Deformation target recovery method based on deep unfolding reconstruction and generative adversarial post-processing

Abstract

This invention discloses a deformation target recovery method based on deep unfolding reconstruction and generative adversarial post-processing, comprising: S1, collecting and preprocessing the HRRP data of the target to be recovered; S2, performing sparse recovery on the preprocessed HRRP data using a Bayesian inference algorithm; S3, obtaining the deformation components to be recovered based on the sparse representation coefficients of the target; S4, inputting the recovered components into a trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation components for the current iteration; S5, obtaining the recovered HRRP data when the preset maximum number of iterations is reached; otherwise, returning to S2 and iterating again until the preset maximum number of iterations is reached. This invention employs a model based on deep unfolding reconstruction and generative adversarial post-processing, improving computational efficiency and significantly enhancing the realism and recovery accuracy of the recovered HRRP data.

Description

Technical Field

[0001] This invention belongs to the field of radar target processing technology, specifically relating to a method for deformable target recovery based on deep unfolding reconstruction and generative adversarial post-processing. Background Technology

[0002] High-resolution range profile (HRRP) is a one-dimensional form of radar echo data. It represents the superposition of projected echoes from the scattering centers of a target acquired by a broadband radar along the radar's line-of-sight direction. It can reflect the target's scale information and the spatial distribution characteristics of the scattering centers along the range direction. Because this type of echo data is relatively easy to acquire, has low data dimensionality, and a relatively mature processing flow, it is widely used in the field of radar automatic target recognition technology.

[0003] In the research and application of existing radar target recognition methods, the relevant models are usually designed based on ideal detection conditions, such as assuming that the target is in a high signal-to-noise ratio environment, the target structure is complete and unobstructed, and the number of training samples is sufficient and the target types are complete. However, in actual radar detection scenarios, the above ideal conditions are often difficult to meet simultaneously. Especially in complex environments, the target may be obstructed by external objects, or its own attitude may change, or the target configuration may be altered, resulting in the loss or change of scattering information of some range cells, thus causing a mismatch between the actual high-resolution range image and the training samples.

[0004] The aforementioned target deformation mainly originates from the following situations: First, changes in target configuration, such as the target being equipped with different types of additional devices or weapon systems according to different mission requirements, thereby causing changes in the target's shape and scattering characteristics; second, the target being obscured by the environment or by human camouflage, causing partial obscuring of the target's local structure in the radar line of sight. Due to the presence of deformation factors, there is a significant difference between the high-resolution range image echo of deformed targets and the echo of standard targets, thus reducing the recognition accuracy based on template matching or data-driven models.

[0005] Under low signal-to-noise ratio (SNR) conditions, the superposition of noise interference and deformation factors further weakens the effective scattering information of the target, leading to a more significant performance degradation of existing recognition models. Therefore, how to effectively recover high-resolution range images of deformed targets under low SNR conditions, in order to reduce the impact of deformation and noise on target recognition performance, has become one of the urgent technical problems to be solved in the field of radar target recognition. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a method for deformable target recovery based on deep unfolding reconstruction and generative adversarial post-processing.

[0007] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for deformable target recovery based on deep unfolding reconstruction and generative adversarial post-processing, the method comprising: S1. Collect the HRRP data of the target to be recovered and preprocess it to obtain the preprocessed HRRP data to be recovered; S2. Perform sparse recovery on the preprocessed HRRP data to be recovered according to the Bayesian inference algorithm to obtain the sparse representation coefficients of the target to be recovered; S3. Obtain the deformation component to be recovered based on the sparse representation coefficients of the target to be recovered; S4. Input the deformation component to be recovered into the trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation component of the current round; wherein, the model based on deep unfolding reconstruction and generative adversarial post-processing includes a deep unfolding network and a generative adversarial network; S5. When the preset maximum iteration round is reached, the recovered HRRP data is obtained based on the target deformation component of the current round and the sparse representation coefficient of the target to be recovered; otherwise, return to S2 and iterate again until the preset maximum iteration round is reached.

[0008] Optionally, the training process of the deep unfolded network includes: Obtain the original deformation components; The sparse signal of the training block is divided to obtain the sparse signal of the divided training block. Construct a training dictionary matrix based on the structure of the sparse signals of the divided training blocks; The model of the sparse signal of the divided training block is obtained based on the sparse signal of the divided training block and the training dictionary matrix; The model of the sparse signal of the divided training block is solved by gradient descent, and the recursive formula of the sparse signal of the divided training block is obtained. The sparse signal of the training block after iteration is obtained according to the recursive formula; The sparse signal of the iterative training block is filtered using an adaptive soft thresholding method to obtain the filtered sparse signal of the training block; wherein the threshold of the adaptive soft thresholding method is obtained based on a noise estimation factor. A loss function is constructed based on the filtered sparse signal of the training block and the original deformation components, and the deep unfolded network is trained to obtain the trained deep unfolded network.

