Deep learning tomographic sar three-dimensional imaging method based on order dimension reduction and model selection

Abstract

The application discloses a high-precision and high-efficiency deep learning tomographic SAR three-dimensional imaging method based on order dimension reduction and model selection, which is used in the field of tomographic SAR three-dimensional imaging, solves the problems of difficult selection of a model selection penalty term, difficult parameter optimization and slow traversal optimization, and improves the high-precision three-dimensional imaging efficiency of the tomographic SAR. The method is a hybrid imaging processing framework driven by signal processing and a deep learning network, and proposes a tomographic SAR three-dimensional imaging scheme based on four parts of order dimension reduction, norm reconstruction, model selection based on a feature selection network and least square reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of tomographic SAR three-dimensional imaging technology, and relates to a high-precision and high-efficiency deep learning tomographic SAR three-dimensional imaging method based on order dimensionality reduction and model selection. Background Technology

[0002] Synthetic Aperture Radar (SAR) 3D imaging represents a significant advancement in traditional 2D SAR imaging technology within the field of refined radar information acquisition and perception. Its fundamental principle involves creating a new one-dimensional synthetic aperture along the height direction, in addition to the traditional range and azimuth axes, allowing the resolution of multiple overlapping targets within a single pixel. This technology effectively addresses imaging geometric distortions such as overlay and top-to-bottom inversion that can occur in traditional 2D SAR images. SAR 3D imaging technology can meet the high-precision imaging requirements of areas with dramatic terrain undulations and densely built-up areas, and is therefore considered a cutting-edge direction in SAR imaging technology.

[0003] In 3D SAR imaging, tomographic SAR faces challenges such as non-uniform baselines, where the height aperture is often non-uniformly distributed, leading to high complexity in coherent processing. Height Veraile resolution is often insufficient for imaging requirements, necessitating the use of nonlinear super-resolution imaging methods, such as sparse reconstruction algorithms. These methods can achieve 1.5-25 times super-resolution capability under high signal-to-noise ratio conditions. In TomoSAR 3D imaging applications, due to the high column coherence of the sensing matrix and the influence of random noise, artifacts may appear in the estimation results, resulting in insufficient accuracy in height estimation.

[0004] Currently, model selection methods are generally used for artifact identification and removal. However, traditional model selection methods that combine likelihood models and penalty terms suffer from problems such as difficulty in selecting the optimal penalty term, difficulty in tuning the penalty term coefficient parameters, and slow traversal optimization. As a result, they are inefficient when performing high-precision 3D tomographic SAR imaging of large-scale scenes.

[0005] In recent years, in-depth research on artificial intelligence has shown that deep learning methods have powerful capabilities for solving nonlinear problems. Through training with a large amount of simulation or experimental data, they can achieve high-precision approximate fitting of nonlinear models and can be used to solve large-scale nonlinear complex problems quickly and efficiently. This provides a new approach to solving the problems existing in model selection methods. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and propose a deep learning-based tomographic SAR 3D imaging method based on order dimensionality reduction and model selection. This method is a hybrid imaging processing framework jointly driven by signal processing and deep learning networks, proposing a tomographic SAR 3D imaging scheme based on four parts: order dimensionality reduction, norm reconstruction, model selection based on feature selection networks, and least squares reconstruction. Specifically, order dimensionality reduction reduces the dimension of the target scattering vector through compressed sensing to obtain lower-dimensional elevation scattering characteristics; norm reconstruction obtains a coarse estimate of the elevation scattering characteristics by calculating the minimum norm of the lower-dimensional elevation scattering characteristics; model selection based on order features utilizes a deep learning network to perform feature recognition, feature fusion, and feature selection on the autocorrelation amplitude and phase feature maps of the coarse estimate of the elevation scattering characteristics to obtain the model order selection result; and least squares reconstruction uses least squares to reconstruct the model selection result to obtain a fine estimate of the elevation scattering characteristics. The proposed method can overcome the limitations of traditional model selection methods based on likelihood models and penalty terms, which involve traversal and trial and error, and helps to improve the imaging efficiency of high-precision SAR 3D imaging. The technical approach of this invention is described in [link to invention]. Figure 1 .

