Through-the-wall radar target high-resolution imaging method based on conditional diffusion model

Abstract

The invention provides a through-the-wall radar target high-resolution imaging method based on a conditional diffusion model, and the method breaks through the limitation of the resolution of a conventional algorithm of a through-the-wall radar system through the building of a Markov process of forward diffusion and reverse sampling, the design of a noise reduction network, and the control of the iteration generation of a high-resolution optical image through the radar image information. The shape and contour information of the target is effectively recovered, the identifiability of the result is enhanced, and subsequent operation and use of the imaging result are facilitated; that is to say, compared with other imaging methods, high-resolution imaging can be carried out on the target in the shielded space and scene, the appearance and contour information of the target is recovered, the resolution of a traditional through-the-wall radar imaging method is improved and broken through, and a visual and recognizable imaging result is provided.

Description

Technical Field

[0001] The invention belongs to the technical field of radar signal processing and through-wall radar imaging, and in particular relates to a through-wall radar target high-resolution imaging method based on a conditional diffusion model. Background Art

[0002] With the acceleration of the world's urbanization process, the layout and structure of urban buildings are becoming increasingly complex, the building density is increasing, and some urban internal spaces are seriously obscured, making it easy for targets to hide, posing a serious threat and posing a major security risk. At present, there are many practical and effective target detection methods, including infrared signals, X-rays, ultrasonic detection, etc., but each has different defects in actual scenarios. For example, although X-rays have extremely strong penetration, they pose a huge radiation hazard to the human body; ultrasonic detection has high accuracy and sensitivity, but is easily affected by ambient temperature noise; infrared signals have strong anti-interference ability but almost no penetration ability, etc. Therefore, it is of great significance to study solutions for efficient detection of targets through walls.

[0003] Through-the-wall radar imaging technology is a non-contact and non-destructive electromagnetic sensing technology with many advantages, such as strong penetration ability, strong anti-interference ability, and a wider detection range. The target echo signal behind the obstacle is received by using the penetration of the transmitted signal, and it is analyzed, displayed or imaged to obtain the target information of the shielded area, thereby achieving target positioning and tracking. It has a wide range of applications in disaster relief, military detection, medical detection, etc.

[0004] At present, many high-resolution algorithms have been proposed in the field of through-wall radar imaging. However, for complex extended targets in actual scenarios, performance limitations such as radar aperture and bandwidth make the improvement in target imaging resolution very weak. The imaging results are mostly unrecognizable spots that lose the target's outline, which greatly reduces the usability of the imaging results. Through-wall radar high-resolution imaging technology is still a major challenge in this field. Summary of the invention

[0005] In order to solve the problem that through-the-wall radar imaging has low resolution and target shape and contour information is missing, resulting in difficulty in identification, the present invention provides a through-the-wall radar target high-resolution imaging method based on a conditional diffusion model. By establishing a Markov process of forward diffusion and reverse sampling, a noise reduction network is designed, and radar image information is used to control the iterative generation of high-resolution optical images, the resolution limitation of traditional algorithms of through-the-wall radar systems is broken through, the shape and contour information of the target is effectively restored, the recognizability of the results is enhanced, and the subsequent operation and use of the imaging results are convenient.

[0006] A through-wall radar target high-resolution imaging method based on a conditional diffusion model comprises the following steps:

[0007] S1: Obtain a 3D radar image of the target behind the wall I radar and target depth image I optical ;

[0008] S2: target depth image I optical Add noise moment by moment until the target depth image I optical Transformed into a target noise image x that conforms to a random Gaussian distribution T , where T is the target depth image I optical The total number of noise-adding moments required to transform into the target noise image;

[0009] S3: 3D radar image I radar , time T, target noise image x T Input the trained noise prediction network to obtain the predicted noise;

[0010] S4: Use the predicted noise to predict the target noise image x T Perform reverse denoising to obtain the target noise image required for reverse denoising at the next moment;

[0011] S5: Decrement the current time by 1 to get the next time, and then convert the three-dimensional radar image I radar , the latest target noise image at the next moment is input into the trained noise prediction network to obtain the predicted noise required for the next reverse denoising;

[0012] S6: re-execute steps S4 to S5 using the latest predicted noise and the latest target noise image until T decreases to 0, and the target noise image obtained when the iteration time is 0 is the final imaging of the target to be measured.

