Flexible and robust phase-free terahertz aperture-coded three-dimensional imaging method and device
The phase-free terahertz aperture-coded imaging method addresses limitations of traditional methods by using distance domain slicing and a flexible deep network to iteratively solve an optimization model, achieving high-resolution and efficient reconstruction of random targets.
Patent Information
- Application Number
- CN202411273200.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-12
- Publication Date
- 2024-12-27
- Estimated Expiration
- 2044-09-12
AI Technical Summary
Existing terahertz aperture-coded three-dimensional imaging technologies are constrained by generative models that limit target reconstruction to within their data range and do not account for system errors, leading to inefficiencies and limitations in reconstructing random targets.
A phase-free terahertz aperture-coded three-dimensional imaging method using distance domain slicing, a detection array, and a flexible deep network to iteratively solve an optimization model with an alternating direction multiplier method, reducing computational complexity and system costs while enhancing robustness and flexibility.
Enables high-resolution reconstruction of random targets with reduced computational burden and system complexity, improving imaging efficiency and flexibility beyond the scope of traditional generative models.
Smart Images
Figure CN118859200B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of radar forward-looking imaging technology, and in particular to a flexible and robust phase-free terahertz aperture-coded three-dimensional imaging method and device. Background Art
[0002] With the rapid development of science and technology and the continuous maturity of radar imaging technology, high-resolution radar imaging technology has an urgent application demand in military and civilian fields such as seeker terminal guidance, proximity fuze, safety detection and non-destructive testing. Traditional high-resolution radar imaging technology relies on the relative motion between the radar and the target, and is constrained by the relative motion between the radar and the target, making it difficult to achieve high-resolution imaging under forward-looking and staring conditions. Real aperture array radar or phased array radar requires a large-scale array to obtain azimuth resolution. Although it avoids the phase error caused by target motion, its structure is complex, the volume is large, and the manufacturing cost is high. In summary, in order to meet the needs of practical applications and make up for the shortcomings of traditional radar imaging, it is urgent to develop a high-resolution forward-looking imaging technology that does not rely on the Doppler effect.
[0003] Terahertz (0.1-10THz) aperture coded imaging is a new radar imaging technology proposed in recent years. It draws on the ideas of optical aperture coded imaging, microwave correlation imaging and computational imaging. It uses metamaterial aperture coded antennas to modulate the amplitude or phase of the incident terahertz wave, forming a spatial and temporal independent radiation field distribution in the target area, and then receives the echo signal in a coherent or incoherent manner, and then uses computational imaging to extract and decouple the target information in the beam, ultimately achieving high-resolution, forward-looking, staring, and all-day imaging. In terms of imaging principles, it no longer relies on large-scale antenna arrays and the relative rotation of the radar and the target to obtain high resolution in azimuth, which not only reduces the complexity of the system, but also avoids the problems caused by "motion imaging". In terms of imaging frequency band, the wavelength of terahertz waves is between light waves and microwaves, and has stronger penetration than light waves. Compared with microwaves, its wavelength is shorter and its bandwidth is larger, and the short wavelength can bring higher angular resolution, and the large bandwidth can provide more abundant coding modulation modes for aperture coded imaging. From the perspective of the imaging system, terahertz devices are easy to miniaturize, and aperture coded imaging uses metamaterial antennas to achieve phase or amplitude modulation, avoiding the problem of multi-channel synchronization. The system is simple and easy to array and integrate.
[0004] In terahertz aperture-coded three-dimensional imaging, the current research mainly focuses on phase imaging based on coherent reception. It generally divides the imaging space into grids and uses "distance-delay" deduction analysis to establish an imaging model. In order to reduce computational complexity and improve imaging efficiency, pulse compression is currently used to convert complex terahertz aperture-coded three-dimensional imaging into two-dimensional imaging. Although computational redundancy is greatly reduced, in the terahertz frequency band, the imaging grid is finely divided, the imaging equation is large in scale, and the burden of solving the target scattering coefficient is still heavy. Recently, there is a terahertz aperture-coded imaging solution based on a generative model, which can be applied to both phase imaging and non-phase imaging. It mainly uses generative adversarial networks for learning, learns prior information from data, and uses prior information to reduce the demand for echo data for target reconstruction, thereby reducing the scale of the imaging equation and the complexity of solving the target scattering coefficient. Under compressed measurement, it realizes the parsing of target information from echo signals with noise and clutter.
