Attribute scattering center parameter inversion and radar echo reconstruction method based on neural network

Abstract

The invention provides an attribute scattering center parameter inversion and radar echo reconstruction method based on a neural network. The method comprises the following steps: S1, constructing a simulation parameter data set; s2, taking the attribute scattering center parameters as input, fitting a mapping relation from the scattering center parameters to radar echoes through Fourier position coding calculation, FiLM physical conditional calculation and decoder up-sampling, and outputting a radar echo predicted value; s3, taking the radar echo as input, performing downsampling dimension reduction through an encoder, outputting an attribute scattering center parameter prediction value through a parameter predictor, sending the attribute scattering center parameter prediction value to the neural network decoder pre-trained in the step S2 for radar echo reconstruction, outputting the radar echo prediction value, and optimizing network parameters until the encoder and the predictor converge; and S4, inputting the test set into an encoder, predictor and decoder network, and carrying out attribute scattering center parameter inversion and radar echo reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of attribute scattering center extraction, specifically involving a method for attribute scattering center parameter inversion and radar echo reconstruction based on neural networks. Background Technology

[0002] Radars primarily operate in high-frequency regions. Both theoretical derivations and practical applications confirm that for radars operating in this frequency band, the overall scattering response of a target can be characterized by the coherent superposition of responses from multiple local equivalent scattering sources. These local scattering sources are defined as the target's equivalent scattering center, often simply referred to as the scattering center. The value of the scattering center lies in several aspects: it provides crucial information such as amplitude, location, and structure required for target classification and identification, and it describes the target's total radar echo in a concise and complete manner, significantly reducing the storage overhead of the target scattering characteristic database. The concept of the scattering center provides a clear and convenient technical path for characterizing the electromagnetic scattering characteristics of complex radar targets, forming the basis for applications such as automatic radar target identification, synthetic aperture radar imaging, and target feature signal compression and reconstruction.

[0003] Currently, publicly published classic scattering center models include the ideal point scattering center model, the attenuation exponential model, the scattering center model based on geometric diffraction theory, and the attribute scattering center model. Among them, the attribute scattering center model, based on the ideal point scattering center model, introduces a frequency-dependent factor to characterize different scattering structures, and introduces exponential and sinc functions to describe the dependence of the scattering center amplitude on the radar incident wave azimuth angle. It includes rich physical and geometric properties of the scattering center and has the best fitting effect on radar echoes. Methods for extracting scattering centers from radar echoes can be divided into three types: image domain extraction methods, parametric model fitting methods, and sparse representation and reconstruction methods. Image domain extraction methods require radar imaging of the radar echoes, exemplified by the CLEAN algorithm. This algorithm is an iterative deconvolution process, the core of which is to decompose the target by repeatedly identifying and subtracting the response of the strongest scattering point in the complex SAR image domain. Methods based on parameterized model fitting treat the extraction problem as a nonlinear parameter estimation problem, directly fitting the scattering center model through optimization techniques, and estimating the distance and azimuth of the scattering center through rotation-invariant techniques (such as ESPRIT) or search methods (such as MUSIC). Methods based on sparse representation and reconstruction utilize the natural sparsity of scattering centers in the imaging space, treating the observation data as a linear measurement process, with its measurement matrix consisting of an overcomplete dictionary of scattering centers. The extraction problem is then transformed into the problem of recovering the sparse signal (i.e., the scattering center parameters).

[0004] Traditional backscattering center extraction algorithms suffer from high computational complexity. Parameterized methods based on global search and sparse reconstruction methods using large, overcomplete dictionaries are computationally burdensome when dealing with high-resolution, large-scene data, making them unsuitable for real-time processing. Furthermore, they require pre-defined parameters and lack adaptive capabilities. Deep neural networks, however, can learn complex functional mapping relationships, transforming backscattering center extraction from a mathematical inversion problem into a data-driven feature learning problem, making radar echo reconstruction and backscattering center parameter extraction possible. Currently, research applying deep neural networks to backscattering center extraction is limited. The complex functional mapping relationship between attributed backscattering center parameters and radar echoes, along with the discreteness of frequency-dependent factors in the backscattering center parameters, poses challenges to gradient descent in neural networks, rendering simple deep neural networks ineffective in modeling. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a neural network-based method for inverting attribute scattering center parameters and reconstructing radar echoes. This method designs an encoder based on radar echo composite reconstruction loss, a parameter predictor combining Gumbel-Softmax differentiable classification, and a decoder based on FiLM conditionalization and Fourier position coding for physical conditionalization. An end-to-end framework for attribute scattering center radar echo reconstruction and parameter inversion is constructed, achieving excellent performance in the task of inverting attribute scattering center parameters.