[0009] Optionally, the process of obtaining the noise estimation factor includes: Obtain the training deformation components; Calculate the residual signal corresponding to the sparse signal of the partitioned training block based on the training deformation component and the recursive formula; The residual signal is fused with the training deformation component to obtain the fused signal; The fused signal is passed through three fully connected layers and two ReLU activation functions of the deep unfolded network, then activated by Sigmoid, and finally truncated numerically to obtain the noise estimation factor.

[0010] Optionally, the process of obtaining the training deformation components includes: Collect training HRRP data and preprocess it to obtain preprocessed training HRRP data; The preprocessed training HRRP data is superimposed with block sparse signals of different sparsity to obtain the initial training deformation components. Gaussian white noise is superimposed on the initial training deformation components to obtain the noisy deformation components. The trained deformation component is obtained by multiplying the noise-added deformation component with the identity matrix.

[0011] Optionally, the training process of the discriminator in the generative adversarial network includes: The sparse signal of the initial recovery block output by the trained deep unfolded network is input into the generator of the generative adversarial network to obtain generated samples; The generated sample and the original deformation component are input into the discriminator of the generative adversarial network to obtain the first discriminant value of the generated sample and the second discriminant value of the original deformation component, respectively. A first adversarial loss is constructed based on the first discriminant value and the second discriminant value; The discriminator of the generative adversarial network is trained based on the first adversarial loss to obtain the trained discriminator of the generative adversarial network.

[0012] Optionally, the training process of the generator in the generative adversarial network includes: The generated sample is input into the generator of the generative adversarial network to obtain the corrected generated sample; The mean square error loss is recovered based on the corrected generated sample and the original deformation component. The initial recovery constraint loss is constructed based on the sparse signal of the preliminary recovery block and the original deformation components; Calculate the second adversarial loss obtained by inputting the corrected generated sample into the discriminator of the trained generative adversarial network; Calculate the sparsity regularization loss of the corrected generated samples; The total loss function is obtained based on the recovery mean square error loss, the initial recovery constraint loss, the second adversarial loss, and the sparsity regularization loss. The generator of the generative adversarial network is trained according to the total loss function to obtain the trained generator of the generative adversarial network.

[0013] Optionally, the construction process of the trained model based on deep unfolding reconstruction and generative adversarial post-processing includes: The trained generative adversarial network is obtained based on the discriminator and the generator of the trained generative adversarial network. Based on the trained deep unfolding network and the trained generative adversarial network, a trained model based on deep unfolding reconstruction and generative adversarial post-processing is constructed.

[0014] Secondly, the present invention provides a deformable target recovery device based on depth unfolding reconstruction and generative adversarial post-processing, the device comprising: The preprocessing module is used to collect the HRRP data to be recovered from the target to be recovered and to preprocess it to obtain the preprocessed HRRP data to be recovered. The sparse recovery module is used to perform sparse recovery on the preprocessed HRRP data to be recovered according to the Bayesian inference algorithm to obtain the sparse representation coefficients of the target to be recovered. The module for obtaining the deformation components to be recovered is used to obtain the deformation components to be recovered based on the sparse representation coefficients of the target to be recovered. The module for processing the deformation components to be recovered is used to input the deformation components to be recovered into a trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation components of the current round; wherein, the model based on deep unfolding reconstruction and generative adversarial post-processing includes a deep unfolding network and a generative adversarial network; The data recovery module is used to recover HRRP data based on the target deformation component of the current cycle and the sparse representation coefficient of the target to be recovered when the preset maximum iteration cycle is reached; otherwise, it returns to the sparse recovery module and iterates again until the preset maximum iteration cycle is reached.