[0005] However, this scheme benefits from the generative prior, but is also constrained by it. It cannot reconstruct targets outside the range of the generative model, which limits its scope of application. In actual terahertz aperture coded imaging, the target to be solved has strong randomness, and the generative model in this scheme can only generate new data similar to the real data based on the latent variables. It does not consider more targets outside the data set, and it is difficult to include all targets in the real data set. This shortcoming makes this technical solution only suitable for the reconstruction of targets within the generation range, and it is powerless to reconstruct random targets in actual terahertz aperture coded imaging scenarios. In addition, this scheme does not take into account the impact of terahertz aperture coded imaging system errors on imaging performance.
[0006] Therefore, there is an urgent need for a method and device with greater flexibility, robustness and imaging efficiency to achieve fast and high-resolution three-dimensional imaging. Summary of the invention
[0007] In view of the deficiencies in the prior art, the present invention proposes a flexible and robust phase-free terahertz aperture coded three-dimensional imaging method and device to achieve high-resolution reconstruction of three-dimensional targets.
[0008] A flexible and robust phase-free terahertz aperture-coded three-dimensional imaging method comprises the following steps:
[0009] Step 1: Construct a phase-free terahertz aperture-coded mathematical imaging model based on distance domain slicing and detection array.
[0010] Step 1.1, build a terahertz aperture coded imaging system, use the detection array in the terahertz aperture coded imaging system to obtain echo signals, perform pulse compression processing on the echo signals to obtain a one-dimensional range image, extract and separate the range units with impulse responses according to the one-dimensional range image, and divide the three-dimensional imaging area into multiple two-dimensional imaging areas in different distance domains;
[0011] Step 1.2, in the two-dimensional imaging area, obtain a reference signal matrix at the range domain slice by measurement and calculation, and construct a phase-free terahertz aperture coded mathematical imaging model based on the range domain slice and the detection array using the reference signal matrix and the range domain echo signal;
[0012] Step 2, combining the denoising mechanism with the phase-free terahertz aperture coded mathematical imaging model to obtain a phase-free terahertz aperture coded imaging optimization model based on a depth denoising prior and a detection array;
[0013] Step 3, using an alternating direction multiplier method to iteratively solve the phase-free terahertz aperture coded imaging optimization model, and adding a pre-trained flexible deep network as a submodule of the alternating direction descent multiplier algorithm during the solving process;
[0014] Step 4: Solve the target scattering information in azimuth and elevation for all range domain slices, and achieve fast high-resolution reconstruction of the three-dimensional target through joint reconstruction.
[0015] Preferably, in step 1.2, the phase-free terahertz aperture coding mathematical imaging model based on the detection array at a certain distance domain slice r is:
[0016] (1)
[0017] in, is the target scattering coefficient vector to be solved in the range domain r, is the interference noise vector received by the echo signal during the M detection process, and is the matrix of the echo signal and reference signal received by the detection array M times. For the mth sampling reception, the echo intensity received by the entire detection array is It can be expressed as:
[0018] (2)
[0019] in, is the echo vector detected by H detection array elements, is the corresponding reference signal matrix, , It is the reference signal corresponding to the nth grid unit and the hth detection array element.
[0020] The modeling method for the reference signal matrix is as follows:
[0021] First, through a small number of known targets Build the target dataset Then, different targets are used to obtain different echo vectors through the imaging system, and the echo data set matrix can be obtained by combining them. , then the mathematical expression for modeling the reference signal matrix can be written as
[0022] (3)
[0023] in, represents transpose, is the echo data set matrix The mth column of Represents the matrix to be modeled in the distance domain r The mth column of Representation Matrix The mth column of , by solving formula (3) through the inverse problem solving algorithm, we can get the mth row of the actual reference signal matrix. After M cycles, we can get the reference signal matrix at the distance domain r .
[0024] Preferably, in step 2, the phase-free terahertz aperture coded imaging optimization model based on the depth denoising prior and the detection array is expressed as follows:
[0025] (4)
[0026] in, is the target scattering coefficient to be solved in the distance domain r, N is the number of target grids, is the echo signal vector obtained by the entire detection array after M samplings, is the reference signal matrix in the distance domain r obtained by measurement and computational modeling, is a priori term learned through flexible deep network optimization, It is a preset very small number.
[0027] Preferably, the flexible deep network adopts a denoising convolutional neural network, which mainly includes three parts, the first part consists of a convolutional layer and a non-linear activation layer, the second part consists of a convolutional layer, batch normalization and a non-linear activation layer, and the third part is a convolutional layer.
[0028] Preferably, step 2 further includes acquiring a training data set, and using the training data set to train a flexible deep network, wherein the training data set includes a noise-free target data set and a noisy target data set solved by a phase-free terahertz aperture coded mathematical imaging model.