[0006] Technical Solution: To achieve the aforementioned objectives, this invention proposes a method for inverting attribute scattering center parameters and reconstructing radar echoes based on neural networks, comprising the following steps:

[0007] S1. Constructing the simulation parameter dataset ,in, The attribute scattering center parameters contain five dimensions: , The distance to the attribute scattering center is the distance in the direction of the scattering. The azimuth distance of the attribute scattering center. The magnitude of the attribute scattering center. It is a frequency-dependent factor. The length of the attribute scattering center. It's a radar echo. This represents the entire dataset with 8000 samples, of which 7000 samples are divided into the training set. The 1000 samples were divided into a test set. ;

[0008] S2, a pre-trained neural network decoder, performs radar echo reconstruction. Attribute scattering center parameters As input, data preprocessing is first performed through Fourier position coding calculation, then FiLM physical conditionalization calculation generates factors to modulate the decoder network, and finally, the decoder upsamples the data to fit the parameters from the scattering center. The mapping relationship between radar echo and radar echo is used to output the radar echo prediction value. The network parameters are optimized by calculating the radar echo loss function until the decoder converges. After convergence, the decoder parameters are frozen for use in the training of the encoder and predictor in step S3.

[0009] S3. Train the neural network encoder and predictor, and obtain the attribute scattering center parameters. Inversion, will radar echoes As input, the encoder performs downsampling and dimensionality reduction, and the parameter predictor outputs the predicted values ​​of the attribute scattering center parameters. , The radar echo is fed into the pre-trained neural network decoder in step S2 for radar echo reconstruction, and the predicted radar echo value is output. By calculating the radar echo composite reconstruction loss function, the network parameters are optimized until the encoder and predictor converge.

[0010] S4. Save the optimal parameters from the training process in step S3, and then save the test set. Input the encoder, predictor, and decoder network to perform attribute scattering center parameter inversion and radar echo reconstruction.

[0011] Furthermore, the attribute scattering center and the radar echo have the following correspondence:

[0012]

[0013] in, For radar echo, Let f be the attribute scattering center parameter and f' be the property scattering center parameter. The sampling points represent the frequency and azimuth angle, respectively. denoted as the center frequency, and c as the speed of light.

[0014] Furthermore, in step S1, the method for constructing the simulation parameter dataset is as follows:

[0015] The simulated radar parameters are set as follows: radar center frequency is 9GHz, bandwidth is 1.2GHz, azimuth range is ±3.8°, number of sampling points in both frequency domain and azimuth domain is 128, and imaging space is 16m x 16m.

[0016] Based on the simulated radar parameter settings, the attribute scattering center is generated. and radar echo Among them, distance to position follow Uniform distribution This is the lower limit of the distance to the position. The upper limit of the position in the distance direction, and the position in the azimuth direction. follow Uniform distribution This represents the lower limit of the azimuth position. The upper limit of the azimuth position, amplitude follow Uniform distribution This is the lower limit of the amplitude. Upper limit of amplitude, frequency dependent factor follow Discrete distribution, length parameter exist Based on a uniform distribution, negative values ​​are set to 0 to simulate both uniformly distributed local and distributed scattering centers. This is the lower limit of the length. This represents the upper limit of length.

[0017] Furthermore, the method for calculating the Fourier position code in step S2 is as follows:

[0018] Parameters of the property scattering center The following calculations were performed to obtain the results. , , , , :

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] Where B is the bandwidth, , , , , Combination As the input to the neural network decoder, the input dimension is... :

[0025] .