[0015] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: In the above technical solution, a deep unfolded network (Ada-BlockLISTA) is adopted, which is combined with deep learning to improve computational efficiency while preserving physical interpretability. A generative adversarial network (GAN) is introduced for residual post-processing. The adversarial learning mechanism is used to refine the residual errors in the initial recovered signal, which significantly improves the realism and recovery accuracy of the recovered HRRP data details, thereby reducing the impact of deformation and noise on target recognition performance.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart of a deformable target recovery method based on deep unfolding reconstruction and generative adversarial post-processing provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the framework of a deep unfolded network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the framework of a noise estimation module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a generative adversarial network framework provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of raw aircraft HRRP data provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a deformation component provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of HRRP data after superimposing deformation components and noise, provided by an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the recovery effect of the present invention as provided in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating the recovery of HRRP data according to an embodiment of the present invention; Figure 10 This is a schematic diagram of HRRP data reconstructed using an existing method according to an embodiment of the present invention; Figure 11 This is a schematic diagram of a deformation component reconstructed by a conventional method according to an embodiment of the present invention; Figure 12 This invention provides a deformable target recovery device based on deep unfolding reconstruction and generative adversarial post-processing. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] In recent years, scholars have conducted research on deformable target recognition. Existing HRRP (Human Resonance Target Recognition) methods for deformable targets include the following: 1. A Variant Target Reconstruction and Recognition Method Based on Long Short-Term Memory Networks. In his 2020 master's thesis, "Research on Variant Target Recognition Method Based on Deep Learning," published at Xi'an University of Electronic Science and Technology, Song Xiaolong proposed a variant target reconstruction and recognition method based on Long Short-Term Memory networks. This method assumes that the variant components (deformation components) conform to the characteristics of structured sparsity. It utilizes the sequence modeling capability of LSTM (Long Short-Term Memory network) to automatically learn and capture the spatial dependencies between non-zero elements in the block sparse signal to perform structured sparse modeling of the deformation components. However, this method only focuses on the spatial dependencies between non-zero elements during inference, resulting in poor recovery performance of the deformation components in low signal-to-noise ratio (SNR) scenarios. This method only focuses on spatial dependencies, ignoring the impact of noise levels on the signal recovery threshold. In low SNR scenarios, due to the lack of adaptive perception of noise intensity, its recovery performance is poor, and it is easy to misjudge noise as a signal or an over-smoothed signal.

[0020] 2. A High-Resolution Echo Recovery Method for Variant Aircraft Based on Convolutional Neural Networks. Patent document CN201910201549.X discloses a "High-Resolution Echo Recovery Method for Variant Aircraft Based on Convolutional Neural Networks," which utilizes convolutional neural networks to recover high-resolution echoes from deformed aircraft into high-resolution echoes from undeformed aircraft. A one-dimensional convolutional neural network suitable for high-resolution echoes is constructed and trained for high-resolution echo recovery of deformed aircraft. This method significantly reduces the impact of deformation on target recognition methods. However, the training data used in this method is based on ideal conditions and does not consider complex environments, resulting in poor generalization performance. This method is a typical black-box model, with training data typically based on ideal environment assumptions. When faced with complex background noise or unseen occlusion situations, the network's generalization performance is poor, and the physical basis of its recovery is difficult to explain, making it prone to failure under strong noise.

[0021] Based on the above analysis, the existing technology mainly has the following shortcomings: (1) Poor noise adaptability: The lack of a dynamic sensing mechanism for the signal-to-noise ratio of the input signal results in a fixed soft threshold parameter, making it impossible to balance denoising and signal preservation under different noise levels.

[0022] (2) Weak detail recovery capability: Traditional reconstruction algorithms or simple networks are unable to realistically recover the high-frequency details and weak scattering points of the deformed signal, resulting in differences in texture between the recovered signal and the real signal.

[0023] (3) Lack of physical constraints: Pure neural network methods ignore the block sparse prior knowledge of the signal, resulting in a large number of model parameters and poor interpretability.

[0024] Therefore, this invention proposes a deformable target recovery method based on deep unfolding reconstruction and generative adversarial post-processing to solve the above-mentioned technical problems.

[0025] Figure 1 This is a flowchart of a deformable target recovery method based on depth unfolding reconstruction and generative adversarial post-processing provided by an embodiment of the present invention, such as... Figure 1 As shown, the method may include the following steps: S1. Collect the HRRP data of the target to be recovered and preprocess it to obtain the preprocessed HRRP data to be recovered.

[0026] Understandably, preprocessing operations on the HRRP data to be recovered for the target recovery can include normalization and symmetric operations.

[0027] S2. Perform sparse recovery on the preprocessed HRRP data to be recovered using the Bayesian inference algorithm to obtain the sparse representation coefficients of the target to be recovered.

[0028] S3. Obtain the deformation component to be restored based on the sparse representation coefficients of the target to be restored.

[0029] It is understandable that the process of obtaining the deformation component to be recovered can be represented as follows: ; in, This represents the preprocessed HRRP data to be recovered. For the preset dictionary matrix, The sparse representation coefficients of the target to be restored This represents the deformation component to be recovered.