[0029] Preferably, the training method of the flexible deep network is:
[0030] Inputting the noisy target data into the flexible deep network to obtain a residual image;
[0031] Subtracting the noisy target data from the residual image to obtain a denoised target;
[0032] A reconstruction loss is obtained by calculating a mean square error according to the denoising target and the noise-free target;
[0033] The weight parameters of the denoising convolutional neural network are continuously learned from the training data set according to the reconstruction loss, and finally a trained flexible deep network is obtained.
[0034] Preferably, in step 3, when solving the terahertz aperture coded imaging optimization model, the optimization problem needs to be converted into a multivariable augmented Lagrangian function as shown in the following formula, and a relaxation variable is added to the terahertz aperture coded imaging optimization model: and dual variables , the expression is as follows:
[0035] (5)
[0036] In the above formula, is the scattering coefficient vector of the two-dimensional target to be solved in the range domain r, is a priori term learned through flexible deep network optimization, is the echo intensity information vector obtained by the entire detection array after M samplings, is the reference signal matrix at the distance domain r.
[0037] Preferably, in step 3, the iterative solution of formula (5) by the alternating direction multiplier method mainly includes the solution of three variables. First, fix and , using the gradient descent algorithm to update the slack variables , the optimization function can be expressed as:
[0038] (6)
[0039] in, is the estimated value at step t+1 of the optimization iteration, , , is the estimated value at step t, is the echo intensity signal vector obtained by the entire detection array after M samplings, is the reference signal matrix at the distance domain r obtained by the measurement fusion modeling method, and the updated relaxation variable can be obtained by the gradient descent method. ;
[0040] Then fix and , using the pre-trained flexible deep network to approximate the solution of the optimization problem, the optimization function can be expressed as:
[0041] (7)
[0042] in, and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, It is a priori term learned through flexible deep network optimization. According to the viewpoint of maximum likelihood estimation of posterior distribution, a denoising convolutional neural network with good denoising effect is used to approximate the solution of the above formula without complex optimization iterations.
[0043] fixed and , update the dual variable :
[0044] (8)
[0045] in, , and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, parameter .
[0046] Preferably, the deep denoising network is used to approximate The update optimization function is as follows:
[0047] (9)
[0048] in, and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, is the deep denoising network obtained by the training, is the denoising coefficient to be set.
[0049] The present invention also provides a device for realizing a flexible and robust phase-free terahertz aperture coded three-dimensional imaging method, comprising:
[0050] An information acquisition module based on a detection array is used to obtain echo signals, simplify the complexity of the imaging system, reduce the number of encoding and sampling of the imaging system, and reduce the refresh rate requirements of the metamaterial coding antenna in the imaging system;
[0051] The range domain slicing module is used to convert complex three-dimensional range images into two-dimensional range images in different distance domains;
[0052] A reference signal matrix modeling module is used to achieve high-precision reference signal matrix modeling with low workload; a training sample acquisition module is used to acquire training data sets;
[0053] A flexible deep network training module, used for inputting the data set into the flexible deep network until a trained flexible deep network is obtained;
[0054] The target image inversion module is used to input the echo signal received by the detection array into the optimization model, and use the trained flexible deep network to obtain high-resolution reconstruction of the target at different distance domain slices through alternating optimization iterations, and realize fast high-resolution reconstruction of three-dimensional targets under low signal-to-noise ratio through joint reconstruction.
[0055] Preferably, the terahertz aperture coded imaging system comprises a terahertz transmitting system and an array detection receiving system, wherein the terahertz transmitting system comprises a transmitting antenna and a metamaterial aperture coding antenna, wherein the transmitting antenna transmits terahertz waves to the metamaterial aperture coding antenna, wherein the terahertz waves reflected by the metamaterial aperture coding antenna are randomly reflected in time and space to a detection target, wherein the detection target reflects the terahertz waves to the array detection receiving system, and the array detection receiving system acquires an echo signal.
[0056] Preferably, the main beam of the terahertz wave emitted by the transmitting antenna completely covers the metamaterial aperture coding antenna.
[0057] The present invention has the following characteristics and beneficial effects:
[0058] (1) Compared with the existing phased terahertz aperture coded 3D imaging based on coherent reception, the above phase-free imaging technology based on the detection array can simplify the complexity of the receiving imaging system and reduce the system cost. In addition, the detection array can use spatial sampling instead of temporal sampling, which can reduce the number of system sampling and encoding, thereby reducing the imaging system's demand for the refresh rate of the metamaterial aperture coded antenna and improving the imaging rate.