[0026] Furthermore, the method for calculating the FiLM physical conditions in step S2 is as follows:

[0027] Attribute scattering center parameter Input the FiLM Generator module to generate scaling and translation factors that modulate the decoder network. , , The FiLMGenerator module represents the number of network layers in the decoder and consists of fully connected layers. The dimensions are progressively increased, with each layer followed by a non-linear activation function, ultimately outputting scaling and translation factors. ;

[0028] By scaling and translation factors The process of modulating the decoder network is represented as:

[0029]

[0030] in, For decoder number Parameters before layer modulation, For decoder number Parameters after layer modulation This is a dot product operation.

[0031] Furthermore, the method for optimizing the neural network parameters in step S2 is as follows:

[0032] Will Attribute scattering center parameters Input to neural network, output radar echo prediction value By calculating the radar echo loss function, the network parameters are optimized until the decoder converges. Losses from reconstruction Smoothing constraint regularization term constitute:

[0033]

[0034] in, These are the weighting coefficients. represent and Mean square error between:

[0035]

[0036] Where M and N represent the number of sampling points of the radar echo in the range and azimuth directions, respectively. This represents the i-th scattered field value in the actual radar echo. This represents the i-th scattered field value in the predicted radar echo. Scaling and translation factors during modulation composition;

[0037]

[0038] The optimizer is set to Adam, the learning rate is 1e-3, the batch size is 64, and the training iteration cycle is 1200 rounds. The neural network parameters are updated through code propagation and backpropagation optimization algorithms until the network converges.

[0039] Furthermore, the calculation method for the parameter predictor in step S3 is as follows:

[0040] Attribute scattering center parameters middle, For continuous values, the parameter predictor calculates the predicted values ​​directly using a fully connected neural network and activation functions. , For discrete values, follow The discrete distribution, the predictor does not directly output The predicted values ​​are calculated using Gumbel-Softmax differentiable classification, and the output is... The class probabilities are selected, and the value with the highest class probability is chosen as the class probability. final predicted value The Gumbel-Softmax differentiable class predictor calculates class probabilities using the following formula:

[0041]

[0042] in, It is the original class probability, satisfying , 5 represents Five possible values, They are independent and identically distributed. Noise is generated in the following ways:

[0043]

[0044] in, It's a temperature parameter that controls the smoothness of the approximation. An annealing strategy is used during model training, and a preset temperature is used at the start of training. The intensity gradually decreased as training progressed. .

[0045] Furthermore, the method for optimizing the neural network parameters in step S3 is as follows:

[0046] Will radar echoes As input, the encoder, predictor, and decoder output radar echo prediction values. The network parameters are optimized by calculating the loss function of radar echo composite reconstruction. The parameters of the decoder network are pre-trained and do not need to be updated.

[0047] Composite reconstruction loss function Losses from reconstruction and KL divergence loss constitute:

[0048]

[0049] in, These are the weighting coefficients. Radar echo representing the model output Compared to real radar echoes The differences between them are addressed by employing a multi-component composite loss function based on the complex characteristics of radar echoes, constraining the reconstruction quality from different perspectives. It consists of four parts: real part loss, imaginary part loss, amplitude loss, and phase loss.

[0050]

[0051] in, and Representing radar echoes The real and imaginary parts, and These represent the amplitude and phase of the radar echo, respectively. , , and With radar echo The following correspondence exists:

[0052]

[0053]

[0054] in, and The weights representing the magnitude loss and phase loss are respectively. The Huber loss is calculated as follows:

[0055]

[0056] in, It is the actual value. It is a predicted value. It is a threshold parameter;

[0057] The overall network training was counted in training cycles, with 1200 training iterations, 128 samples per batch, and a learning rate of 0.001. The Adam optimizer was set, and the network parameters were updated through the backpropagation algorithm. After several training cycles, the loss function decreased, and the network converged.

[0058] Furthermore, the method for step S4 is as follows:

[0059] Save the optimal parameters during training and the test set. Input encoder, predictor, decoder, and get and .

[0060] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0061] (1) The decoder designed in this invention effectively solves the problems of "poor high-frequency fitting" and "conditional dilution" in radar echo reconstruction of traditional MLP. By applying affine modulation to the features channel by channel in each deconvolution upsampling layer through FiLM physical conditionalization, the attribute scattering center parameters are ensured. Conditional information is integrated throughout all layers of the network to prevent information "diluted" during deep propagation. Furthermore, physical conditional modulation is incorporated into key resolution enhancement stages to ensure that each feature amplification step conforms to actual physical laws, thereby improving the accuracy and physical consistency of the final radar echo output. Fourier position coding is introduced to incorporate attribute scattering center parameters... Each dimension of the scalar is expanded into a multi-frequency sine and cosine basis, explicitly improving the network's ability to fit high-frequency structures, thus realizing the transformation from attribute scattering center parameters. Modeling the controllable and precise physical coding process of high-dimensional replication of radar echoes.