[0030] S4. Input the deformation component to be recovered into the trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation component of the current round; wherein, the model based on deep unfolding reconstruction and generative adversarial post-processing includes a deep unfolding network and a generative adversarial network.

[0031] S5. When the preset maximum iteration round is reached, the recovered HRRP data is obtained based on the target deformation component of the current round and the sparse representation coefficient of the target to be recovered; otherwise, return to S2 and iterate again until the preset maximum iteration round is reached.

[0032] Optionally, the process of obtaining the training deformation components includes: Collect training HRRP data and preprocess it to obtain preprocessed training HRRP data; The preprocessed training HRRP data is superimposed with block sparse signals of different sparsity to obtain the initial training deformation components. Gaussian white noise is superimposed on the initial training deformation components to obtain the noisy deformation components; The training deformation component is obtained by multiplying the noise-added deformation component with the identity matrix.

[0033] It is understandable that, by utilizing the structured characteristics of deformation components, the preprocessed training HRRP data is superimposed with block sparse signals of different sparsities to obtain the initial training deformation components, and then Gaussian white noise is superimposed to obtain the noisy deformation components under the real simulation environment. The noisy deformation components are multiplied with the identity matrix to obtain the training deformation components, and finally the target category is used as the label.

[0034] Optionally, the training process for a deep unfolded network includes: Obtain the original deformation components; The sparse signal of the training block is divided to obtain the sparse signal of the divided training block. Construct a training dictionary matrix based on the structure of the sparse signals of the divided training blocks; The model of the sparse signal of the divided training block is obtained based on the sparse signal of the divided training block and the training dictionary matrix; The model of the sparse signal of the divided training block is solved by gradient descent, and the recursive formula of the sparse signal of the divided training block is obtained. The sparse signal of the training block after iteration is obtained according to the recursive formula; The sparse signal of the training block after iteration is filtered using an adaptive soft thresholding method to obtain the filtered sparse signal of the training block; wherein, the threshold of the adaptive soft thresholding method is obtained based on the noise estimation factor. The loss function is constructed based on the filtered sparse signal of the training block and the original deformation components, and the deep unfolded network is trained to obtain the trained deep unfolded network.

[0035] Understandable, Figure 2 This is a schematic diagram of a deep unfolded network framework provided in an embodiment of the present invention, for obtaining sparse signals from training blocks. And divide it into A series of consecutive blocks. Each block contains The element, denoted as the nth element. block is Its training block sparse signal element The first sub-block Each element is represented as Then the entire training block has sparse signal (in The structure of ) can be represented as: ; in, This indicates the transpose operation. Represents the real number field.

[0036] To match the block structure of the signal, a dictionary matrix is ​​trained. (in Indicates the training deformation components. The dimension representing the sparse signal of the training block also needs to be partitioned accordingly. Let the training dictionary matrix... Depend on Composed of sub-matrix blocks: .

[0037] Sparse signal of the divided training blocks To solve this, but in order to limit... The structured sparsity property allows the model of the sparse signal after the training block to be divided into the following representations: ; in, Indicates regularization sparsity. Norm is defined as each block The sum of norms, that is: ; Then, gradient descent is used to solve the model, resulting in a model of the sparse signal of the divided training blocks: ; in, This represents the learnable weight matrix. Indicates the first Block sparse signal in the next iteration The A small block, Indicates the first The block sparse signal obtained from the next iteration calculation The A small block, Then it means the first The update step size factor of the layer.

[0038] To ensure that the sparse signal of the training block obtained after iteration has structured sparsity, an adaptive soft thresholding method is used for filtering. The specific formula is as follows: ; in, This represents the threshold, which is related to the noise estimation factor. Related, The positive operator is represented as follows: .

[0039] Repeat the above steps sequentially to construct a loss function based on the filtered sparse signal of the training blocks and the original deformation components, and then train the deep unfolded network to obtain the trained deep unfolded network. This loss function can be expressed as follows: .

[0040] in, This represents the original deformation component.

[0041] Optionally, the process of obtaining the noise estimation factor includes: Obtain the training deformation components; The residual signal corresponding to the sparse signal of the divided training block is calculated based on the training deformation components and the recursive formula. The residual signal is fused with the training deformation component to obtain the fused signal; The fused signal is passed through three fully connected layers and two ReLU activation functions of a deep unfolded network, then activated by Sigmoid, and finally truncated numerically to obtain the noise estimation factor.

[0042] Understandable, Figure 3 This is a schematic diagram of the framework of a noise estimation module provided in an embodiment of the present invention, which utilizes training deformation components. and Calculate the residual signal corresponding to the sparse signal of the partitioned training block. : ; Due to residual signal It contains important noise information, so the training deformation component will be used. With residual signal The fusion is performed to obtain the fused signal. ; for fused signals The noise estimation factor is obtained by using three fully connected layers and two ReLU activation functions, followed by Sigmoid activation and numerical truncation. .

