A wall-penetrating radar high-resolution imaging method with less forgetting

Abstract

The application discloses a few-forgetting through-wall radar high-resolution imaging method. The specific steps comprise the following steps: training the optical image assisted high-resolution imaging network of a source domain, so that the shape of the target of the source domain can be restored and used as a pre-training network; loading the pre-training model to initialize the corresponding part parameters of the target domain network LFHRNet with a multi-branch structure, then training the target domain network LFHRNet on a target domain data set, one branch learns new knowledge of the target domain during training, the other branch is frozen to retain the knowledge of the source domain, and a routing network is introduced to perform accurate expert allocation. The application effectively reduces the forgetting of the knowledge of the source domain, and enables the network to have the ability of high-resolution imaging of static targets in the source domain and the target domain.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to a high-resolution imaging method for through-wall radar with minimal forgetting. Background Technology

[0002] With the continuous advancement of urbanization, the demand for concealed space detection has increased significantly, playing an important role in fields such as security, rescue, and medical care.

[0003] Through-wall radar (TWR) uses low-frequency electromagnetic waves to penetrate buildings and image targets in obscured space, making it a primary technology for detecting targets in obscured space.

[0004] Current traditional through-wall radar imaging algorithms, limited by antenna aperture and array spacing, produce blurry, spot-like images regardless of the target's shape, failing to directly represent target shape information and hindering direct identification and use. Deep learning models cannot generalize to new targets with significant morphological differences, and sufficient labeled data is difficult to collect for training in reality. Transfer learning can improve imaging performance in target domains with few samples, but the network forgets knowledge from the source domain, such as... Figure 1 As shown, the network no longer possesses the ability to correctly recover the shape of the target in the source domain. Therefore, mitigating the network's forgetting of source domain knowledge and enabling the network to simultaneously perform high-resolution imaging of targets in both the source and target domains is a problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention proposes a less-forgetting high-resolution imaging method for through-wall radar. Addressing the problems of insufficient imaging resolution and network forgetting source domain knowledge after fine-tuning in the target domain, this invention designs a less-forgetting high-resolution imaging network (LFHRNet) for the target domain, specifically a multi-branch network structure. First, this invention trains an optically assisted high-resolution imaging network in the source domain, enabling it to recover the shape of the target in the source domain, and uses this network as a pre-trained network. For training the target domain network LFHRNet, one branch learns new target domain knowledge, while another branch remains unchanged to retain source domain knowledge. A routing network is also introduced for expert allocation. Ultimately, this mitigates the forgetting of source domain knowledge, enabling the network to recover the shape of the target in the target domain as well as the shape of the target in the source domain.

[0006] The technical solution of this invention is:

[0007] A high-resolution imaging method for through-wall radar with minimal forgetting includes the following steps:

[0008] Step 1: Obtain the BP radar imaging results and corresponding optical images, and construct a dataset:

[0009] Consider as Figure 3 The size shown is N, which includes a single reflection. x ×N y In the imaging scenario, the radar transmitted signal s(t) is a stepped-frequency signal with a starting frequency of f0 and a step frequency interval of Δf, containing K frequency points. Therefore, the frequencies within the bandwidth are:

[0010] f k =f0+kΔf,k=0,...,K-1.

[0011] Assume the wall thickness is d wall The dielectric constant is ε r If the antenna is moved at intervals d1, then the position of the m-th antenna is (x m Considering the direct and primary reflection signals from the target, the direct reflection signal from the wall, and the reflection signals between the targets, the received echo signal at the m-th antenna position and the k-th frequency can be regarded as the target reflection signal. Direct signal reflection from the wall Reflected signals between targets and noise v m,k The superposition of can be expressed as:

[0012]

[0013] Where P is the number of targets, W is the number of direct reflection paths from the wall, and R is the number of multiple primary reflection paths from the target. σ p and σ pq These are the reflection coefficients of the direct reflection path of the w-th wall, the path between the p-th target, and the path between the p-th and q-th targets, respectively. Figure 3 Taking Path-A as an example, τ pm It is the two-way delay from the m-th antenna to the p-th target. The two-way delay of the direct reflection path of the w-th wall can be expressed as:

[0014]

[0015] Where l1 is the propagation distance of the electromagnetic wave in the wall, and l2 and l3 are the propagation distances of the electromagnetic wave in the air. Let be the length of the direct reflection path of the w-th wall, and v be the propagation speed of the electromagnetic wave in the wall. According to Snell's law, v can be expressed as:

[0016]

[0017] Where c is the speed of light.