[0062] (2) The encoder designed in this invention ensures the quality of radar echo reconstruction and the generalization ability of the model from multiple dimensions. The encoder uses progressive downsampling to extract high-dimensional radar echo features, and constrains the network output distribution to be close to the standard normal distribution through KL divergence loss, avoiding the discretization of the potential space and improving the generalization performance. For the complex characteristics of radar echo, the reconstruction loss adopts Huber loss, and constructs constraints from the real and imaginary parts, channel by channel, amplitude, and phase. Among them, the real and imaginary part loss ensures the accuracy of waveform structure, the amplitude loss ensures the energy distribution characteristics, and the phase loss captures subtle phase changes, taking into account both numerical stability and training efficiency, realizing high-fidelity reconstruction of radar echo, and providing a reliable optimization target for parameter inversion.

[0063] (3) The predictor designed in this invention uses Gumbel-Softmax differentiable classification, which efficiently solves the problem of accurate prediction of discrete and continuous parameters. For continuous parameters such as position (x / y), amplitude (A), and length (L), a differentiable activation function is used for range scaling; for frequency-dependent factors... By using the Gumbel-Softmax technique to approximate a discrete distribution with a continuous distribution, and in conjunction with a training annealing strategy, the gradient non-differentiability problem of traditional softmax+argmax is solved, achieving end-to-end accurate inversion of all-dimensional scattering parameters.

[0064] (4) This invention adopts a "staged training" strategy, which balances the accuracy of physical process fitting and the efficiency of parameter inversion. In the first stage, the decoder is trained separately, focusing on fitting the physical encoding process of "scattering parameters → complex radar echo", resulting in a differentiable surrogate network with high-fidelity reconstruction capability; in the second stage, the encoder parameters are frozen, and only the encoder and predictor are trained to achieve rapid inversion of "complex radar echo → scattering parameters". This strategy avoids the coupling interference between the encoding physical process and the inversion process during training, and reduces the training complexity of the inversion stage by freezing the mature decoder, achieving stable convergence of loss and ensuring the training efficiency and parameter inversion accuracy of the model. Attached Figure Description

[0065] Figure 1 This is a flowchart of the method of the present invention;

[0066] Figure 2 This is a comparison of the loss functions in the FiLM conditionalization and Fourier feature ablation experiments of this invention;

[0067] Figure 3 This is a schematic diagram illustrating the RCS fitting effect of FiLM conditionalization and Fourier feature ablation in this invention.

[0068] Figure 4 This is a graph showing the decreasing trend of the loss function of the attribute scattering center parameter inversion network in this invention. Detailed Implementation

[0069] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described.

[0070] like Figure 1 As shown, this invention proposes a method for inverting attribute scattering center parameters and reconstructing radar echoes based on neural networks. This method includes the following steps:

[0071] S1. Constructing the simulation parameter dataset ,in, The attribute scattering center parameters contain five dimensions: , The distance to the attribute scattering center is the distance in the direction of the scattering. The azimuth distance of the attribute scattering center. The magnitude of the attribute scattering center. It is a frequency-dependent factor. The length of the attribute scattering center. It's a radar echo. This represents the entire dataset with 8000 samples, of which 7000 samples are divided into the training set. The 1000 samples were divided into a test set. ;

[0072] S2, a pre-trained neural network decoder, is used to reconstruct radar echoes. Attribute scattering center parameters As input, the data is first preprocessed using Fourier position encoding, then the factors for modulating the decoder network are generated using FiLM physical conditionalization, and finally, the parameters from the scattering center are fitted by upsampling through the decoder. The mapping relationship between radar echo and radar echo is used to output the radar echo prediction value. The network parameters are optimized by calculating the radar echo loss function until the decoder converges. After convergence, the decoder parameters are frozen for use in the training of the encoder and predictor in step S3.