[0059] (2) Compared with the existing optimization model solving method based on sparse prior, the alternating direction multiplier method based on the flexible deep network can reduce the demand for the number of echo measurements for target reconstruction, thereby reducing the scale of the matrix imaging equation, reducing the computational complexity of the entire three-dimensional imaging, and improving imaging efficiency.
[0060] (3) Compared with the optimization model solving method based on generative prior, the above technical solution has strong flexibility and robustness, and is no longer restricted by the scope of the generative model. In addition, the above technical solution takes into account the impact of the actual terahertz aperture coded imaging system error on the imaging performance, and can achieve high-resolution reconstruction of random targets in actual terahertz aperture coded imaging scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0062] FIG1 is a schematic flow chart of a flexible and robust phase-free terahertz aperture-coded three-dimensional imaging method in this embodiment.
[0063] FIG. 2 is an analysis diagram of the number of array elements received by the array and the number of system encoding and sampling times in this embodiment.
[0064] FIG3 is a schematic diagram of a phase-free terahertz aperture-coded three-dimensional imaging system based on range domain slicing and a detection array in this embodiment.
[0065] FIG4 is a flow chart of an algorithm for solving the target scattering coefficient at a certain range domain slice in this embodiment.
[0066] FIG. 5 is a graph showing the performance analysis experimental results of the method of this embodiment and the existing method at a distance domain slice.
[0067] FIG6 is a diagram showing the experimental results of the flexibility comparison between the method of this embodiment and the existing depth prior method at a distance domain slice.
[0068] FIG. 7 is a schematic diagram of jointly reconstructing three-dimensional imaging based on information of distance domain slices in this embodiment.
[0069] FIG8 is a structural block diagram of a flexible and robust phase-free terahertz aperture-coded three-dimensional imaging device in this embodiment. DETAILED DESCRIPTION
[0070] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0071] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.
[0072] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.
[0073] Example 1
[0074] The present invention provides a flexible and robust phase-free terahertz aperture coded three-dimensional imaging method, such as Figure 1 As shown, the following steps are included:
[0075] Step 1, setting control parameters of the terahertz transmitting system and the detection array receiving system; transmitting terahertz waves so that the main beam covers all the metamaterial aperture coding antennas; designing the detection array;
[0076] It can be understood that the number of receiving elements in the detection array can be comprehensively considered by analyzing the impact of the number of receiving elements on the sampling and encoding times of the system, and adjusted according to actual needs and hardware conditions.
[0077] Step 2, analyzing the terahertz aperture coded imaging system based on array reception, acquiring echo signals by using the array detection, performing pulse compression processing on the echo signals to acquire a one-dimensional range image, extracting and separating range units with impulse responses according to the one-dimensional range image, and dividing the three-dimensional range image into two-dimensional range images of different distance domains;
[0078] Step 3: in the two-dimensional imaging area, obtain a reference signal matrix at the range domain slice by measuring and calculating, and construct a phase-free terahertz aperture coded mathematical imaging model based on the range domain slice and the detection array using the reference signal matrix and the range domain echo signal;
[0079] Step 4: applying the denoising mechanism to the mathematical imaging model of the phase-free terahertz aperture coding to obtain a phase-free terahertz aperture coding imaging optimization model based on the depth denoising prior and the detection array;
[0080] Step 5, obtaining a training data set of a flexible deep network, wherein the training data set includes a noise-free target data set and a noisy target data set obtained by a mathematical imaging model of phase-free terahertz aperture coding;
[0081] Step 6: input the training data set into the flexible deep network for training until a trained flexible deep network is obtained;
[0082] Step 7: using the alternating direction multiplier method to iteratively solve the optimization problem, and adding a pre-trained flexible deep network as a submodule of the alternating direction descent multiplier algorithm during the solving process;
[0083] Step 8: Reconstruct the target scattering information in azimuth and elevation directions of all range domain slices using the method, and achieve rapid high-resolution reconstruction of the three-dimensional target through joint reconstruction.
[0084] In the above steps, step 1 is the control parameter setting, steps 2-4 are the imaging system modeling and optimization model construction process, steps 5 and 6 are the flexible deep network training process, step 7 is the process of applying the trained flexible deep network to the alternating multiplier method to solve the target scattering coefficient at the two-dimensional range domain slice, and step 8 is the process of three-dimensional target joint reconstruction.
[0085] It is understandable that in step 1, a phase-free terahertz aperture-coded 3D imaging system based on distance domain slicing and detection array is constructed. This is mainly because when coherently detecting high-frequency and large-bandwidth terahertz signals, the receiver still faces great challenges in down-conversion and bandwidth reduction. Therefore, exploring phase-free 3D imaging can reduce system costs and complexity. In addition, array reception is adopted under the premise of phase-free imaging, and spatial sampling is used instead of temporal sampling, which greatly improves the efficiency of 3D imaging while reducing the refresh rate requirements of metamaterial coding antennas.