[0013] Furthermore, in step S2, the target noise image x that conforms to the random Gaussian distribution T The method to obtain is:

[0014]

[0015] Among them, x0 is the target depth image I optical ,∈ T is the random noise sampled from the standard Gaussian distribution at the Tth noise adding moment, is the variance variable at the Tth noise adding moment, and α k Represents the random noise ∈ sampled at the kth noise adding moment k The noise variance β k The difference between α and 1, and α k =1-β k , k=1,…,T;

[0016] At the same time, the target intermediate noise image x at any noise adding time t t The method to obtain is:

[0017]

[0018] Among them, ∈ t is the random noise sampled from the standard Gaussian distribution at the t-th noise adding moment, t=1,…,T-1, is the variance variable at the tth noise adding moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t.

[0019] Further, in step S4, a method for acquiring any target noise image required for reverse noise reduction at the next moment is:

[0020] Get the mean value μ of the target noise image required for reverse denoising at the next moment θ (x t ,t,c):

[0021]

[0022] Among them, t represents the current reverse noise reduction time, f θ (x t ,t,c) represents the latest predicted noise at the current reverse denoising time t, and c represents the three-dimensional radar image I radar , θ represents the network parameter set of the prediction network; x t represents the latest target noise image obtained at the current reverse denoising time t; is the variance variable at the t-th noise addition moment corresponding to the t-th reverse noise reduction moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-βi ， i=1,…,t;

[0023] Fit the mean to μ θ (x t ,t,c), the variance conforms to (1-α t )I image is used as the target noise image x required for reverse denoising at the next time t-1 t-1 .

[0024] Furthermore, in step S3, the training method of the noise prediction network is:

[0025] S31: Collect 3D radar images and target depth images of different targets behind the wall;

[0026] S32: Acquire a target noise image of each target according to the target depth image of each target;

[0027] S33: inputting the total number of noise addition moments T required for converting the three-dimensional radar image of each target and the target depth image of each target into the target noise image and the target noise image of each target into the trained noise prediction network to obtain predicted noise;

[0028] S34: Construct the loss function L according to the predicted noise as follows:

[0029]

[0030] Among them, the reverse noise reduction time t=1,…,T, represents the predicted noise obtained at the reverse denoising time t, and c represents the three-dimensional radar image I radar , θ represents the network parameter set of the prediction network; ∈ t represents the random noise sampled from the standard Gaussian distribution at the tth noise adding moment in the noise adding process of converting the target depth image into the target noise image; x0 represents the target depth image without noise; x t represents the target noise image obtained at any reverse denoising time t, and is the variance variable at the tth noise adding moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t; Indicates the p-norm; Represents the acquisition of random noise ∈ t As a variable, change the random noise ∈ t The different values ​​of Expectations; E x,c Expressing expectations expectations;

[0031] S35: Determine whether the loss function L is less than a set threshold. If yes, obtain the final noise prediction network; if no, proceed to step S36;

[0032] S36: changing the network parameter values ​​in the network parameter set θ of the noise prediction network, and then re-executing steps S33 to S35 using the noise prediction network after the network parameters are changed.

[0033] Furthermore, in step S1, the three-dimensional radar image I of the target to be detected radar The method to obtain is:

[0034] Assume that there is a wall with a thickness of d and a dielectric constant of ∈ in the x-axis direction, the antenna and the target to be measured are respectively in the positive and negative directions of the y-axis, and everything except the wall is free space;

[0035] The radar works in synthetic aperture mode, using N transceiver antenna units with a spacing of d1 to transmit a stepped frequency continuous wave signal containing M frequency points. Assuming that there are Q targets to be measured on the other side of the wall, the frequency domain echo signal S(m,n) received by the nth transceiver antenna unit at the mth frequency point is:

[0036]

[0037] Among them, σ wall and σ q represents the scattering coefficient of the wall and the qth target to be measured, q=1,2,…,Q,S noise (m,n) represents the noise signal of the nth transmitting and receiving antenna unit at the mth frequency point, f m is the frequency of the mth frequency point, τ wall represents the round-trip delay between N transmitting and receiving antenna units and the wall, τ q,n represents the round-trip delay between the nth antenna and the qth target to be measured;