[0073] S3. Train the neural network encoder and predictor, and obtain the attribute scattering center parameters. Inversion. radar echoes As input, the encoder performs downsampling and dimensionality reduction, and the parameter predictor outputs the predicted values ​​of the attribute scattering center parameters. , The radar echo is fed into the pre-trained neural network decoder in step S2 for radar echo reconstruction, and the predicted radar echo value is output. By calculating the radar echo composite reconstruction loss function, the network parameters are optimized until the encoder and predictor converge.

[0074] S4. Save the optimal parameters from the training process in step S3, and then save the test set. Input the encoder, predictor, and decoder network to perform attribute scattering center parameter inversion and radar echo reconstruction.

[0075] Furthermore, the attribute scattering center and the radar echo have the following correspondence:

[0076]

[0077] in, For radar echo, Let f be the attribute scattering center parameter and f' be the property scattering center parameter. These represent sampling points for frequency and azimuth, respectively.

[0078] Furthermore, in step S1, the method for constructing the simulation parameter dataset is as follows:

[0079] The simulated radar parameters are set as follows: radar center frequency is 9GHz, bandwidth is 1.2GHz, azimuth range is ±3.8°, number of sampling points in both frequency domain and azimuth domain is 128, and imaging space is 16m x 16m.

[0080] Based on the simulated radar parameter settings, the attribute scattering center is generated. and radar echo Among them, distance to position follow Uniform distribution This is the lower limit of the distance to the position. The upper limit of the position in the distance direction, and the position in the azimuth direction. follow Uniform distribution This represents the lower limit of the azimuth position. The upper limit of the azimuth position, amplitude follow Uniform distribution This is the lower limit of the amplitude. Upper limit of amplitude, frequency dependent factor follow Discrete distribution, length parameter exist Based on a uniform distribution, negative values ​​are set to 0 to simulate both uniformly distributed local and distributed scattering centers. This is the lower limit of the length. This represents the upper limit of length.

[0081] Furthermore, the method for calculating the Fourier position code in step S2 is as follows:

[0082] Parameters of the property scattering center The following calculations were performed to obtain the results. , , , , :

[0083]

[0084]

[0085]

[0086]

[0087]

[0088] Where B is the bandwidth, , , , , Combination As the input to the neural network decoder, the input dimension is... :

[0089] .

[0090] Furthermore, the method for calculating the FiLM physical conditions in step S2 is as follows:

[0091] Attribute scattering center parameter Input the FiLM Generator module to generate scaling and translation factors that modulate the decoder network. , , The FiLMGenerator module represents the number of network layers in the decoder and consists of fully connected layers. The dimensions are progressively increased, with each layer followed by a non-linear activation function, ultimately outputting scaling and translation factors. ;

[0092] By scaling and translation factors The process of modulating the decoder network is represented as:

[0093]

[0094] in, For decoder number Parameters before layer modulation, For decoder number Parameters after layer modulation This is a dot product operation.

[0095] Furthermore, the method for optimizing the neural network parameters in step S2 is as follows:

[0096] Will Attribute scattering center parameters Input to neural network, output radar echo prediction value By calculating the radar echo loss function, the network parameters are optimized until the decoder converges. Losses from reconstruction Smoothing constraint regularization term constitute:

[0097]

[0098] in, These are the weighting coefficients. represent and Mean square error between:

[0099]

[0100] Where M and N represent the number of sampling points of the radar echo in the range and azimuth directions, respectively. This represents the i-th scattered field value in the actual radar echo. This represents the i-th scattered field value in the predicted radar echo. Scaling and translation factors during modulation composition;

[0101]

[0102] The optimizer is set to Adam, the learning rate is 1e-3, the batch size is 64, and the training iteration cycle is 1200 rounds. The neural network parameters are updated through code propagation and backpropagation optimization algorithms until the network converges.

[0103] Furthermore, the calculation method for the parameter predictor in step S3 is as follows:

[0104] Attribute scattering center parameters middle, For continuous values, the parameter predictor calculates the predicted values ​​directly using a fully connected neural network and activation functions. , For discrete values, follow The discrete distribution, the predictor does not directly output The predicted values ​​are calculated using Gumbel-Softmax differentiable classification, and the output is... The class probabilities are selected, and the value with the highest class probability is chosen as the class probability. final predicted value The Gumbel-Softmax differentiable class predictor calculates class probabilities using the following formula:

[0105]

[0106] in, It is the original class probability, satisfying , 5 represents Five possible values, They are independent and identically distributed. Noise is generated in the following ways:

[0107]

[0108] It's a temperature parameter that controls the smoothness of the approximation. An annealing strategy is used during model training, and a preset temperature is used at the start of training. The intensity gradually decreased as training progressed. .