[0043] Optionally, the training process of the discriminator in a generative adversarial network includes: The sparse signal of the initial recovery block output from the trained deep unfolded network is input into the generator of the generative adversarial network to obtain generated samples. The generated sample and the original deformation component are input into the discriminator of the generative adversarial network to obtain the first discriminant value of the generated sample and the second discriminant value of the original deformation component, respectively. The first adversarial loss is constructed based on the first discriminant value and the second discriminant value; The discriminator of the generative adversarial network is trained based on the first adversarial loss to obtain the trained discriminator of the generative adversarial network.

[0044] Specifically, Figure 4 This is a schematic diagram of a generative adversarial network framework provided by an embodiment of the present invention, which converts the original deformation components... The sparse signal of the initial recovered block output from the trained deep unwrap network is input into the generator of the generative adversarial network to obtain generated samples. The original deformation components and generated samples are first passed through a convolutional layer with spectral normalization to perceive the local morphology of non-zero locations of real samples and the energy changes of samples within a small range, while ensuring the stability of subsequent GAN training. The output of the spectral normalized convolutional layer is then input into the LeakyReLU activation function. The LeakyReLU activation function effectively alleviates the neuron inactivation problem commonly encountered in sparse signal discrimination by preserving non-zero gradients in negative activation regions, while also enhancing the discriminator's ability to detect weak non-zero structures and pseudo-sparseness. The sensitivity of the pattern is considered. To suppress discriminator overfitting and control the upper limit of discriminator capability, the output of the LeakyReLU activation function is further constrained by a random deactivation layer. Finally, through a series of downsampling steps, the feature dimension of the discriminator's convolutional network output is automatically calculated, and a fully connected layer is constructed accordingly to map the high-dimensional discriminative features into a single scalar output for discriminating against the input real samples. Using the original deformation components and the discriminant values ​​corresponding to the generated samples obtained above, the first adversarial loss between the original deformation components and the generated samples is calculated, and its corresponding first adversarial loss is expressed as follows: ; in, This represents the discriminant value corresponding to the generated sample. This represents the discriminant value corresponding to the original deformation component. This represents the expectation. The generator of the generative adversarial network is forward-propagated using this first adversarial loss.

[0045] Optionally, the training process of the generator in a generative adversarial network includes: The generated samples are input into the generator of the generative adversarial network to obtain the corrected generated samples; The mean square error loss is recovered based on the corrected generated sample and the original deformation components. The initial recovery constraint loss is constructed based on the sparse signal of the preliminary recovery block and the original deformation components; Calculate the second adversarial loss obtained by inputting the corrected generated sample into the discriminator of the trained generative adversarial network; Calculate the sparsity regularization loss of the corrected generated samples; The total loss function is obtained by considering the recovery mean square error loss, the initial recovery constraint loss, the second adversarial loss, and the sparsity regularization loss. The generator of the generative adversarial network is trained based on the total loss function to obtain the trained generator of the generative adversarial network.

[0046] Specifically, the generated samples are mapped to a high-dimensional feature space to effectively capture local block structure and edge information while maintaining the signal length and structure, providing a stable and structured input representation for error modeling in the subsequent residual module. The extracted local block structure and edge information are then processed by the residual module to learn and accumulate modeling of residual systematic errors and nonlinear distortions in the recovered signal layer by layer without destroying the sparse structure already recovered by the Learned Iterative Shrinkage-Thresholding Algorithm (LISTA). Then, the high-dimensional error information extracted by the multi-layer residual module is mapped to a pointwise error estimate of the same dimension as the original signal through the error mapping tail layer. Finally, the error estimate is subtracted from the recovered signal to obtain the corrected generated sample. The mean square error loss of the recovery between the corrected generated sample and the original deformation component is calculated. It is used to constrain the numerical consistency between the generated signal and the real clean signal; it calculates the initial recovery constraint loss between the sparse signal of the preliminary recovery block and the original deformation components. This is used to maintain a stable relationship between the Ada-BlockLISTA network output and the real signal; the second adversarial loss of the corrected generated samples after passing through the discriminator is calculated. It is used to guide the generator to produce high-quality signals that can deceive the discriminator; the sparsity regularization loss is calculated on the corrected generated samples. This is used to suppress non-zero redundant components and improve the sparse representation capability of the signal. The total loss function is obtained by linearly fusing the recovery mean square error loss, initial recovery constraint loss, second adversarial loss, and sparsity regularization loss. This total loss function is expressed as follows: : ; in, To recover the weighting coefficients of the mean squared error loss, These are the weighting coefficients for the initial recovery constraint loss. The weighting coefficients for the second adversarial loss. These are the weight coefficients for the sparsity regularization loss. The generator parameters of the generative adversarial network are iteratively optimized using this total loss function.