[0018] For 3D BP imaging, the imaging region is divided into a grid along the xyz axes. Echo data from M antennas at different locations are collected. The echo data is then subjected to IFFT transformation to obtain s(m,k). The two-way delay τ between the m-th antenna position and the h-th pixel in the imaging region is calculated. mh Then the imaging result of the h-th pixel is:

[0019]

[0020] Calculate the image of the entire region by measuring all pixels.

[0021] Repeat the above steps to collect echo data from different targets at different rotation angles and perform imaging to construct a complete simulation dataset. Then conduct field experiments to construct a complete experimental dataset.

[0022] Step 2: Build and train a source-domain optical image-assisted high-resolution imaging network:

[0023] This invention uses Conditional Generative Adversarial Networks (cGANs) for high-resolution imaging. A cGAN consists of a generator and a discriminator, such as... Figure 4 As shown. The network input is a 3D radar image of size 32×32×32 and a 2D tag of size 128×128.

[0024] The generator comprises an encoder and a decoder. The encoder-decoder structure extracts abstract features from the input 3D radar image and uses the decoder to reconstruct a high-quality 2D image from this compressed representation. The encoder consists of five sequentially connected 3D convolutional layers, mapping the 3D radar image into a 2048×1 vector. Each 3D convolutional layer is followed by a BatchNorm layer and a LeakyReLU layer. The BatchNorm layer stabilizes training and improves network convergence, while the LeakyReLU layer introduces non-linearity to prevent gradient vanishing or descent. The decoder consists of five sequentially connected 2D transposed convolutional layers, a BatchNorm layer, and a ReLU layer, upsampling the vector to reconstruct a 128×128 image. Simultaneously, to preserve high-frequency information, the network employs skip connections, concatenating the generated image output from the decoder with the high-frequency channels of the radar image, and then passing the concatenated image through two transposed convolutional layers to obtain the final output.

[0025] The discriminator simultaneously inputs a 3D radar image and a 2D label or generated image, and uses two different networks to transform both into vectors. The structure of the 3D input is the same as the encoder structure of the generator, while the 2D input consists of eight sequentially connected 2D convolutional layers, a BatchNorm layer, and a LeakyReLU layer. The outputs of both are concatenated into a vector, which is then classified through a Linear layer. The classification result D(x,y) or D(x,G(x)) represents the probability that the 2D input is the true label. During the training phase, the discriminator distinguishes between the true label y and the generated image G(x), while the generator generates images that meet the constraints, making them as indistinguishable as possible to the discriminator. This is a game theory process, and the loss function of this process can be expressed as:

[0026]

[0027] To improve the quality of the generated images, this invention introduces root mean square error (MSE) and perceptual loss during the training of the cGAN network. The generator loss function during training is then:

[0028] L(G)=L GAN (G)+λ MSE L MSE +λ p L p

[0029] Among them, L GAN (G) is the adversarial loss of the generative adversarial network itself, L MSE It is the MSE loss between the real label and the generated image, to ensure the accuracy and consistency of the generated image and label at the pixel level. p The perceptual loss is calculated using a pre-trained VGG network to make the generated images more consistent with human visual perception. λ MSE and λ p This is a constant that ensures the generated image is accurate in detail and directly recognizable by the human eye. During training, the generator is optimized to maximize the probability that the generated image is classified as the true label, while minimizing L. MSE and L p The discriminator is optimized to minimize the probability of misclassification between the generated image and the real label.