[0086] The detection array is designed, as shown in the following figure. Figure 2 As shown in the figure, the effect of different numbers of receiving array elements on the encoding and sampling times of the imaging system is analyzed. From the results, we can see that as the number of echo measurements M decreases and the number of detection array elements H increases, This can not only reduce the difficulty of hardware implementation of the terahertz aperture coding system, but also greatly improve the rate of three-dimensional imaging.
[0087] In step 2, the terahertz aperture coded imaging system is analyzed, an echo signal is acquired using a detection array, the echo signal is pulse compressed to acquire a one-dimensional range image, and a range domain echo signal containing target scattering information is acquired based on the range image distribution. The three-dimensional imaging area is divided into two-dimensional imaging areas of different distance domains by gridding only the distance units containing target scattering information.
[0088] Specifically, Figure 3As shown, a schematic diagram of an array receiving terahertz aperture coding three-dimensional imaging system based on distance domain slicing is shown, which mainly includes a control processing terminal, a transmitting antenna, a coding module, a metamaterial aperture coding antenna, and a detection array. In the imaging system, the transmitting antenna transmits a terahertz signal to irradiate the surface of the metamaterial aperture coding antenna. At the same time, the control processing terminal controls the coding module to load a random modulation factor to the metamaterial aperture coding antenna to perform phase modulation or amplitude modulation on the incident terahertz signal, and then a spatially uncorrelated random radiation field is formed in the target area. The different modulation factors loaded at different times make the radiation field time-independent, and finally a detection signal with significant time-space non-correlation characteristics is formed in the imaging area. The detection array is used to detect the initial echo signal reflected by the target, and then the initial echo signal is pulse-compressed to obtain a one-dimensional range image of the target. According to the one-dimensional range image result, the echo signal with impulse response is extracted and separated, and the range unit with target scattering is determined as a range domain slice, thereby dividing the three-dimensional imaging area into multiple two-dimensional imaging areas of different range domains.
[0089] In step 3, a phase-free terahertz aperture coded mathematical imaging model based on distance domain slicing and detection array is constructed.
[0090] Specifically, in this embodiment, the mathematical imaging model expression of the phase-free terahertz aperture coding after pulse pressure processing at a certain distance domain slice is as follows:
[0091] (1)
[0092] in, is the target scattering coefficient vector at range r, is the interference noise vector received by the echo signal during the M detection process, and is the matrix of the echo signal and reference signal received by the detection array M times. For the mth sampling reception, the echo intensity received by the entire detection array is It can be expressed as:
[0093] (2)
[0094] in, is the echo intensity vector detected by H detection array elements, is the corresponding reference signal matrix, , It is the reference signal corresponding to the nth grid unit and the hth detection array element.
[0095] Further, the reference signal matrix at the distance domain r is modeled by firstly using a small number of known targets Build the target dataset Then, different targets are used to obtain different echo vectors through the imaging system, and the echo data set matrix can be obtained by combining them. , then the mathematical expression for modeling the reference signal matrix can be written as
[0096] (3)
[0097] in, represents transpose, is the echo data set matrix The mth column of Represents the matrix to be modeled in the distance domain r The mth column of Representation Matrix The mth column of , by solving formula (3) through the inverse problem solving algorithm, we can get the mth row of the actual reference signal matrix. After M cycles, we can get the reference signal matrix at the distance domain r .
[0098] In step 4, a phase-free terahertz aperture coded imaging optimization model based on deep denoising prior and detection array is constructed, and the expression is as follows:
[0099] (4)
[0100] in, is the target scattering coefficient to be solved in the distance domain r, N is the number of target grids, is the echo signal vector obtained by the entire detection array after M samplings, is the reference signal matrix in the distance domain r obtained by measurement and modeling, is a priori term learned through flexible deep network optimization, It is a preset very small number.
[0101] In step 5, when acquiring the flexible deep network training data set, the number and position of the target scattering points are randomly changed. The training data set includes a noise-free target data set and a noisy target dataset solved by the terahertz aperture coded imaging model ;
[0102] In step 6, the flexible deep network adopts a denoising convolutional neural network, which mainly includes three parts, the first part consists of a convolutional layer and a non-linear activation layer, the second part consists of a convolutional layer, a batch normalization and a non-linear activation layer, and the third part is a convolutional layer;
[0103] In this embodiment, the training data set is input into the flexible deep network for training until a trained flexible deep network is obtained, which includes:
[0104] Inputting the noisy target into the flexible deep network to obtain a residual image;
[0105] Subtracting the noisy target from the residual image output by the flexible deep network to obtain a denoised noise target;
[0106] A reconstruction loss is obtained by performing a mean square error calculation based on the denoising target and the noise-free target;
[0107] According to the reconstruction loss, the weight parameters of the denoising convolutional neural network are continuously learned from the data set, and finally a trained deep denoising network is obtained.