[0047] Optionally, the construction process of the trained model based on deep unfolding reconstruction and generative adversarial post-processing includes: The trained Generative Adversarial Network (GAN) is obtained by combining the discriminator and the generator of the trained GAN. Based on the trained deep unfolding network and the trained generative adversarial network, a trained model based on deep unfolding reconstruction and generative adversarial post-processing is constructed.

[0048] It is worth mentioning the alternative to deep unfolding networks: This invention is based on deep unfolding networks. In fact, deep unfolding can also be based on AMP (Approximate Message Passing Algorithm) or ADMM (Alternating Direction Multiplier Method), and similar adaptive recovery effects can be achieved by introducing a noise estimation module to adjust the parameters.

[0049] Post-processing alternatives to generative networks: Although this invention uses generative adversarial networks (GANs) for post-processing, variational autoencoders (VAEs) or the latest diffusion models can be used to replace the GAN modules for fine-tuning the distribution of the initially recovered signal and denoising.

[0050] Input feature substitution for the noise estimation module: Currently, noise estimation utilizes "signal + residual". Alternatively, the frequency domain features or wavelet transform features of the input signal can be used as input to the noise estimation sub-network, achieving the same goal of estimating noise levels and guiding threshold generation.

[0051] In one embodiment, the recovery effect of the method of the present invention is detected using one-dimensional high-resolution range echoes of three types of aircraft. The specific parameters of the radar that recorded the high-resolution echoes of the three types of aircraft targets, and the parameters of the three types of aircraft targets themselves, are shown in Table 1. Table 1

[0052] The An-26's flight path was divided into 6 segments, the Citation's flight path into 7 segments, and the Yak-42's flight path into 5 segments. Only the fifth segment of the Yak-42's data had 17,950 samples; the remaining segments each had 26,000 samples. Regarding the selection of training samples, to ensure the completeness of the template library, the training samples should cover as many attitude angles as possible. Therefore, the fifth and sixth segments of the An-26's flight path, the sixth and seventh segments of the Citation's flight path, and the second and fifth segments of the Yak-42's flight path were selected as training samples, while the data in the remaining segments were used as test samples.

[0053] The experiments related to this invention were conducted on the NVIDIA GeForce RTX 4060 graphics processor platform manufactured by NVIDIA Corporation. The experimental program was written in the open-source Python language, using Python version 3.9, and implemented based on the PyTorch deep learning framework. The specific experimental environment configuration is shown in Table 2.

[0054] Table 2

[0055] In this invention, during the block sparse recovery process, the 128-bit block sparse signal is divided into 16 consecutive blocks, each consisting of 8 consecutive elements. The sparsity is set to 10%, 30%, and 50%, respectively, generating a total of 100 batches of data, with each batch containing 32 samples. 80% of these samples are used as training data, and 20% as validation data. Regarding the Ada-BlockLISTA network, this invention sets it to 5 layers, with a learnable step size. The initial value was set to 0.1, and the Adam optimizer was used during training with a learning rate of 1e-3 and a training duration of 100 epochs. Furthermore, the GAN adversarial network was also trained for 100 epochs in this invention, during which the following parameters were used: Set to 100, The weight is set to 10. Set to 2, Set the learning rate to 5, and set both the generator and discriminator learning rates to 2e-4.