[0030] Step 3: Build and train the target domain network LFHRNet:

[0031] The structure of the target domain network LFHRNet designed in this invention is as follows: Figure 5As shown, the target domain network has the same encoder structure as the source domain network and integrates two parallel branches as experts specializing in handling different tasks. The tasks in the source and target domains are respectively high-resolution imaging of the source domain input and high-resolution imaging of the target domain input. When training the target domain network on target domain data, one branch is initialized using the source domain network and frozen during training to retain source domain knowledge, while the other expert's parameters are updated to learn new target domain knowledge. The training objective can be represented as:

[0032]

[0033] Among them, D t This is the target domain dataset, containing 3D radar images x and 2D labels y. Θ encoder These are parameters pre-trained in the source domain, containing knowledge learned in the source domain. The encoder of the target domain network uses Θ. encoder Initialize and in D t The parameters are frozen during training as a ready-made feature extractor. ΔW is the parameter learned on the target domain data, obtained by minimizing the loss L. t The optimal ΔW is obtained.

[0034] To integrate two experts into a single network, this invention introduces a routing network to enable expert collaboration and precise allocation. When a 3D radar image is input into the target domain network, this image may originate from either the source or target domain. Due to the different data characteristics and distributions of the source and target domains, this invention selects one expert instead of weighting the results of different experts. The routing network is implemented using a simple CNN and incorporates domain labels y. i To balance expert selection, for a given 3D radar image input, the routing network determines the source of the input and assigns an expert to process the input data, obtaining the corresponding high-resolution imaging result. The output can then be expressed as:

[0035] o=αW0x encoder +βΔWx encoder

[0036] Where W0 represents the parameters of the frozen branch, and ΔW represents the parameters of the trainable branch. α and β are the classification results of the routing network, which can be expressed as:

[0037]

[0038] Where ROUND(·) is the rounding operation, C(·) is the linear layer of the binary classification operation, and W g These are the trainable parameters of the routing network. The final output can be represented as:

[0039] o=αW0x encoder+βΔWx encoder

[0040] =ROUND(C(xW g ))×[W0x encoder ,ΔWx encoder ] T

[0041] Through the above steps, the routing network can assign different experts to process inputs from different domains. Since the parameters of the expert recovering the target shape from the source domain are not updated, the knowledge of the source domain is preserved. Simultaneously, the network learns new knowledge from the target domain.

[0042] Step 4: Input the BP radar image of the source domain or target domain sample into the target domain network to obtain a high-resolution imaging result containing shape and contour information.

[0043] Beneficial effects:

[0044] The network designed in this invention can learn new knowledge in the target domain while reducing the forgetting of knowledge in the source domain, and simultaneously possesses the ability to perform high-resolution imaging of targets in both the source and target domains. This represents an effective, low-forgetting, through-wall radar high-resolution imaging method.

[0045] 1. Compared to the blurry, spot-like imaging results of traditional imaging algorithms, the network designed in this invention can achieve high-resolution imaging through-wall radar. The imaging results include the shape and category information of the imaged target, which can be directly identified and used.

[0046] 2. Compared to deep learning methods that train from scratch, this invention uses transfer learning, which can generate higher quality images with fewer iterations when the number of samples in the target domain is small. These images have clearer shapes and realistic details and are closer to the labels.

[0047] 3. Compared to ordinary neural networks, which forget the knowledge of the source domain after transfer learning and can only perform high-resolution imaging of the target domain input but no longer can perform high-resolution imaging of the source domain input, the network designed in this invention can alleviate the forgetting of the source domain knowledge and has the ability to perform high-resolution imaging of the target in both the source and target domains. Attached Figure Description

[0048] Figure 1 This explains the phenomenon that the network forgets source domain knowledge after fine-tuning the target domain;

[0049] Figure 2 A flowchart for the implementation method;

[0050] Figure 3 , is a schematic diagram of a signal scenario;

[0051] Figure 4 The generator and discriminator structures for optically assisted high-resolution imaging networks;

[0052] Figure 5 A schematic diagram of the target domain network LFHRNet structure;

[0053] Figure 6 Here are examples of simulation datasets, where (a) is a simulation model diagram, (b) is the dataset label, (c) is a 3D BP radar chart, and (d) and (e) are the top view and front view, respectively.

[0054] Figure 7 Figure 1 is a schematic diagram of the actual test scenario, where (a) is the scenario without wall penetration, (b) is the radar behind the wall in the scenario with wall penetration, and (c) is a schematic diagram of the target placement in the scenario with wall penetration.