[0007] The main steps of the proposed method include:

[0008] Step 1: Based on the SAR two-dimensional imaging model, use the measured data to perform two-dimensional SAR imaging of the observed scene, and generate tomographic SAR measured data and two-dimensional SAR complex image data.

[0009] Step 2: Extract echoes and model observations from the measured 2D SAR image data generated in Step 1, using each azimuth-range resolution cell, to generate an echo set {g}. 1 ,...,g I}, Observation matrix set {L 1 ,...,L I};

[0010] Step 3: Utilize the echo set {g} generated in Step 2 1 ,...,g I}, Observation matrix set {L 1 ,...,L I The dimensionality-reduced low-dimensional vector set is obtained by using an iterative soft-threshold compressed sensing method.

[0011] Step 4: Use the data from Step 3 to obtain the set of low-dimensional vectors after dimensionality reduction. Solve for its minimum norm-based solution and obtain its set of autocorrelation matrix magnitude feature maps and phase feature maps {(Abs,Phase)}. 1 ,...,(Abs,Phase) I};

[0012] Step 5: Utilize the actual spatial observation geometric priors to generate a training dataset, train the proposed order feature selection network, and obtain the order feature selection network model.

[0013] Step 6: Input the magnitude feature map and phase feature map of the autocorrelation matrix obtained in Step 4 into the order feature selection network obtained in Step 5 to obtain the model selection result;

[0014] Step 7: Using the model order selection results obtained in Step 6, obtain the precise estimation results of elevation scattering characteristics through least squares reconstruction.

[0015] Step 8: Using the high-altitude scattering characteristic estimation results obtained in Step 7, map them into the three-dimensional scattering characteristic space to obtain the tomographic SAR three-dimensional image.

[0016] In step one, the Fast Decomposition Back Projection (FFBP) imaging algorithm is used when performing two-dimensional SAR imaging of the observation scene using measured data based on the SAR two-dimensional imaging model.

[0017] The FFBP algorithm flow is described below:

[0018] Fast Factorized Back-Projection (FFBP) is a temporal imaging method used for 2D SAR imaging. Its basic principle is to divide the aperture into multiple sub-apertures and perform recursive fusion operations on these sub-apertures, ultimately generating an image with high range and angular resolution on a polar coordinate grid. This image is then projected onto an imaging grid on the ground plane.

[0019] The implementation steps of the FFBP algorithm are as follows: Figure 2 As shown, it can be mainly divided into three stages:

[0020] Sub-aperture projection stage: The entire aperture is divided into several sub-apertures, and the echo data after range pulse compression is projected onto a polar coordinate grid established with the center of each sub-aperture as the pole, to obtain a polar coordinate sub-image.

[0021] Recursive fusion stage: On the polar coordinate grid, radix-2 recursive fusion is performed on two adjacent sub-images, and the sub-images are continuously merged until an image with high distance resolution and high angular resolution is obtained.

[0022] Polar coordinate to rectangular coordinate conversion stage: The image located on the polar coordinate grid is projected onto the imaging grid of the ground distance plane, and the polar coordinates are converted to rectangular coordinates to obtain the final two-dimensional SAR imaging result.

[0023] The FFBP algorithm, by employing sub-aperture projection and recursive fusion operations, can generate high-resolution SAR imaging results in a short computation time, making it one of the fastest time-domain imaging methods currently available.