[0109] Furthermore, the method for optimizing the neural network parameters in step S3 is as follows:

[0110] Will radar echoes As input, the encoder, predictor, and decoder output radar echo prediction values. The network parameters are optimized by calculating the loss function of radar echo composite reconstruction. The parameters of the decoder network are pre-trained and do not need to be updated.

[0111] Composite reconstruction loss function Losses from reconstruction and KL divergence loss constitute:

[0112]

[0113] in, These are the weighting coefficients. Radar echo representing the model output Compared to real radar echoes The differences between them are addressed by employing a multi-component composite loss function based on the complex characteristics of radar echoes, constraining the reconstruction quality from different perspectives. It consists of four parts: real part loss, imaginary part loss, amplitude loss, and phase loss.

[0114]

[0115] in, and Representing radar echoes The real and imaginary parts, and These represent the amplitude and phase of the radar echo, respectively. , , and With radar echo The following correspondence exists:

[0116]

[0117]

[0118] and The weights representing the magnitude loss and phase loss are respectively. The Huber loss is calculated as follows:

[0119]

[0120] in, It is the actual value. It is a predicted value. It is a threshold parameter;

[0121] The overall network training was counted in training cycles, with 1200 training iterations, 128 samples per batch, and a learning rate of 0.001. The Adam optimizer was set, and the network parameters were updated through the backpropagation algorithm. After several training cycles, the loss function decreased, and the network converged.

[0122] Furthermore, the method for step S4 is as follows:

[0123] Save the optimal parameters during training and the test set. Input encoder, predictor, decoder, and get and The network performance was evaluated on a test set using root mean square error (RMSE), RCS fitting, and SAR images, with a sample size of [number missing]. The formula for calculating RMSE is:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] To reconstruct the amplitude of the radar echo, the RCS calculation process is as follows:

[0130]

[0131] Win is the windowing function, IFFT is the inverse Fourier transform, and the calculation process for SAR images is as follows:

[0132]

[0133] As shown in Table 1, in order to verify the effectiveness of the FiLM-based physical conditional decoder, this section adopts a comparative experiment of four structural modes. In order to verify the effectiveness of FiLM modulation and Fourier position coding, except for the difference between the FiLM layer and the Fourier position coding layer, the neural network architecture and training parameters of the four structural modes are kept consistent.

[0134]

[0135] Table 1. Characteristics of comparative experimental models of physically conditional decoders based on FiLM.

[0136] Reference Figure 2 Experimental comparisons revealed that the Baseline method exhibited the most gradual loss decrease and ultimately the highest loss value, indicating that the baseline method has a weak fitting ability in Fourier feature ablation scenarios. The FiLM-only method showed a better loss decrease trend than the Baseline method, but remained at a high level in the later stages, suggesting that using only FiLM conditionalization can help with feature processing to some extent, but the improvement is limited. The PE-only method showed a significantly faster loss decrease than the previous two, and its final loss was lower than that of FiLM-only and Baseline, verifying the effectiveness of positional encoding in feature ablation scenarios. The FiLM+PE method showed the largest loss decrease, with the loss in the later stages of iteration being much lower than other methods, indicating that the combination strategy of FiLM conditionalization and positional encoding (PE) performed best in this experiment, reducing loss more efficiently and improving the model's adaptability to Fourier feature ablation scenarios.

[0137] To verify the fitting performance of the FiLM-based physical conditionalization decoder on the test set, the radar cross section (RCS) corresponding to the real radar echo and the RCS corresponding to the radar echo output by the neural network with four structural modes were calculated. The results are referenced from... Figure 3 As can be seen, the calculation results of FiLM+PE best fit the actual RCS curve changes.