[0055] Figure 8 A comparison of image generation results from LFHRNet and those trained from scratch with the number of training iterations;

[0056] Figure 9 Comparison of cosine similarity and PSNR with different training sample numbers for LFHRNet and training from scratch;

[0057] Figure 10 Here are some examples of test results for LFHRNet on simulation data in the target domain;

[0058] Figure 11 Here are some examples of test results for LFHRNet on real-world data in the target domain;

[0059] Figure 12 Here are some examples of test results for LFHRNet on simulation and experimental data in the source domain. Detailed Implementation

[0060] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0061] The purpose of this invention is to overcome the limitations of existing traditional through-wall radar imaging algorithms, such as insufficient resolution, difficulty in generalizing deep learning methods to new targets, poor image quality due to insufficient data on new targets, and the inability of fine-tuning the network to perform high-resolution imaging of source domain targets. This invention provides a high-resolution through-wall radar imaging method with less forgetting. Figure 2 This is a flowchart of an embodiment of the present invention, the specific steps of which include:

[0062] Step 1: Obtain the BP radar imaging results and corresponding optical images, and construct the data:

[0063] Consider as Figure 3The size shown is N, which includes a single reflection. x ×N y In the imaging scenario, the radar transmitted signal s(t) is a stepped-frequency signal with a starting frequency of f0 and a step frequency interval of Δf, containing K frequency points. Therefore, the frequencies within the bandwidth are:

[0064] f k =f0+kΔf,k=0,...,K-1.

[0065] Assume the wall thickness is d wall The dielectric constant is ε r If the antenna is moved at intervals d1, then the position of the m-th antenna is (x m Considering the direct and primary reflection signals from the target, the direct reflection signal from the wall, and the reflection signals between the targets, the received echo signal at the m-th antenna position and the k-th frequency can be regarded as the target reflection signal. Direct signal reflection from the wall Reflected signals between targets and noise v m,k The superposition of can be expressed as:

[0066]

[0067] Where P is the number of targets, W is the number of direct reflection paths from the wall, and R is the number of multiple primary reflection paths from the target. σ p and σ pq These are the reflection coefficients of the direct reflection path of the w-th wall, the path between the p-th target, and the path between the p-th and q-th targets, respectively. Figure 3 Taking Path-A as an example, τ pm It is the two-way delay from the m-th antenna to the p-th target. The two-way delay of the direct reflection path of the w-th wall can be expressed as:

[0068]

[0069] Where l1 is the propagation distance of the electromagnetic wave in the wall, and l2 and l3 are the propagation distances of the electromagnetic wave in the air. Let be the length of the direct reflection path of the w-th wall, and v be the propagation speed of the electromagnetic wave in the wall. According to Snell's law, v can be expressed as:

[0070]

[0071] Where c is the speed of light.

[0072] For 3D BP imaging, the imaging region is divided into a grid along the xyz axes. Echo data from M antennas at different locations are collected. The echo data is then subjected to IFFT transformation to obtain s(m,k). The two-way delay τ between the m-th antenna position and the h-th pixel in the imaging region is calculated. mh Then the imaging result of the h-th pixel is:

[0073]

[0074] Calculate the image of the entire region by measuring all pixels.

[0075] Repeat the above steps to collect echo data from different targets at different rotation angles and perform imaging to construct a complete simulation dataset. Then conduct field experiments to construct a complete experimental dataset.

[0076] Step 2: Build and train a source-domain optical image-assisted high-resolution imaging network:

[0077] This invention uses Conditional Generative Adversarial Networks (cGANs) for high-resolution imaging. A cGAN consists of a generator and a discriminator, such as... Figure 4 As shown. The network input is a 3D radar image of size 32×32×32 and a 2D tag of size 128×128.

[0078] The generator comprises an encoder and a decoder. The encoder-decoder structure extracts abstract features from the input 3D radar image and uses the decoder to reconstruct a high-quality 2D image from this compressed representation. The encoder consists of five sequentially connected 3D convolutional layers, mapping the 3D radar image into a 2048×1 vector. Each 3D convolutional layer is followed by a BatchNorm layer and a LeakyReLU layer. The BatchNorm layer stabilizes training and improves network convergence, while the LeakyReLU layer introduces non-linearity to prevent gradient vanishing or descent. The decoder consists of five sequentially connected 2D transposed convolutional layers, a BatchNorm layer, and a ReLU layer, upsampling the vector to reconstruct a 128×128 image. Simultaneously, to preserve high-frequency information, the network employs skip connections, concatenating the generated image output from the decoder with the high-frequency channels of the radar image, and then passing the concatenated image through two transposed convolutional layers to obtain the final output.