[0024] In step two, the echo set {g} 1 ,...,g I The measured data observation matrix set {L} is generated by stacking the registered 2D SAR images of the data at position i on the same range-azimuth plane. 1 ,...,L I The method is as follows: based on the vertical baseline B of the SAR system in space. s_all The slant distance path r is calculated based on the pixel coordinates at position i, and then based on B... s_all Calculate the measured data observation matrix L for each imaging point using r. s as follows:

[0025]

[0026] Where λ is the wavelength of the emitted electromagnetic wave, and v is the elevation sampling vector;

[0027] In step three, the ISTA algorithm is used to perform a point-by-point order reduction and inversion to obtain a low-dimensional vector. in The following iterative calculation is performed to obtain the result:

[0028]

[0029]

[0030] Where ρ is the regularization parameter, which can be set to (L i ) T (L i The maximum singular value;

[0031] In step four, the minimum norm solution Obtained through the following methods:

[0032]

[0033] in, From low-dimensional vectors The position corresponding to the non-zero value of L i It consists of a portion of column vectors.

[0034] The set of amplitude and phase characteristic maps of the autocorrelation matrix of the minimum norm solution {(Abs 1 Phase 1 ),...,(AbsI Phase I )} is obtained through the following method:

[0035]

[0036]

[0037] Where H represents taking the conjugate transpose, and angle represents taking the phase.

[0038] In step five, the proposed order feature selection network consists of four parts: an amplitude feature extraction module, a phase feature extraction module, a feature fusion module, and a feature selection module. Figure 3 As shown.

[0039] Each feature extraction layer consists of one convolutional layer (Conv), one batch normalization layer (BN), and one nonlinear activation layer (ReLU). The amplitude and phase feature extraction modules together comprise 12 feature extraction layers, the structure of which is shown in Table 1.

[0040] Table 1. Composition structure of the feature extraction module

[0041]

[0042] The feature fusion selection module consists of 6 feature extraction layers, and the structure of these 6 layers is shown in Table 2:

[0043] Table 2 Feature Extraction Layer Structure

[0044]

[0045]

[0046] The feature selection module consists of two fully connected layers, Fc1 and Fc2. Fc1 reconstructs the input feature map into a one-dimensional feature map and maps it to the same-dimensional feature map. The current function of Fc2 is to select the feature map. Therefore, the input feature map needs to be mapped to the output map with the same dimension as the scattering vector.

[0047] The proposed training data generation and network training are carried out through the following steps:

[0048] (1) Randomly generate scattering vector γ label , using γ label Set the target location to 1 and generate the data label P. label In the set of measured data observation matrices {L 1 ,...,L I Randomly select L = L from} i Using data labels γ labelCombine the observation matrix L with 0-40dB random complex Gaussian noise to obtain the height-dimensional data vector g.

[0049] g=L×γ label +ε.

[0050] (2) Using the method in step three, obtain the dimensionality-reduced low-dimensional vector γ. R

[0051] (3) Using the method for obtaining the autocorrelation matrix amplitude feature map and phase feature map in step four, obtain the low-dimensional vector γ. R Amplitude feature map and phase feature map {Abs trian Phase trian} as training input data;

[0052] (4) Generate training data {Abs} based on (1)-(3). trian Phase trian} and P label Several groups.

[0053] The loss function for network training is:

[0054] J loss =MSE(P out ,P label )

[0055] In step six, the set of autocorrelation matrix amplitude feature map and phase feature map {(Abs 1 Phase 1 ),...,(Abs I Phase I After selecting the network by inputting the order features, the network input becomes the model selection result.

[0056] In step seven, the set of precise estimation results for elevation scattering characteristics is obtained. The value is obtained through the following least squares estimation:

[0057]

[0058] in, By P out The position corresponding to the non-zero value of L i It consists of a portion of column vectors.

[0059] In step eight, the set of precise estimation results of elevation scattering characteristics is used. By mapping to a three-dimensional scattering characteristic space, a tomographic SAR three-dimensional image is obtained.