[0056] This invention is used to recover high-resolution echoes from deformed aircraft: Extract a high-resolution echo from the high-resolution echo data of the An-26 aircraft in its undeformed state. Figure 5 This is a schematic diagram of raw aircraft HRRP data provided in an embodiment of the present invention, as shown below. Figure 5 As shown, deformation components are superimposed in the distance element 70-90. Figure 6 This is a schematic diagram of a deformation component provided in an embodiment of the present invention. The specific deformation component is as follows: Figure 6 Furthermore, noise is superimposed on it to make the signal-to-noise ratio of the superimposed HRRP 10dB; Figure 7 This is a schematic diagram of HRRP data after superimposing deformation components and noise, provided by an embodiment of the present invention. Figure 7 As shown, the results have changed significantly, especially between 70 and 90 units. Then, the method of this invention is used to restore them. The specific restoration process is as shown in this invention. Figure 8 This is a schematic diagram illustrating the recovery effect of the present invention as provided in an embodiment of the present invention, such as... Figure 8 As shown, Figure 9 This is a schematic diagram of HRRP data recovery provided by an embodiment of the present invention, as shown below. Figure 9 As shown, the method of this invention can effectively recover the high-resolution echo of a deformed aircraft into the corresponding high-resolution echo of an undeformed aircraft even in low signal-to-noise ratio environments. Furthermore, a comparative analysis with the "variant target reconstruction and recognition method based on long short-term memory networks" demonstrates that… Figure 10 This is a schematic diagram of HRRP data reconstructed by an existing method according to an embodiment of the present invention, as shown in the figure. Figure 10 As shown, Figure 11 This is a schematic diagram of a deformation component reconstructed by a conventional method according to an embodiment of the present invention, as shown below. Figure 11 As shown in Table 3, the existing method 1 has a large recovery error and contains significant noise in a low signal-to-noise ratio environment. To further verify the recovery performance of the method of the present invention, NMSE will be used as a metric to detect the recovery error.

[0057] Table 3

[0058] The core of this invention lies in the innovative embedding of a noise estimation module that calculates noise estimation factors within a deep network. This mechanism allows the network parameters to change dynamically with the signal-to-noise ratio of the input data, rather than being fixed. Unlike traditional GANs that directly generate target images, this invention proposes a generator that learns the distribution of "residual noise" and uses subtraction to correct the initial signal recovery, thereby improving the detail recovery accuracy of one-dimensional radar signals. The key protection lies in the cascaded framework that combines a physically interpretable deep unfolded network as the pre-stage "coarse recovery" stage with a GAN with strong distribution fitting capabilities as the post-stage "fine recovery" stage, as well as the alternating iterative process that incorporates Bayesian inference during the testing phase.

[0059] Figure 12 This invention provides a deformable target recovery device based on depth unfolding reconstruction and generative adversarial post-processing, such as... Figure 12 As shown, the device 1200 may include: Preprocessing module 1201 is used to collect the HRRP data to be recovered from the target to be recovered and to preprocess it to obtain the preprocessed HRRP data to be recovered; The sparse recovery module 1202 is used to perform sparse recovery on the preprocessed HRRP data to be recovered according to the Bayesian inference algorithm to obtain the sparse representation coefficients of the target to be recovered. The module 1203 for obtaining the deformation components to be recovered is used to obtain the deformation components to be recovered based on the sparse representation coefficients of the target to be recovered. The deformation component processing module 1204 is used to input the deformation component to be recovered into the trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation component of the current round; wherein, the model based on deep unfolding reconstruction and generative adversarial post-processing includes a deep unfolding network and a generative adversarial network. The data recovery module 1205 is used to recover HRRP data based on the target deformation component and the sparse representation coefficient of the target to be recovered in the current round when the preset maximum iteration round is reached; otherwise, it returns to the sparse recovery module and iterates again until the preset maximum iteration round is reached.

[0060] It is understood that the device embodiments are basically similar to the method embodiments, so the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0061] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0062] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0063] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0064] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for deformed target recovery based on deep unfolding reconstruction and generative adversarial post-processing, characterized in that, The method includes: S1. Collect the HRRP data of the target to be recovered and preprocess it to obtain the preprocessed HRRP data to be recovered; S2. Perform sparse recovery on the preprocessed HRRP data to be recovered according to the Bayesian inference algorithm to obtain the sparse representation coefficients of the target to be recovered; S3. Obtain the deformation component to be recovered based on the sparse representation coefficients of the target to be recovered; S4. Input the deformation component to be recovered into the trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation component of the current round; wherein, the model based on deep unfolding reconstruction and generative adversarial post-processing includes a deep unfolding network and a generative adversarial network; S5. When the preset maximum iteration round is reached, the recovered HRRP data is obtained based on the target deformation component of the current round and the sparse representation coefficient of the target to be recovered; otherwise, return to S2 and iterate again until the preset maximum iteration round is reached.

2. The deformed target recovery method based on depth unfolding reconstruction and generative adversarial post-processing according to claim 1, characterized in that, The training process of the deep unfolded network includes: Obtain the original deformation components; The sparse signal of the training block is divided to obtain the sparse signal of the divided training block. Construct a training dictionary matrix based on the structure of the sparse signals of the divided training blocks; The model of the sparse signal of the divided training block is obtained based on the sparse signal of the divided training block and the training dictionary matrix; The model of the sparse signal of the divided training block is solved by gradient descent, and the recursive formula of the sparse signal of the divided training block is obtained. The sparse signal of the training block after iteration is obtained according to the recursive formula; The sparse signal of the iterative training block is filtered using an adaptive soft thresholding method to obtain the filtered sparse signal of the training block; wherein the threshold of the adaptive soft thresholding method is obtained based on a noise estimation factor. A loss function is constructed based on the filtered sparse signal of the training block and the original deformation components, and the deep unfolded network is trained to obtain the trained deep unfolded network.