[0079] The discriminator simultaneously inputs a 3D radar image and a 2D label or generated image, and uses two different networks to transform both into vectors. The structure of the 3D input is the same as the encoder structure of the generator, while the 2D input consists of eight sequentially connected 2D convolutional layers, a BatchNorm layer, and a LeakyReLU layer. The outputs of both are concatenated into a vector, which is then classified through a Linear layer. The classification result D(x,y) or D(x,G(x)) represents the probability that the 2D input is the true label. During the training phase, the discriminator distinguishes between the true label y and the generated image G(x), while the generator generates images that meet the constraints, making them as indistinguishable as possible to the discriminator. This is a game theory process, and the loss function of this process can be expressed as:

[0080]

[0081] To improve the quality of the generated images, this invention introduces root mean square error (MSE) and perceptual loss during the training of the cGAN network. The generator loss function during training is then:

[0082] L(G)=L GAN (G)+λ MSE L MSE +λ p L p

[0083] Among them, L GAN (G) is the adversarial loss of the generative adversarial network itself, L MSE It is the MSE loss between the real label and the generated image, to ensure the accuracy and consistency of the generated image and label at the pixel level. p The perceptual loss is calculated using a pre-trained VGG network to make the generated images more consistent with human visual perception. λ MSE and λ p This is a constant that ensures the generated image is accurate in detail and directly recognizable by the human eye. During training, the generator is optimized to maximize the probability that the generated image is classified as the true label, while minimizing L. MSE and L p The discriminator is optimized to minimize the probability of misclassification between the generated image and the real label.

[0084] Step 3: Build and train the target domain network LFHRNet:

[0085] The structure of the target domain network LFHRNet designed in this invention is as follows: Figure 5As shown, the target domain network has the same encoder structure as the source domain network and integrates two parallel branches as experts specializing in handling different tasks. The tasks in the source and target domains are respectively high-resolution imaging of the source domain input and high-resolution imaging of the target domain input. When training the target domain network on target domain data, one branch is initialized using the source domain network and frozen during training to retain source domain knowledge, while the other expert's parameters are updated to learn new target domain knowledge. The training objective can be represented as:

[0086]

[0087] Among them, D t This is the target domain dataset, containing 3D radar images x and 2D labels y. Θ encoder These are parameters pre-trained in the source domain, containing knowledge learned in the source domain. The encoder of the target domain network uses Θ. encoder Initialize and in D t The parameters are frozen during training as a ready-made feature extractor. ΔW is the parameter learned on the target domain data, obtained by minimizing the loss L. t The optimal ΔW is obtained.

[0088] To integrate two experts into a single network, this invention introduces a routing network to enable expert collaboration and precise allocation. When a 3D radar image is input into the target domain network, this image may originate from either the source or target domain. Due to the different data characteristics and distributions of the source and target domains, this invention selects one expert instead of weighting the results of different experts. The routing network is implemented using a simple CNN and incorporates domain labels y. i To balance expert selection, for a given 3D radar image input, the routing network determines the source of the input and assigns an expert to process the input data, obtaining the corresponding high-resolution imaging result. The output can then be expressed as:

[0089] o=αW0x encoder +βΔWx encoder

[0090] Where W0 represents the parameters of the frozen branch, and ΔW represents the parameters of the trainable branch. α and β are the classification results of the routing network, which can be expressed as:

[0091]

[0092] Where ROUND(·) is the rounding operation, C(·) is the linear layer of the binary classification operation, and W g These are the trainable parameters of the routing network. The final output can be represented as:

[0093] o=αW0x encoder+βΔWx encoder

[0094] =ROUND(C(xW g ))×[W0x encoder ,ΔWx encoder ] T

[0095] Through the above steps, the routing network can assign different experts to process inputs from different domains. Since the parameters of the expert recovering the target shape from the source domain are not updated, the knowledge of the source domain is preserved. Simultaneously, the network learns new knowledge from the target domain.