[0108] In step 7, the iterative solution of the optimization model using the alternating direction multiplier method includes:
[0109] Including slack variables in the optimization model and dual variables The augmented Lagrangian function is obtained
[0110] (5)
[0111] In the above formula, is the scattering coefficient vector of the two-dimensional target to be solved in the range domain r, is a priori term learned through flexible deep network optimization, is the echo intensity information vector obtained by the entire detection array after M samplings, is the reference signal matrix at the distance domain r obtained by measurement and computational modeling. The iterative solution of formula (5) by the alternating direction multiplier method mainly includes the solution of three variables. First, fix and , using the gradient descent algorithm to update the slack variables , the optimization function can be expressed as:
[0112] (6)
[0113] in, is the estimated value at step t+1 of the optimization iteration, , , is the estimated value at step t, is the echo intensity signal vector obtained by the entire detection array after M samplings, is the reference signal matrix at the distance domain r obtained by the measurement fusion modeling method, and the updated relaxation variable can be obtained by the gradient descent method. ;
[0114] Then fix and , using the pre-trained flexible deep network to approximate the solution of the optimization problem, the optimization function can be expressed as:
[0115] (7)
[0116] in, and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, It is a priori term learned through flexible deep network optimization. Based on the viewpoint of maximum likelihood estimation of posterior distribution, a denoising convolutional neural network with good denoising effect is used to approximate the solution of the above formula without complex optimization iterations;
[0117] fixed and , update the dual variable :
[0118] (8)
[0119] in, , and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, parameter .
[0120] Finally, after T steps of alternating optimization, the target scattering coefficient at the distance domain r is obtained.
[0121] (9)
[0122] Figure 5 5 is a curve diagram of simulation experimental results of the method of this embodiment and the existing method in a distance domain. (a) and (b) in FIG5 are curves obtained with PSNR and MSE as performance evaluation indicators, respectively. The solid line with a square is the method of this embodiment, and the dashed line with a five-pointed star, the dashed line with a triangle, and the dashed line with a circle are the classic soft threshold-based alternating direction multiplier method (ST-ADMM), the WFOS algorithm, and the SWFOS algorithm, respectively. Figure 6 Comparison chart of reconstruction experimental results of different targets between this implementation method and existing methods. Figure 6 (a) shows the reconstruction results of targets within the generator range by different methods, where the first row is targets of different sparsity, the second row is the existing imaging method based on depth prior, and the third row is the method of this embodiment. Figure 6(b) shows the reconstruction results of different methods for targets outside the generator range, where the first row is for targets with different sparsity, the second row is for existing imaging methods based on depth priors, and the third row is for the method of this embodiment. Figure 5 and Figure 6 The method of this embodiment can reconstruct various targets with high quality at low signal-to-noise ratio, and has good flexibility and robustness. In addition, the method has low requirements for echo measurement data, which reduces the scale of the matrix imaging equation and the computational burden. It is combined with the detection array to further improve the three-dimensional imaging rate.
[0123] In step 8, the three-dimensional target reconstruction is achieved through joint reconstruction. Figure 7 First, the alternating direction multiplier method based on deep learning is used to reconstruct the target scattering information in the azimuth and elevation directions of all range domain slices, and then the three-dimensional target is quickly reconstructed with high resolution through joint reconstruction.
[0124] Example 2
[0125] This embodiment provides a device for implementing a flexible and robust phase-free terahertz aperture coded three-dimensional imaging method provided in Embodiment 1, such as Figure 8 As shown, including:
[0126] An information acquisition module based on a detection array is used to obtain echo signals, simplify the complexity of the imaging system, reduce the number of encoding and sampling of the imaging system, and reduce the refresh rate requirements of the metamaterial coding antenna in the imaging system;
[0127] The range domain slicing module is used to convert complex three-dimensional range images into two-dimensional range images in different distance domains;
[0128] Reference signal matrix modeling module, used to achieve high-precision reference signal matrix modeling with low workload;
[0129] The reference signal matrix modeling module is used to achieve low-workload and high-precision radiation field modeling, correct the imaging model, and improve the target reconstruction quality from the root.