[0060] The advantages of this invention are:

[0061] (1) This invention models the model selection problem in SAR 3D imaging based on compressed sensing into a deep learning feature selection model, utilizing the nonlinear solution capability of deep learning networks. Through end-to-end mapping, it avoids the complex traversal optimization, parameter tuning, and penalty term selection of traditional methods, which can greatly improve the efficiency of model selection.

[0062] (2) This invention designs a new paradigm for three-dimensional imaging of tomographic SAR. Based on tomographic SAR data, three-dimensional images are acquired through four main steps: order dimensionality reduction, norm reconstruction, model selection based on feature selection network, and least squares reconstruction. The proposed method has high three-dimensional imaging efficiency and imaging accuracy. Attached Figure Description

[0063] Figure 1 A roadmap for deep learning-based tomographic SAR 3D imaging methods based on order dimensionality reduction and model selection;

[0064] Figure 2 The flowchart is for the FFBP algorithm;

[0065] Figure 3 Selecting network structure diagrams for order features;

[0066] Figure 4 This is a schematic diagram of the geometry of SAR data acquisition and target distribution in an implementation example;

[0067] Figure 5 A schematic diagram of two-dimensional SAR imaging results with different baselines in the implementation example;

[0068] Figure 6 This is a schematic diagram of a low-dimensional scattering vector in an implementation example;

[0069] Figure 7 The diagram shows the minimum norm solution of the low-dimensional scattering vector in the implementation example.

[0070] Figure 8 The amplitude and phase feature maps of the autocorrelation matrix are shown in the implementation example.

[0071] Figure 9 The image shown is a tomographic SAR three-dimensional image obtained in the example. Detailed Implementation

[0072] Computer simulation of P-band tomographic SAR three-dimensional imaging. The geometry of SAR data acquisition and target distribution are as follows: Figure 4 As shown.

[0073] The radar transmits signals at a center frequency of 800MHz (band), with a total baseline length of 170-260m in the altitude direction, generating a total of 7 data tracks. The signal bandwidth is 150MHz, the shortest slant range from the radar to the center of the observation scene is 917m, and the signal-to-noise ratio (SNR) is set to 10dB. Figure 4 The target is a scattering surface target with a height of 0m and 5m.

[0074] Following the method in step one, the Fast Front Back Projection (FFBP) algorithm is used for two-dimensional imaging to obtain two-dimensional imaging results along 21 baselines, as follows: Figure 5 As shown.

[0075] Following the method in step two, the echo set {g} is obtained sequentially by distance and azimuth. 1 ,...,g I} and calculate the observation matrix set {L} based on SAR observation priors. 1 ,...,L I}

[0076] Following the method in step three, using the echo set {g 1 ,...,g I}, Observation matrix set {L 1 ,...,L I A low-dimensional set of scattering vectors is calculated using a compressed sensing method based on iterative soft thresholding. like Figure 6 The diagram shows a low-dimensional scattering vector of a pixel with overlapping points in the range-azimuth direction.

[0077] Following the method in step four, solve for its minimum norm solution, see... Figure 7 And obtain the amplitude and phase feature maps of the autocorrelation matrix, see... Figure 8 .

[0078] Following the method in step five, simulation data and data labels were generated. A total of 110,000 sets of data were generated, of which 100,000 sets were used as the training set and 10,000 sets were used as the test set. The proposed order feature selection network was pre-trained using the training data to obtain the order feature selection network model.

[0079] Following the method in step six, the amplitude and phase feature maps of the autocorrelation function of the measured data are input into the order feature selection network model to obtain the model selection results.

[0080] Following the method in step seven, the model selection results are reconstructed using least squares to obtain a precise estimation result of the elevation scattering characteristics.

[0081] Following the method in step eight, the obtained high-altitude scattering characteristic estimation results are mapped onto the three-dimensional scattering characteristic space to obtain the tomographic SAR three-dimensional image, such as... Figure 9 As shown.