[0096] Step 4: Input the BP radar image of the source domain or target domain sample into the target domain network to obtain a high-resolution imaging result containing shape and contour information.

[0097] Thus, a high-resolution imaging method for through-wall radar with minimal forgetting has been completed.

[0098] Example

[0099] Following step 1, the simulated transmitted signal is set to a stepped-frequency signal with an antenna step of 5cm and an imaging area of ​​3m×3m×4m. A dataset containing tables, chairs, people, and RPGs at different rotation angles is constructed within the simulation scenario. The source domain dataset includes Chair-1. sim ), Chair-2 sim Table 1 sim ) and RPG, the target domain dataset contains Table-2 sim ) and people (Human) sim The transmitted signal was set to a stepped-frequency signal with a frequency interval of 2MHz and a frequency range of 1.7GHz-2.2GHz. The antenna array consisted of 10 transmitters and 10 receivers, approximately 40cm × 40cm in size. In a non-wall-penetrating scenario, the distance between the radar and the target was 2m. In a wall-penetrating scenario, the wall thickness was 0.2m, with the radar close to one side of the wall and the target 2m away from the other side. Data sets containing tables, chairs, and people at different rotation angles were collected. The source domain included Chair-1. exp ) and table (Table-1) exp The target domain includes Chair-2. exp ) and people (Human) exp ).

[0100] According to step 2, set the batch size to 4, the optimizer to the Adam optimizer with an initial learning rate of 1e-2, train the source domain cGAN network so that it can recover the shape of the source domain target and use it as a pre-trained network.

[0101] According to step 3, some parameters of the target domain network LFHRNet are initialized using the cGAN network obtained in step 2, and then the target domain network is trained using the target domain dataset. For comparison, this invention also uses two other strategies: training a cGAN network from scratch using the target domain dataset, and using a cGAN network in the target domain, initializing it, and then fine-tuning it on the target domain dataset.

[0102] According to step 4, the target domain network is used as the final high-resolution imaging network to process the input radar images in the source and target domains. The size of the target domain training dataset is an important factor affecting the quality of the generated images. Therefore, this invention analyzes the impact of the target domain training dataset size, using 42, 84, 206, 330, and 412 samples in the target domain, respectively, and training is performed from scratch on the same dataset. The results are then compared.

[0103] Figure 8 To train generated images on 42 target domain samples for different epochs, LFHRNet was able to generate higher quality images with sharper shapes and realistic details, and closer to the labels, with fewer iterations compared to training from scratch. The fewer the iterations, the more pronounced LFHRNet's advantage in generated image quality. As the number of iterations increased, images generated by LFHRNet gradually approached those generated by training from scratch, but LFHRNet still maintained its advantage at the end of training. Figure 9 The graph shows the changes in cosine similarity and PSNR calculated on the training samples during the training process. As the number of iterations increases, both the cosine similarity and PSNR of training from scratch and LFHRNet gradually increase. Figure 9 As shown in (a) and (b), when the number of training samples is limited, LFHRNet's cosine similarity and PSNR are higher than those trained from scratch, indicating that initializing a portion of the target domain network and freezing it during training when the number of training samples is limited is beneficial. Figure 9 As shown in (c)(d), when the training data is sufficient and the number of iterations is small, LFHRNet generates better results than training from scratch, but when the number of iterations increases, the results generated by the two are similar.

[0104] Table 1 shows the results of cosine similarity, PSNR, LPIPS, and SSIM on the target domain test set when the target domain training dataset size is different. When the number of training samples is limited, LFHRNet outperforms training from scratch; when the training data is sufficient, training from scratch outperforms LFHRNet. Regardless of the number of training samples, fine-tuning is the most effective of the three because it initializes and updates all layers, transferring knowledge from the source domain and making the training more comprehensive. Although LFHRNet also utilizes knowledge from the source domain, it freezes some layers during training. Furthermore, the routing network suffers from expert assignment errors, and the input to the target domain may be reconstructed to be closer to the shape of the target in the source domain. Although LFHRNet's performance on the target domain is slightly inferior to fine-tuning, it still meets the requirements for achieving high-resolution imaging. Figure 10 and Figure 11 Examples of test results for simulated and measured target domains are shown, demonstrating that LFHRNet can generate high-resolution imaging results that include shape, contour, and texture information. LFHRNet's main advantage lies in the source domain.