[0138] Reference Figure 4 The loss function of the attribute scattering center parameter inversion network gradually decreases and tends to stabilize, indicating that the network is converging. The test set is fed into the neural network for parameter inversion to obtain the predicted attribute scattering center parameter values. These values ​​are then compared with the true values ​​to calculate the root mean square error of each parameter. To verify the effectiveness of this network, a variational autoencoder with the same number of parameters was trained, and its performance on the test set was calculated. The experimental results are shown in Table 2. It can be seen that this method can achieve more accurate estimation of scattering center parameters.

[0139] Scattering center parameter data were randomly selected from the test set and fed into the network for parameter inversion. The inverted parameters and the original parameters were then used for SAR imaging. The ground truth SAR image and the SAR image reconstructed by parameter inversion showed a high degree of fit, and the parameter inversion achieved good results. It can be used for subsequent SAR imaging, target recognition and other processes.

[0140]

[0141] Table 2. Error Table for Inversion of Scattering Center Parameters of Test Set

[0142] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A neural network-based attribute scattering center parameter inversion and radar echo reconstruction method, characterized in that, Comprising the following steps: S1, Constructing simulation parameter dataset where, is the attribute scattering center parameter, containing five dimensions: , is the distance-to-distance of the attribute scattering center, is the azimuth distance of the attribute scattering center, is the amplitude of the attribute scattering center, is the frequency-dependent factor, is the length of the attribute scattering center, is the radar echo, represents the overall dataset, with 8000 samples, 7000 of which are divided into a training set , and 1000 of which are divided into a test set ; S2, pre-training the neural network decoder to perform radar echo reconstruction, to obtain the attribute scattering center parameters in the scattering center parameter set As input, data preprocessing is performed through Fourier position coding calculation, factors for modulating the decoder network are generated through FiLM physical conditioning calculation, and finally the mapping relationship between the scattering center parameters and the radar echo is fitted through decoder upsampling to output the radar echo prediction value The network parameter optimization is guided by calculating the radar echo loss function until the decoder converges, and after convergence, the decoder parameters are frozen for the training of the encoder and the predictor in step S3. S3, training neural network encoder and predictor, performing attribute scattering center parameter inversion, will radar echo in As input, down-sampling dimension reduction through the encoder, output attribute scattering center parameter prediction value through the parameter predictor , into the neural network decoder pre-trained in step S2 for radar echo reconstruction, output radar echo prediction value , through the calculation of radar echo complex reconstruction loss function, guide network parameter optimization, until the encoder and predictor converge; S4, save the optimal parameters in the training process of step S3, and test set Input encoder, predictor, decoder network, attribute scattering center parameter inversion and radar echo reconstruction.

2. The neural network based attribute scattering center parameter inversion and radar echo reconstruction method according to claim 1, characterized in that, The attribute scattering center has the following corresponding relationship with the radar echo: wherein, is a radar return, is a property scattering center parameter, f and represent frequency and azimuthal sampling points, respectively, is a center frequency, is the speed of light.

3. The neural network based property scattering center parameter inversion and radar echo reconstruction method according to claim 1, characterized in that, In step S1, the method for constructing the simulation parameter data set is as follows: The simulation radar parameters are set as follows: the radar center frequency is 9GHz, the bandwidth is 1.2GHz, the azimuth range is ±3.8°, the sampling point number in the frequency domain and the azimuth domain is 128, and the imaging space is 16mx16m; Based on the simulated radar parameter settings, the attribute scattering center is generated. and radar echo Among them, distance to position follow Uniform distribution This is the lower limit of the distance to the position. The upper limit of the position in the distance direction, and the position in the azimuth direction. follow Uniform distribution This represents the lower limit of the azimuth position. The upper limit of the azimuth position, amplitude follow Uniform distribution This is the lower limit of the amplitude. Upper limit of amplitude, frequency dependent factor follow Discrete distribution, length parameter exist Based on a uniform distribution, negative values ​​are set to 0 to simulate both uniformly distributed local and distributed scattering centers. This is the lower limit of the length. This represents the upper limit of length.