[0038] Perform inverse Fourier transform on the frequency domain echo signals S(m,n) received by the N transceiver antenna units respectively to obtain the time domain echo corresponding to each transceiver antenna unit;

[0039] Perform BP imaging on the time domain echo corresponding to each transceiver antenna unit to obtain the BP imaging I of each transceiver antenna unit in the target area. n ;

[0040] The BP image I of each transceiver antenna unit in the target area n Perform coherent superposition to obtain a superimposed image

[0041] The phase coherence factor PCF is used to weight the superimposed image to obtain the three-dimensional radar image I of the target to be measured. radar =PCF·I BP .

[0042] Furthermore, the method for obtaining the phase coherence factor PCF is:

[0043] PCF=1-std(e -jφ )

[0044]

[0045] in, represents the calculated standard deviation; j represents the imaginary part; φ l Indicates the phase of the echo signal reflected by the target to be measured to the lth transceiver antenna unit. The serial number of the transceiver antenna unit is l = 0, 1, ..., N-1, φ n It represents the phase of the echo signal reflected by the target to be measured to the nth transceiver antenna unit, and φ represents the phase of the total echo signal reflected by the target to be measured to all transceiver antenna units.

[0046] Furthermore, in step S2, the target depth image I of the target to be measured optical The method to obtain is:

[0047] According to the perspective principle of light, the optical depth camera is placed at the center of the radar's transceiver antenna array to obtain the same viewing angle as the radar and the same as the 3D radar image I radar Matching target depth images I optical .

[0048] Beneficial effects:

[0049] The present invention provides a through-the-wall radar target high-resolution imaging method based on a conditional diffusion model. By establishing a Markov process of forward diffusion and reverse sampling, designing a noise reduction network, and using radar image information to control the iterative generation of high-resolution optical images, the resolution limitation of traditional algorithms of through-the-wall radar systems is broken through, the shape and contour information of the target is effectively restored, the recognizability of the results is enhanced, and the subsequent operation and use of the imaging results are convenient; that is, compared with other imaging methods, the present invention can perform high-resolution imaging of the target in a shielded space and scene, restore the shape and contour information of the target, improve and break through the resolution of traditional through-the-wall radar imaging methods, and provide intuitive and recognizable imaging results. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of an embodiment of the present invention;

[0051] Figure 2 It is a schematic diagram of a signal scenario in the present invention;

[0052] Figure 3 is a schematic diagram of phase coherence factor weighting in the present invention;

[0053] Figure 4 It is the forward diffusion and reverse sampling process of the conditional noise reduction diffusion probability logic adopted by the present invention;

[0054] Figure 5 It is the noise reduction network hierarchical structure designed by the present invention;

[0055] Figure 6 It is the iterative high-resolution imaging process of the present invention;

[0056] Figure 7 It is a schematic diagram of a simulation scenario of the present invention;

[0057] Figure 8 Schematic diagram of the experimental scenario of the present invention, wherein (a) is the MIMO radar used, (b) is a schematic diagram of the radar array element position, (c) is a schematic diagram of a non-through-the-wall scenario, (d) is a schematic diagram of a through-the-wall scenario, (e) is a schematic diagram of the radar wall-attached mode in the through-the-wall scenario, and (f) is a schematic diagram of the target placement behind the wall;

[0058] Fig. 9 : is a simulation result diagram of the present invention, wherein (a) is a simulated target diagram, (b) is a three-dimensional radar imaging diagram, (c) is a radar cross-sectional diagram of the front view and the top view, (d) is a high-resolution imaging diagram of the method proposed in the present invention, and (e) is a target true value label;

[0059] Fig.10 It is the measured result diagram of the present invention, wherein (a) is the measured target diagram, (b) is the three-dimensional radar imaging diagram, (c) is the radar cross-sectional diagram of the front view and the top view, (d) is the high-resolution imaging diagram of the method proposed in the present invention, and (e) is the target true value label. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0061] In order to overcome the problems of insufficient resolution and weak imaging identifiability of existing through-wall radar imaging algorithms, the present invention provides a through-wall radar target high-resolution imaging method based on a conditional diffusion model (Conditional Denoising Diffusion Probabilistic Model, CDDPM). The basic idea is to fit the distribution characteristics of the target data domain under specified conditions through a forward denoising and reverse denoising process based on a Markov chain.