[0105] Table 1. Statistical Comparison of Results on the Target Domain Test Set

[0106]

[0107]

[0108] Although fine-tuning improves the imaging results of the target domain samples, it forgets the knowledge learned in the source domain. To verify the effectiveness of LFHRNet in preventing the forgetting of source domain knowledge, this invention tested it on the source domain. Figure 12The test results are shown on simulated and measured data in the source domain. Column (a) shows target illustrations in the simulated and measured scenarios, column (b) shows the labels, column (c) shows the 3D radar images, columns (d) and (e) are top and front views of the 3D radar images, respectively, and columns (f) to (h) show the generation results of training from scratch, fine-tuning, and LFHRNet, respectively. It can be seen that the network fine-tuned in the target domain forgets the knowledge learned in the source domain and cannot correctly reconstruct the shape of the target in the source domain. The generated results become just as bad as those trained from scratch, even though training from scratch has never been done on source domain data. This is because all parameters are updated during fine-tuning in the target domain, destroying the knowledge originally stored in the network. LFHRNet, on the other hand, can correctly generate high-resolution images close to the labels, demonstrating its effectiveness in avoiding the forgetting of source domain knowledge. This is because when training LFHRNet in the target domain, an expert is frozen to preserve source domain knowledge. Furthermore, considering that LFHRNet can also perform high-resolution imaging of target domain samples, it can be seen that the two experts in LFHRNet have acquired the ability to handle tasks in different domains, and the routing network can assign appropriate experts to different inputs. Table 2 shows the results of cosine similarity, PSNR, LPIPS, and SSIM on the source domain test set when the target domain training dataset size is different, showing that LFHRNet has a significant advantage. Moreover, the size of the target domain training dataset has little impact on the performance of LFHRNet in the source domain.

[0109] Table 2: Statistical Comparison of Results on the Source Domain Test Set

[0110]

[0111] In summary, this invention proposes a high-resolution imaging method for through-wall radar with minimal forgetting. LFHRNet employs a multi-branch structure, where one branch is updated during training to learn new knowledge from target domain samples, and another branch is frozen to retain source domain knowledge. Simulation and field experiments demonstrate that the proposed method effectively alleviates the problem of the network forgetting source domain knowledge after fine-tuning in the target domain, and can simultaneously achieve high-resolution imaging of targets in both the source and target domains.

[0112] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A high-resolution imaging method for through-wall radar with minimal forgetting, characterized in that, include: Step 1: Obtain BP radar imaging results and corresponding optical images, and construct a dataset; Step 2: Build and train a source-domain optical image-assisted high-resolution imaging network, namely a conditional generative adversarial network (cGAN). The cGAN consists of a generator and a discriminator. The input to the network is a 3D radar image of size 32×32×32. And 2D labels with a size of 128×128 ; Step 3: Build and train the target domain network LFHRNet. It has the same encoder structure as the source domain network and integrates two parallel branches as experts specializing in handling different tasks. The tasks in the source and target domains are high-resolution imaging of the source domain input and high-resolution imaging of the target domain input, respectively. When training the target domain network on target domain data, one branch is initialized using the source domain network and frozen during training to retain source domain knowledge, while the other expert's parameters are updated to learn new target domain knowledge. The training objective is represented as follows: ； in, It is a target domain dataset containing 3D radar images. and 2D tags , These are parameters pre-trained in the source domain, containing knowledge learned in the source domain, and are used by the encoder of the target domain network. Initialize and in Freeze during training as a ready-made feature extractor. These are parameters learned from the target domain data, achieved by minimizing the loss. To obtain the optimal ; A routing network is introduced to enable expert collaboration and precise allocation. When a 3D radar image is input into the target domain network, this image may originate from either the source or target domain. Due to the different data characteristics and distributions of the source and target domains, one expert is selected instead of simply adding the results of different experts by weight. The routing network is implemented using a CNN and domain labels are introduced. To balance expert selection; Step 4: Input the BP radar images of the target domain and source domain samples into the target domain network to obtain high-resolution imaging results containing shape and contour information.