4. The neural network based attribute scattering center parameter inversion and radar echo reconstruction method according to claim 3, characterized in that, The method for calculating the Fourier position encoding in step S2 is as follows: Parameters of the scattering centers of the attributes , respectively, as follows , , , , : Wherein, B is bandwidth, will 、 、 、 、 Combination is , as a neural network decoder, the dimension size of the input is : 。 5. The neural network based attribute scattering center parameter inversion and radar echo reconstruction method according to claim 4, characterized in that, The method for calculating the FiLM physical conditioning in step S2 is as follows: Attribute scattering center parameters An input FiLM Generator module generates scaling and shifting factors that modulate the decoder network , , The number of network layers representing the decoder, the FiLM Generator module is composed of fully connected layers Dimensionality is gradually increased, each layer is matched with a nonlinear activation function, and finally the scaling and shifting factors are output ; By scaling and translation factors The process of modulating the decoder network is denoted as: wherein, is the parameter before modulation for the decoder layer, is the parameter after modulation for the decoder layer, is the dot product operation.

6. The neural network based attribute scattering center parameter inversion and radar echo reconstruction method according to claim 5, characterized in that, The neural network parameter optimization method in step S2 is as follows: The attribute scattering center parameters in the formula are determined by the following steps: 1) determining a plurality of initial attribute scattering center parameters; 2) inputting the initial attribute scattering center parameters into a neural network to obtain a radar echo prediction value; 3) calculating a radar echo loss function by using the radar echo prediction value and a radar echo value; 4) guiding optimization of network parameters by using the radar echo loss function until a decoder converges; and 5) determining the attribute scattering center parameters by using the optimized network parameters. ​​​​​​ wherein, is a weight coefficient, represents with the mean square error between where M and N represent the number of sampling points in range and azimuth direction of radar echo respectively, denotes the i-th scattering field value in real radar echo, denotes the i-th scattering field value in predicted radar echo, consists of scaling and translation factors at modulation time ; The optimizer is set as Adam, the learning rate is 1e-3, the batch sample number is 64, the training iteration period is 1200 rounds, the coding propagation and the backward gradient optimization algorithm are passed, the neural network parameter update is realized, and the network converges until the network converges.

7. The neural network based property scattering center parameter inversion and radar echo reconstruction method according to claim 6, characterized in that, The calculation method of the parameter predictor in step S3 is as follows: Attribute scattering center parameter In, For continuous values, the parameter predictor is calculated by a fully connected neural network and an activation function, and directly outputs the predicted value , For discrete values, follow the discrete distribution of The predictor does not directly output the predicted value of Through the Gumbel-Softmax differentiable classification calculation, the category probability is output, and the value with the maximum category probability is selected as the final prediction value of Wherein, the Gumbel-Softmax differentiable classification predictor, the formula for calculating the category probability is as follows:​​ where is the original class probability, satisfying , is 5, representing five possible values, are independent and identically distributed noises, generated by wherein, is a temperature parameter, controlling the smoothness of the approximation, an annealing strategy is taken during the model training, using a pre-set , which is gradually decreased as the training progresses.

8. The neural network based property scattering center parameter inversion and radar echo reconstruction method according to claim 1, characterized in that, The neural network parameter optimization method in step S3 is as follows: Will radar echoes As input, the encoder, predictor, and decoder output radar echo prediction values. The network parameters are optimized by calculating the loss function of radar echo composite reconstruction. The parameters of the decoder network are pre-trained and do not need to be updated. composite reconstruction loss function from the reconstruction loss and the kl divergence loss comprises: wherein, are weight coefficients, represent radar echoes output by the model and the difference between the radar echoes from different angles, according to the complex characteristics of the radar echoes, a multi-component complex loss function is adopted to constrain the reconstruction quality, including four parts: real part loss, imaginary part loss, amplitude loss, and phase loss: wherein and represent the real and imaginary parts of the radar return , and represent the amplitude and phase of the radar return , , and the following correspondences exist between the radar return ​ where, and represent the weight of amplitude loss and phase loss, respectively, and the Huber loss is calculated as follows: wherein, is a true value, is a predicted value, is a threshold parameter; The whole network training is counted by the training period, the training iteration period is 1200 rounds, the batch sample number is 128, the learning rate is 0.001, the Adam optimizer is set, the network parameter is updated through the backward propagation algorithm, the loss function decreases after multiple training periods, and the network converges.

9. The neural network based property scattering center parameter inversion and radar echo reconstruction method according to claim 8, characterized in that, The method of step S4 is as follows: Save the optimal parameters in the training process, test set Input encoder, predictor, decoder, get And .

Images