[0062] Specifically, Figure 1 As shown, the through-wall radar target high-resolution imaging method of the present invention comprises the following steps:

[0063] S1: Obtain a 3D radar image of the target behind the wall I radar and target depth image I optical ;

[0064] Among them, the three-dimensional radar image I of the target to be measured radar The method to obtain is:

[0065] S11: Figure 2 As shown, it is assumed that there is a wall with a thickness of d and a dielectric constant of ∈ in the x-axis direction, the antenna and the target to be measured are respectively in the positive and negative directions of the y-axis, and everything except the wall is free space;

[0066] S12: The radar works in synthetic aperture mode, using N transceiver antenna units with a spacing of d1 to transmit a stepped frequency continuous wave signal (SFCW) containing M frequency points. Assuming that there are Q targets to be measured on the other side of the wall, the frequency domain echo signal S(m,n) received by the nth transceiver antenna unit at the mth frequency point is:

[0067]

[0068] Among them, σ wall and σ q represents the scattering coefficient of the wall and the qth target to be measured, q=1,2,…,Q,S noise (m,n) represents the noise signal of the nth transmitting and receiving antenna unit at the mth frequency point, f m is the frequency of the mth frequency point, τ wall represents the round-trip delay between N transmitting and receiving antenna units and the wall, τ q,n represents the round-trip delay between the nth antenna and the qth target to be measured;

[0069] S13: performing inverse Fourier transform on the frequency domain echo signals S(m,n) received by the N transceiver antenna units respectively to obtain the time domain echo corresponding to each transceiver antenna unit;

[0070] S14: Perform BP imaging on the time domain echo corresponding to each transceiver antenna unit respectively to obtain the BP imaging I of each transceiver antenna unit in the target area. n ;

[0071] S15: BP imaging of each transceiver antenna unit in the target area I n Perform coherent superposition to obtain a superimposed image

[0072] S16: To improve the image quality, reduce the grating lobe energy, such as Figure 3 As shown in the figure, the phase coherence factor PCF is used to weight the superimposed image to obtain the three-dimensional radar image I of the target to be measured. radar =PCF·I BP .

[0073] Among them, the calculation method of the phase coherence factor PCF is:

[0074] PCF=1-std(e -jφ )

[0075]

[0076] in, represents the calculated standard deviation; j represents the imaginary part; φ l Indicates the phase of the echo signal reflected by the target to be measured to the lth transceiver antenna unit. The serial number of the transceiver antenna unit is l = 0, 1, ..., N-1, φ n It represents the phase of the echo signal reflected by the target to be measured to the nth transceiver antenna unit, and φ represents the phase of the total echo signal reflected by the target to be measured to all transceiver antenna units.

[0077] Furthermore, the target depth image I of the target to be measured optical The method to obtain is:

[0078] According to the perspective principle of light, the optical depth camera is placed at the center of the radar's transceiver antenna array to obtain the same viewing angle as the radar and the same as the 3D radar image I radar Matching target depth images I optical .

[0079] S2: target depth image I optical Add noise moment by moment until the target depth image I optical Transformed into a target noise image x that conforms to a random Gaussian distribution T , where T is the target depth image I optical The total number of noise-adding moments required to transform into the target noise image;

[0080] It should be noted that the target noise image x that conforms to the random Gaussian distribution T The method to obtain is:

[0081]

[0082] Among them, x0 is the target depth image I optical ,∈ T is the random noise sampled from the standard Gaussian distribution at the Tth noise adding moment, is the variance variable at the Tth noise adding moment, and α k Represents the random noise ∈ sampled at the kth noise adding moment k The noise variance β k The difference between α and 1, and α k =1-β k , k=1,…,T;

[0083] Based on this, the present invention can also give a target intermediate noise image x at any noise adding time t: t The method to obtain is:

[0084]

[0085] Among them, ∈ t is the random noise sampled from the standard Gaussian distribution at the t-th noise adding moment, where t=1,…,T-1, is the variance variable at the tth noise adding moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t.