3. The deformed target recovery method based on depth unfolding reconstruction and generative adversarial post-processing according to claim 2, characterized in that, The process of obtaining the noise estimation factor includes: Obtain the training deformation components; Calculate the residual signal corresponding to the sparse signal of the partitioned training block based on the training deformation component and the recursive formula; The residual signal is fused with the training deformation component to obtain the fused signal; The fused signal is passed through three fully connected layers and two ReLU activation functions of the deep unfolded network, then activated by Sigmoid, and finally truncated numerically to obtain the noise estimation factor.

4. The deformed target recovery method based on depth unfolding reconstruction and generative adversarial post-processing according to claim 3, characterized in that, The process of obtaining the training deformation components includes: Collect training HRRP data and preprocess it to obtain preprocessed training HRRP data; The preprocessed training HRRP data is superimposed with block sparse signals of different sparsity to obtain the initial training deformation components. Gaussian white noise is superimposed on the initial training deformation components to obtain the noisy deformation components. The trained deformation component is obtained by multiplying the noise-added deformation component with the identity matrix.

5. The deformed target recovery method based on depth unfolding reconstruction and generative adversarial post-processing according to claim 2, characterized in that, The training process of the discriminator in the generative adversarial network includes: The sparse signal of the initial recovery block output by the trained deep unfolded network is input into the generator of the generative adversarial network to obtain generated samples; The generated sample and the original deformation component are input into the discriminator of the generative adversarial network to obtain the first discriminant value of the generated sample and the second discriminant value of the original deformation component, respectively. A first adversarial loss is constructed based on the first discriminant value and the second discriminant value; The discriminator of the generative adversarial network is trained based on the first adversarial loss to obtain the trained discriminator of the generative adversarial network.

6. The deformed target recovery method based on depth unfolding reconstruction and generative adversarial post-processing according to claim 5, characterized in that, The training process of the generator in the generative adversarial network includes: The generated sample is input into the generator of the generative adversarial network to obtain the corrected generated sample; The mean square error loss is recovered based on the corrected generated sample and the original deformation component. The initial recovery constraint loss is constructed based on the sparse signal of the preliminary recovery block and the original deformation components; Calculate the second adversarial loss obtained by inputting the corrected generated sample into the discriminator of the trained generative adversarial network; Calculate the sparsity regularization loss of the corrected generated samples; The total loss function is obtained based on the recovery mean square error loss, the initial recovery constraint loss, the second adversarial loss, and the sparsity regularization loss. The generator of the generative adversarial network is trained according to the total loss function to obtain the trained generator of the generative adversarial network.

7. The deformed target recovery method based on depth unfolding reconstruction and generative adversarial post-processing according to claim 6, characterized in that, The process of constructing the trained model based on deep unfolding reconstruction and generative adversarial post-processing includes: The trained generative adversarial network is obtained based on the discriminator and the generator of the trained generative adversarial network. Based on the trained deep unfolding network and the trained generative adversarial network, a trained model based on deep unfolding reconstruction and generative adversarial post-processing is constructed.

8. A deformable target recovery device based on deep unfolding reconstruction and generative adversarial post-processing, characterized in that, The device includes: The preprocessing module is used to collect the HRRP data to be recovered from the target to be recovered and to preprocess it to obtain the preprocessed HRRP data to be recovered. The sparse recovery module is used to perform sparse recovery on the preprocessed HRRP data to be recovered according to the Bayesian inference algorithm to obtain the sparse representation coefficients of the target to be recovered. The module for obtaining the deformation components to be recovered is used to obtain the deformation components to be recovered based on the sparse representation coefficients of the target to be recovered. The module for processing the deformation components to be recovered is used to input the deformation components to be recovered into a trained model based on deep unfolding reconstruction and generative adversarial post-processing to obtain the target deformation components of the current round; wherein, the model based on deep unfolding reconstruction and generative adversarial post-processing includes a deep unfolding network and a generative adversarial network; The data recovery module is used to recover HRRP data based on the target deformation component of the current cycle and the sparse representation coefficient of the target to be recovered when the preset maximum iteration cycle is reached; otherwise, it returns to the sparse recovery module and iterates again until the preset maximum iteration cycle is reached.

Images