural networks.

4. The terahertz three-dimensional high-resolution imaging method based on deep learning as described in claim 1, characterized in that, Step three includes: acquiring terahertz three-dimensional data using a terahertz time-domain spectroscopy system. ,in These represent the terahertz image length, terahertz image width, and terahertz time window length, respectively. The three-dimensional data is input into the UNet-BiLSTM network architecture according to the terahertz time-domain spectral signals at different locations to obtain impulse response sequences of different time series. This enables high-precision extraction of peak values ​​at the sample interface.

Images