[0086] It should be noted that the essence of step S2 is to establish a forward noise addition process, which belongs to the diffusion process in the conditional denoising diffusion probability model, that is, the optical data is denoised step by step using the definition of the Markov chain, wherein the Markov diffusion operator q(·) is expressed as follows:

[0087]

[0088] x0 represents the original high-resolution target depth image I optical , and a noise variance table β1,…,β1 with a value between 0 and 1 and monotonically increasing is preset t , for the convenience of subsequent representation, define α t =1-β t , So the diffusion operator is obtained as:

[0089]

[0090] The present invention uses t t=1,…,T represents the random noise sampled from the standard Gaussian distribution at each step, that is, noise sampling is performed once at each noise adding moment, and the random noise obtained by each sampling is independent of each other. Then the noisy image at the tth noise adding moment can be expressed as

[0091]

[0092] Through the above recursive formula, the relationship between the image at any time and the initial time can be established as follows:

[0093]

[0094] The present invention uses this forward process to obtain a high-resolution target depth image I optical It is gradually submerged by noise and finally transformed into a random Gaussian distribution.

[0095] S3: 3D radar image I radar , time T, target noise image x T Input the trained noise prediction network to obtain the predicted noise;

[0096] S4: Use the predicted noise to predict the target noise image x T Perform reverse denoising to obtain the target noise image required for reverse denoising at the next moment;

[0097] Furthermore, a method for acquiring a target noise image required for reverse noise reduction at any next moment is:

[0098] Get the mean value μ of the target noise image required for reverse denoising at the next moment θ (x t ,t,c):

[0099]

[0100] Among them, t represents the current reverse noise reduction time, f θ (x t ,t,c) represents the latest predicted noise at the current reverse denoising time t, and c represents the three-dimensional radar image I radar , θ represents the network parameter set of the prediction network; x t represents the latest target noise image obtained at the current reverse denoising time t; is the variance variable at the t-th noise addition moment corresponding to the t-th reverse noise reduction moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t;

[0101] Fit the mean to μ θ (x t ,t,c), the variance conforms to (1-α t )I image is used as the target noise image x required for reverse denoising at the next time t-1 t-1 .

[0102] S5: Decrement the current time by 1 to get the next time, and then convert the three-dimensional radar image I radar, the latest target noise image at the next moment is input into the trained noise prediction network to obtain the predicted noise required for the next reverse denoising;

[0103] S6: re-execute steps S4 to S5 using the latest predicted noise and the latest target noise image until T decreases to 0, and the target noise image obtained when the iteration time is 0 is the final imaging of the target to be measured.

[0104] For example, the target noise image x at the next moment after acquiring time T T-1 When the three-dimensional radar image I radar , time T, target noise image x T Input the trained noise prediction network to get the predicted noise f θ (x T ,T,c); Then, using the prediction noise f θ (x T ,T,c) for the target noise image x T Perform reverse denoising to obtain the target noise image x required for reverse denoising at the next moment T-1 ;

[0105] Then, the time T decreases by 1, and the three-dimensional radar image I radar , time T-1, target noise image x T-1 Input the trained noise prediction network to obtain the predicted noise f required for the next reverse denoising θ (x T-1 ,T-1,c); using the newly obtained prediction noise f θ (x T-1 ,T-1,c) for the target noise image x T-1 Perform reverse denoising to obtain the target noise image x required for the next reverse denoising T-2 ;

[0106] This process is repeated in this way until T decreases to 0, and the image obtained by the last reverse denoising is the final image of the target to be measured.

[0107] It can be seen that steps S3 to S6 of the present invention are actually establishing a reverse denoising process, which belongs to the sampling process in the conditional denoising diffusion probability model; Figure 4 As shown, the reverse process is also modeled by Markov chain, the purpose is to restore the condition c (3D radar image I radar ) under high-resolution image; Different from the forward process, the reverse process needs to predict the noise of the forward noise, so the present invention defines a neural network f θ The noise of each step is predicted, θ represents the parameters of the network, so the operator of the reverse process is defined as

[0108]

[0109] Among them, p(x T ) is a sample from a standard Gaussian distribution. Similar to the forward process, the present invention can use the mean and variance to represent the reverse process operator as follows:

[0110]