2. The high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 1, characterized in that, In step 1, for the size containing a single reflection, Imaging scene, radar transmission signal It is a step frequency signal, with a starting frequency of The step frequency interval is ,Include If there are multiple frequency points, then the frequencies within the bandwidth are: 。 3. The high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 2, characterized in that, In step 1, let the wall thickness be... The dielectric constant is , to interval Moving the antenna, then the first m The antenna positions are Considering the direct reflection and primary reflection signals from the target, the direct reflection signal from the wall, and the reflection signals between the targets, then the... m The antenna position, the first k The received echo signal at each frequency is considered as the target reflection signal. Direct signal reflection from the wall Reflected signals between targets and noise The superposition of these elements is represented as: ； in, P It is the target quantity. W It is the number of direct reflection paths from the wall. R It is the number of multiple primary reflection paths on the target. , and They are the first w The wall's direct reflection path, the first p The first goal and the first p, q Reflection coefficient of the path between targets; It is the first m The antenna to the first p Two-way delay for each target; It is the first w Two-way delay of the direct reflection path from the wall; ； ； in It is the distance that electromagnetic waves travel within the wall. and It is the distance that electromagnetic waves travel in the air. It is the first w The length of the direct reflection path of the wall It is the speed of electromagnetic wave propagation in the wall, according to Snell's law, Represented as ； in It's the speed of light.

4. The high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 2, characterized in that, In step 1, for three-dimensional BP imaging, the imaging area is along... xyz The axis is divided into a grid for data collection. M Echo data from antennas at different locations were analyzed, and IFFT transformation was performed on the echo data. Calculate the first m The antenna position and imaging area h Two-way latency per pixel Then the first h The imaging result of each pixel is 。 5. The high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 1, characterized in that, In step 2, the generator includes an encoder and a decoder. The encoder-decoder structure can extract abstract features from the input 3D radar image and use the decoder to reconstruct a high-quality 2D image from the compressed representation. The encoder consists of 5 sequentially connected 3D convolutional layers, which can map the 3D radar image into a 2048×1 vector. Each 3D convolutional layer is followed by a BatchNorm layer and a LeakyReLU layer. The decoder consists of 5 sequentially connected 2D transposed convolutional layers, BatchNorm layers, and ReLU layers, which upsample the vector to reconstruct a 128×128 image.

6. The high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 5, characterized in that, In step 2, the network uses skip connections to stitch the generated image output by the decoder with the high-frequency channel of the radar image, and then obtains the final output through two transposed convolutional layers.

7. The high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 5, characterized in that, In step 2, the discriminator simultaneously inputs a 3D radar image. and 2D tags Or generate an image The two inputs are transformed into vectors using two different networks. The 3D input has the same structure as the encoder of the generator, while the 2D input consists of eight sequentially connected 2D convolutional layers, a BatchNorm layer, and a LeakyReLU layer. The outputs of both are concatenated into a single vector, which is then classified using a Linear layer. The classification result... or This represents the probability that the 2D input is the true label.

8. A high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 7, characterized in that, In step 2, during the training phase, the discriminator is used to distinguish between real labels. and the generated image The generator is used to generate images that meet the conditional constraints, making them as indistinguishable as possible to the discriminator. The loss function is expressed as: ； When training the cGAN network, the root mean square error (MSE) and perceptual loss are introduced. Therefore, the generator loss function during training is: ； in, It is the adversarial loss of the generative adversarial network itself. It is the MSE loss between the real label and the generated image. The perceptual loss is calculated using a pre-trained VGG network. and It is a constant.

9. A high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 8, characterized in that, In step 2, during training, the generator is optimized to maximize the probability that the generated image is classified as the true label and minimize the probability of the true label being defined. and The discriminator is optimized to minimize the probability of misclassification between the generated image and the real label.

10. A high-resolution imaging method for through-wall radar with minimal forgetting as described in claim 1, characterized in that, In step 3, for a 3D radar image input, the routing network determines the source of the input and assigns an expert to process the input data to obtain the corresponding high-resolution imaging result; the output is represented as: ； in, Parameters for freezing branches, For the parameters of the trainable branch, and This is the classification result of the routing network, represented as: ； in, For rounding operations, For binary classification operations, a linear layer, These are the trainable parameters of the routing network; the final output is represented as... 。

Images