[0130] A training sample acquisition module is used to acquire a training data set, wherein the training data set includes a noise-free target data set. , a noisy target data set solved by the phase-free terahertz aperture-coded mathematical imaging model ;
[0131] A flexible deep network training module, used for inputting the data set into a denoising convolutional neural network until a trained flexible deep network is obtained;
[0132] The target image inversion module is used to input the echo signal received by the detection array into the optimization model, and use the trained flexible deep network to obtain high-resolution reconstruction of the target at different distance domain slices through alternating optimization iterations, and realize fast high-resolution reconstruction of three-dimensional targets under low signal-to-noise ratio through joint reconstruction.
[0133] Further, such as Figure 3 As shown, the terahertz aperture coded imaging system includes a terahertz transmitting system and an array detection receiving system. The terahertz transmitting system includes a transmitting antenna and a metamaterial aperture coding antenna. The transmitting antenna transmits a terahertz wave to the metamaterial aperture coding antenna. The terahertz wave reflected by the metamaterial aperture coding antenna is randomly reflected in time and space to the detection target. The detection target reflects the terahertz wave to the array detection receiving system, and the array detection receiving system acquires an echo signal. The main beam of the terahertz wave emitted by the transmitting antenna covers the entire metamaterial aperture coding antenna.
[0134] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.
Claims
1. A flexible and robust phase-free terahertz aperture-coded three-dimensional imaging method, characterized in that: The steps include: Step 1: Construct a phase-free terahertz aperture-coded mathematical imaging model based on distance domain slicing and detection array. Step 1.1, build a terahertz aperture coded imaging system, use the detection array in the terahertz aperture coded imaging system to obtain echo signals, perform pulse compression processing on the echo signals to obtain a one-dimensional range image, extract and separate the range units with impulse responses according to the one-dimensional range image, and divide the three-dimensional imaging area into multiple two-dimensional imaging areas with different distance domains; Step 1.2, in the two-dimensional imaging area, obtain a reference signal matrix at the range domain slice by measurement and calculation, and construct a phase-free terahertz aperture coded mathematical imaging model based on the range domain slice and the detection array using the reference signal matrix and the range domain echo signal; The mathematical imaging model of phase-free terahertz aperture coding based on the detection array at a certain distance domain slice r is: in, is the target scattering coefficient vector to be solved in the range domain r, is the interference noise vector received by the echo signal during the M detection process, and S r =[S r (t1),S r (t2),…,S r (t M )] T is the matrix of the echo signal and reference signal received by the detection array M times. For the mth sampling reception, the echo intensity received by the entire detection array is It is expressed as: in, is the echo vector detected by H detection array elements, is the corresponding reference signal matrix, is the reference signal corresponding to the nth grid unit and the hth detection array element; The method of modeling the reference signal matrix is as follows: First, a small number of known targets β (i) ∈R N ,i=1,2,…,ΡBuild the target data set Then, different targets are used to obtain different echo vectors through the imaging system, and the echo data set matrix Sr is obtained by combining them. set =[Sr (1) ,Sr (2) ,…,Sr (P) ]∈R M×P , then the mathematical expression for modeling the reference signal matrix is written as in,(·) T represents transpose, is the echo data set matrix (Sr set ) T The mth column of Represents the matrix to be modeled in the distance domain r The mth column of Represents the matrix (ω set ) T The mth column of set =[ω (1) ,ω (2) ,…,ω (Ρ) ]∈R M×P , solve formula (3) through the inverse problem solving algorithm to get the mth row of the actual reference signal matrix, and obtain the reference signal matrix at the distance domain r after M cycles Step 2, combining the denoising mechanism with the phase-free terahertz aperture coded mathematical imaging model to obtain a phase-free terahertz aperture coded imaging optimization model based on a depth denoising prior and a detection array; Step 3, using an alternating direction multiplier method to iteratively solve the phase-free terahertz aperture coded imaging optimization model, and adding a pre-trained flexible deep network as a submodule of the alternating direction descent multiplier algorithm during the solving process; Step 4: Solve the target scattering information in azimuth and elevation for all range domain slices, and achieve fast high-resolution reconstruction of the three-dimensional target through joint reconstruction.
2. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 1, characterized in that: In step 2, a phase-free terahertz aperture coded imaging optimization model based on deep denoising prior and detection array is constructed, and the expression is as follows: Among them, β r is the target scattering coefficient to be solved in the distance domain r, N is the number of target grids, is the echo signal vector obtained by the entire detection array after M samplings, is the reference signal matrix in the distance domain r obtained by measurement and computational modeling, ψ(β r ) is a prior term learned through flexible deep network optimization, σ 2 It is a preset very small number.
3. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 1, characterized in that: The flexible deep network adopts a deep denoising network, which mainly includes three parts. The first part consists of a convolutional layer and a non-linear activation layer, the second part consists of a convolutional layer, batch normalization and a non-linear activation layer, and the third part is a convolutional layer.
4. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 2, characterized in that: The step 2 also includes obtaining a training data set, and using the training data set to train the flexible deep network, wherein the training data set includes a noise-free target data set and a noisy target data set solved by a phase-free terahertz aperture coded mathematical imaging model.
5. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 4, characterized in that: The training method of the flexible deep network is: Inputting the noisy target data into the flexible deep network to obtain a residual image; Subtracting the noisy target data from the residual image to obtain a denoised target; A reconstruction loss is obtained by calculating a mean square error according to the denoising target and the noise-free target; The weight parameters of the denoising convolutional neural network are continuously learned from the training data set according to the reconstruction loss, and finally a trained flexible deep network is obtained.
6. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 4, characterized in that: In step 3, when solving the terahertz aperture coded imaging optimization model, the optimization problem needs to be converted into a multivariable augmented Lagrangian function as shown in the following formula, and a relaxation variable ξ is added to the terahertz aperture coded imaging optimization model: r and the dual variable κ r , the expression is as follows: In the above formula, βr is the scattering coefficient vector of the two-dimensional target to be solved at the distance domain r, ψ(β r ) is a priori term learned through flexible deep network optimization, is the echo intensity information vector obtained by the entire detection array after M samplings, is the reference signal matrix at the distance domain r.
7. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 6, characterized in that: In step 3, the iterative solution of formula (5) by the alternating direction multiplier method mainly includes the solution of three variables. First, fix and Update the slack variables using the gradient descent algorithm The optimization function is expressed as: in, is the estimated value at step t+1 of the optimization iteration, is the estimated value at step t, is the echo intensity signal vector obtained by the entire detection array after M samplings, is the reference signal matrix at the distance domain r obtained by measurement and calculation through modeling, and the updated relaxation variable is obtained by gradient descent method. Then fix and The pre-trained flexible deep network is used to approximate the solution of the optimization problem. The optimization function is expressed as: in, and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, ψ(β r ) is a priori term learned through flexible deep network optimization. According to the viewpoint of maximum likelihood estimation of posterior distribution, a denoising convolutional neural network with good denoising effect is used to approximate the solution of the above formula without complex optimization iterations; fixed and Update the dual variable κ r in, and To optimize the estimated value of the t+1th step of the iteration, is the estimated value of the tth step, with parameter η∈(0,1).
8. The flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 7, characterized in that: Use the deep denoising network to approximate β r The update optimization function is as follows: in, and To optimize the estimated value of the t+1th step of the iteration, is the estimated value at step t, is the deep denoising network obtained by the training, and σ is the set denoising coefficient.
9. A device for implementing the flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to any one of claims 1 to 8, characterized in that: include: An information acquisition module based on a detection array is used to obtain echo signals, simplify the complexity of the imaging system, reduce the number of encoding and sampling of the imaging system, and reduce the refresh rate requirements of the metamaterial coding antenna in the imaging system; The range domain slicing module is used to convert complex three-dimensional range images into two-dimensional range images in different distance domains; Reference signal matrix modeling module, used to achieve high-precision reference signal matrix modeling with low workload; A training sample acquisition module is used to acquire a training data set; A flexible deep network training module, used for inputting the data set into the flexible deep network until a trained flexible deep network is obtained; The target image inversion module is used to input the echo signal received by the detection array into the optimization model, and use the trained flexible deep network to obtain high-resolution reconstruction of the target at different distance domain slices through alternating optimization iterations, and realize fast high-resolution reconstruction of three-dimensional targets under low signal-to-noise ratio through joint reconstruction.
10. The apparatus of the flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 9, characterized in that: The terahertz aperture coded imaging system includes a terahertz transmitting system and an array detection receiving system. The terahertz transmitting system includes a transmitting antenna and a metamaterial aperture coding antenna. The transmitting antenna transmits a terahertz wave to the metamaterial aperture coding antenna. The terahertz wave reflected by the metamaterial aperture coding antenna is randomly reflected in time and space to a detection target. The detection target reflects the terahertz wave to the array detection receiving system, and the array detection receiving system acquires an echo signal.
11. The apparatus of the flexible and robust phase-free terahertz aperture coded three-dimensional imaging method according to claim 10, characterized in that: The main beam of the terahertz wave emitted by the transmitting antenna completely covers the metamaterial aperture coding antenna.