[0111] Through the noise prediction network f θ The noise is predicted and the mean of the reconstructed image is calculated, and then the reverse process can be iterated, where the predicted noise f θ (x t ,t,c) The reconstructed mean calculated is as follows:

[0112]

[0113] The following is a detailed introduction to the training method of the noise prediction network adopted by the present invention, which specifically includes the following steps:

[0114] S31: Collect 3D radar images and target depth images of different targets behind the wall;

[0115] S32: Acquire a target noise image of each target according to the target depth image of each target;

[0116] S33: inputting the total number of noise addition moments T required for converting the three-dimensional radar image of each target and the target depth image of each target into the target noise image and the target noise image of each target into the trained noise prediction network to obtain predicted noise;

[0117] S34: Construct the loss function L according to the predicted noise as follows:

[0118]

[0119] Among them, the reverse noise reduction time t=1,…,T, represents the predicted noise obtained at the reverse denoising time t, and c represents the three-dimensional radar image I radar , θ represents the network parameter set of the prediction network; ∈ t represents the random noise sampled from the standard Gaussian distribution at the tth noise adding moment in the noise adding process of converting the target depth image into the target noise image; x0 represents the target depth image without noise; x t represents the target noise image obtained at any reverse denoising time t, and is the variance variable at the tth noise adding moment, and α i Represents the random noise ∈ sampled at the i-th noise adding momenti The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t; Indicates the p-norm; Represents the acquisition of random noise ∈ t As a variable, change the random noise ∈ t The different values ​​of Expectations; E x,c Indicates expectation of obtaining The expectation of , x represents the target depth image;

[0120] S35: Determine whether the loss function L is less than a set threshold. If yes, obtain the final noise prediction network; if no, proceed to step S36;

[0121] S36: changing the network parameter values ​​in the network parameter set θ of the noise prediction network, and then re-executing steps S33 to S35 using the noise prediction network after the network parameters are changed.

[0122] It should be noted that if Figure 5 As shown, the noise prediction network of the present invention is a full convolution noise prediction network based on the self-attention mechanism. Specifically, firstly, a multi-layer two-dimensional convolution module is used to extract features of radar data and noisy images and then perform feature splicing. The data is reduced in dimension and restored through a multi-layer residual module, a down-sampling module, and an up-sampling module. At the same time, a self-attention mechanism module is added to the low-dimensional part to increase the processing weight of the effective part of the data. The network encodes the random time samples and adds them to the residual module for fusion processing, and finally obtains the predicted noise output of the same size as the optical image.

[0123] like Figure 6 As shown, when the present invention implements high-resolution imaging through iterative noise reduction, a Gaussian white noise image is first sampled from a normal distribution with zero mean and unit variance, and the predicted noise output at the current moment is obtained by inputting the image at the moment, the radar image as a condition, and the moment code into the noise prediction network; the image at the previous moment is calculated through parameterized reconstruction, and the previous steps are repeated, thereby iteratively implementing high-resolution imaging.

[0124] Below Figure 7 Taking the application scenario of as an example, a through-wall radar target high-resolution imaging method based on the conditional diffusion model provided by the present invention is described in detail.

[0125] First, set the simulation parameters as shown in Table 1. The radar works in synthetic aperture mode. The frequency of the transmitted step-frequency signal ranges from 1 GHz to 3.5 GHz, with a frequency point set every 10 MHz. The antenna moving size is 1m*1m, and the antenna step is 10 cm.

[0126] Table 1 Simulation parameter settings

[0127]

[0128]

[0129] Construct a dataset containing tables, chairs, people, and RPG in a simulation scenario.

[0130] The present invention lists Figure 8 An experimental scenario is shown in the figure. The experimental parameters are set as shown in Table 2. A 10-transmit 10-receive MIMO array is used to transmit a 1.7 GHz to 2.2 GHz stepped frequency signal. The array size is approximately 40 cm*40 cm. The concrete brick wall is approximately 20 cm thick. In the wall-attached mode, the target 2 m behind the wall is illuminated.

[0131] Table 2 Experimental parameter settings

[0132]

[0133] In the experimental scenario, a data set containing tables, chairs, and people is constructed and merged with the simulation data set. By establishing diffusion and sampling logic, the data set is divided into a training set and a test set. The designed network is trained using the training set to predict the noise at each step in the sampling process. The image at the previous moment is reconstructed through iterative parameterization, and finally the reconstruction of a high-resolution image is achieved.