[0082] This completes the generation of tomographic SAR 3D images using a deep learning-based tomographic SAR 3D imaging method based on order dimensionality reduction and model selection.

Claims

1. A deep learning-based tomographic SAR three-dimensional imaging method based on order dimensionality reduction and model selection, characterized in that, The method includes the following steps: Step 1: Based on the SAR two-dimensional imaging model, use the measured data to image the observed scene and generate tomographic SAR measured data and two-dimensional SAR complex image data; Step 2: Extract echoes and model observations from the measured 2D SAR images generated in Step 1, using each azimuth-range resolution cell, to generate an echo set. Observation matrix set ; Step 3: Utilize the echo set generated in Step 2 Observation matrix set The set of low-dimensional vectors after dimensionality reduction is obtained by using an iterative soft-threshold compressed sensing method. ; Step 4: Use the data from Step 3 to obtain the set of low-dimensional vectors after dimensionality reduction. and echo set Observation matrix set Solve for the minimum norm solution of the low-dimensional vector set, and obtain its set of autocorrelation matrix magnitude feature maps and phase feature maps. ; Step 5: Using the observation matrix of the actual task, generate a training dataset, train the order feature selection network, and obtain the order feature selection network model. Step 6: Input the magnitude feature map and phase feature map of the autocorrelation matrix obtained in Step 4 into the order feature selection network obtained in Step 5 to obtain the model selection result; Step 7: Using the model order selection results obtained in Step 6, obtain the precise estimation results of elevation scattering characteristics through least squares reconstruction. ; Step 8: Map the high-altitude scattering characteristics estimation results obtained in Step 7 onto the three-dimensional scattering space to obtain the tomographic SAR three-dimensional image.

2. The deep learning tomographic SAR three-dimensional imaging method based on order dimensionality reduction and model selection as described in claim 1, characterized in that... In step four, the minimum norm solution Obtained through the following methods: ； in, From low-dimensional vectors There are non-zero value positions corresponding to It consists of a portion of column vectors; Set of amplitude and phase characteristic maps of the autocorrelation matrix of the minimum norm solution Obtained through the following methods: ； in, H 'angle' indicates taking the conjugate transpose, and 'angle' indicates taking the phase.

3. The deep learning tomographic SAR three-dimensional imaging method based on order dimensionality reduction and model selection as described in claim 1, characterized in that... In step five, the order feature selection network consists of four parts: an amplitude feature extraction module, a phase feature extraction module, a feature fusion module, and a feature selection module.

4. The deep learning tomographic SAR three-dimensional imaging method based on order dimensionality reduction and model selection as described in claim 1, characterized in that... In step five, the feature extraction layer consists of one convolutional layer (Conv), one batch normalization layer (BN), and one nonlinear activation layer (ReLU). The amplitude feature extraction module and the phase feature extraction module consist of 12 feature extraction layers; the feature fusion and selection module consists of 6 feature extraction layers; the feature selection module consists of two fully connected layers, Fc1 and Fc2. Fc1 reconstructs the input feature map into one dimension and maps it to features of the same dimension, while Fc2 performs feature selection on the features output by Fc1.

5. The deep learning tomographic SAR three-dimensional imaging method based on order dimensionality reduction and model selection as described in claim 1, characterized in that... In step five, the training data generation and network training are performed through the following steps: (1) Randomly generate scattering vectors ,use Set the target location to 1 and generate a data label. In the set of measured data observation matrices Random selection Using data tags With observation matrix L And add 0-40dB random complex Gaussian noise to obtain a high-dimensional data vector. g ； (2) Using the method in step three, obtain the dimensionality-reduced low-dimensional vector. ; (3) Using the method for obtaining the autocorrelation matrix amplitude feature map and phase feature map in step four, obtain the low-dimensional vector. amplitude feature map and phase feature map As training input data; (4) Generate training data based on (1)-(3). and Several groups, The loss function for network training is: 。

Images