[0134] The present invention selects some simulation and experimental results for analysis. The simulation results are as follows: Fig. 9 As shown in the figure, the target is a wooden table. (b) shows the 3D radar data obtained by BP imaging, and (c) is the cross-sectional view of the radar data from the front and top. Due to the low resolution of the radar data, the target's outline and type are basically indistinguishable. (d) and (e) respectively show the high-resolution imaging results of the proposed method and the corresponding true value image, which restores the target's true outline and texture information to a great extent and achieves high-resolution imaging of the target. The measured structure is shown in Fig.10 As shown in Figure 1, the target is an iron chair placed behind a 20 cm thick wall. (b) and (c) show the 3D radar image and cross-sectional view obtained by BP imaging. Since the actual measurement is affected by many complex factors, such as clutter caused by multipath reflection in the environment, the imaging accuracy of the target drops sharply. Through the processing of the proposed method, as shown in (d), the target contour information is restored, achieving high-resolution imaging.

[0135] In summary, compared with other imaging methods, the present invention can improve the accuracy of through-wall radar imaging results, restore the target contour and texture information, and improve the target's recognizability.

[0136] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for high-resolution imaging of through-wall radar targets based on a conditional diffusion model, characterized in that: The following steps are involved: S1: Obtain a 3D radar image of the target behind the wall I radar and target depth image I optical ; S2: target depth image I optical Add noise moment by moment until the target depth image I optical Transformed into a target noise image x that conforms to a random Gaussian distribution T , where T is the target depth image I optical The total number of noise-adding moments required to transform into the target noise image; S3: 3D radar image I radar , time T, target noise image x T Input the trained noise prediction network to obtain the predicted noise; S4: Use the predicted noise to predict the target noise image x T Perform reverse denoising to obtain the target noise image required for reverse denoising at the next moment; S5: Decrement the current time by 1 to get the next time, and then convert the three-dimensional radar image I radar , the latest target noise image at the next moment is input into the trained noise prediction network to obtain the predicted noise required for the next reverse denoising; S6: re-execute steps S4 to S5 using the latest predicted noise and the latest target noise image until T decreases to 0, and the target noise image obtained when the iteration time is 0 is the final imaging of the target to be measured.

2. A through-wall radar target high-resolution imaging method based on a conditional diffusion model as claimed in claim 1, characterized in that: In step S2, the target noise image x that conforms to the random Gaussian distribution T The method to obtain is: Among them, x0 is the target depth image I optical ,∈ T is the random noise sampled from the standard Gaussian distribution at the Tth noise adding moment, is the variance variable at the Tth noise adding moment, and α k Represents the random noise ∈ sampled at the kth noise adding moment k The noise variance β k The difference between α and 1, and α k =1-β k , k=1,…,T; At the same time, the target intermediate noise image x at any noise adding time t t The method to obtain is: Among them, ∈ t is the random noise sampled from the standard Gaussian distribution at the t-th noise adding moment, t=1,…,T-1, is the variance variable at the tth noise adding moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t.

3. The method for high-resolution through-wall radar target imaging based on a conditional diffusion model as claimed in claim 1, characterized in that: In step S4, the method for acquiring any target noise image required for reverse noise reduction at the next moment is: Get the mean value μ of the target noise image required for reverse denoising at the next moment θ (x t ,t,c): Among them, t represents the current reverse noise reduction time, f θ (x t ,t,c) represents the latest predicted noise at the current reverse denoising time t, and c represents the three-dimensional radar image I radar , θ represents the network parameter set of the prediction network; x t represents the latest target noise image obtained at the current reverse denoising time t; is the variance variable at the t-th noise addition moment corresponding to the t-th reverse noise reduction moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t; Fit the mean to μ θ (x t ,t,c), the variance conforms to (1-α t )I image is used as the target noise image x required for reverse denoising at the next time t-1 t-1 .

4. The method for high-resolution through-wall radar target imaging based on a conditional diffusion model as claimed in claim 1, characterized in that: In step S3, the training method of the noise prediction network is: S31: Collect 3D radar images and target depth images of different targets behind the wall; S32: Acquire a target noise image of each target according to the target depth image of each target; S33: inputting the total number of noise addition moments T required for converting the three-dimensional radar image of each target and the target depth image of each target into the target noise image and the target noise image of each target into the trained noise prediction network to obtain predicted noise; S34: Construct the loss function L according to the predicted noise as follows: Among them, the reverse noise reduction time t=1,…,T, represents the predicted noise obtained at the reverse denoising time t, and c represents the three-dimensional radar image I radar , θ represents the network parameter set of the prediction network; ∈ t represents the random noise sampled from the standard Gaussian distribution at the tth noise adding moment in the noise adding process of converting the target depth image into the target noise image; x0 represents the target depth image without noise; x t represents the target noise image obtained at any reverse denoising time t, and is the variance variable at the tth noise adding moment, and α i Represents the random noise ∈ sampled at the i-th noise adding moment i The noise variance β i The difference between α and 1, and α i =1-β i , i=1,…,t; Indicates the p-norm; Represents the acquisition of random noise ∈ t As a variable, change the random noise ∈ t The different values ​​of Expectations; E x,c Expressing expectations expectations; S35: Determine whether the loss function L is less than a set threshold. If yes, obtain the final noise prediction network; if no, proceed to step S36; S36: changing the network parameter values ​​in the network parameter set θ of the noise prediction network, and then re-executing steps S33 to S35 using the noise prediction network after the network parameters are changed.

5. The method for high-resolution through-wall radar target imaging based on a conditional diffusion model as claimed in claim 1, characterized in that: In step S1, the three-dimensional radar image I of the target to be detected radar The method to obtain is: Assume that there is a wall with a thickness of d and a dielectric constant of ∈ in the x-axis direction, the antenna and the target to be measured are respectively in the positive and negative directions of the y-axis, and everything except the wall is free space; The radar works in synthetic aperture mode, using N transceiver antenna units with a spacing of d1 to transmit a stepped frequency continuous wave signal containing M frequency points. Assuming that there are Q targets to be measured on the other side of the wall, the frequency domain echo signal S(m,n) received by the nth transceiver antenna unit at the mth frequency point is: Among them, σ wall and σ q represents the scattering coefficient of the wall and the qth target to be measured, q=1,2,…,Q,S noise (m,n) represents the noise signal of the nth transmitting and receiving antenna unit at the mth frequency point, f m is the frequency of the mth frequency point, τ wall represents the round-trip delay between N transmitting and receiving antenna units and the wall, τ q,n represents the round-trip delay between the nth antenna and the qth target to be measured; Perform inverse Fourier transform on the frequency domain echo signals S(m,n) received by the N transceiver antenna units respectively to obtain the time domain echo corresponding to each transceiver antenna unit; Perform BP imaging on the time domain echo corresponding to each transceiver antenna unit to obtain the BP imaging I of each transceiver antenna unit in the target area. n ; The BP image I of each transceiver antenna unit in the target area n Perform coherent superposition to obtain a superimposed image The phase coherence factor PCF is used to weight the superimposed image to obtain the three-dimensional radar image I of the target to be measured. radar =PCF·I BP .

6. A through-wall radar target high-resolution imaging method based on a conditional diffusion model as claimed in claim 5, characterized in that: The method for obtaining the phase coherence factor PCF is: PCF=1-std(e -jφ ) in, represents the calculated standard deviation; j represents the imaginary part; φ l Indicates the phase of the echo signal reflected by the target to be measured to the lth transceiver antenna unit. The serial number of the transceiver antenna unit is l = 0, 1, ..., N-1, φ n It represents the phase of the echo signal reflected by the target to be measured to the nth transceiver antenna unit, and φ represents the phase of the total echo signal reflected by the target to be measured to all transceiver antenna units.

7. The method for high-resolution through-wall radar target imaging based on a conditional diffusion model as claimed in claim 1, characterized in that: In step S2, the target depth image I of the target to be measured optical The method to obtain is: According to the perspective principle of light, the optical depth camera is placed at the center of the radar's transceiver antenna array to obtain the same viewing angle as the radar and the same as the 3D radar image I radar Matching target depth images I optical .

Images