A SAR target recognition method based on scattering center feature extraction

Abstract

The present invention discloses a SAR target recognition method based on scattering center feature extraction, including: Step 1, obtaining a SAR image to be recognized; Step 2, inputting the SAR image to be recognized into a trained integrated network for scattering center extraction and target recognition to obtain a recognition result. By embedding the physical model of the SAR image into a deep neural network, the present invention constructs a deeply interpretable network model, realizes a SAR target recognition network driven jointly by mechanism and data, and improves the interpretability of the existing deep neural network by extracting scattering center features with clear physical meanings.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and particularly relates to a SAR target recognition method based on scattering center feature extraction. Background Art

[0002] In recent years, with the rapid development of Synthetic Aperture Radar (SAR) imaging technology, SAR images contain information of targets in the azimuth dimension and the range dimension, and can intuitively reflect characteristics such as the shape and contour of the targets. These information play an important role in target category judgment. Therefore, radar target recognition based on SAR images has received extensive attention. Due to the powerful feature mining ability of deep neural networks, methods based on deep neural networks have made significant progress in the field of SAR target recognition. However, the deep neural network is a pure data-driven model, with poor network interpretability and limited generalization ability. In order to combine the physical mechanism of SAR images and improve the interpretability and generalization ability of the deep neural network, SAR target recognition methods based on scattering centers have gradually received more and more attention.

[0003] Traditional SAR target recognition methods based on deep neural networks ignore the physical characteristics of SAR images, with poor network interpretability, and the generalization of this pure data-driven model is poor. The present invention combines the physical model of SAR images, embeds electromagnetic scattering characteristics into the deep neural network, extracts physically interpretable scattering center features with SAR images, thereby improving the interpretability of the deep neural network, and the scattering center features have robust electromagnetic scattering characteristics, and the target recognition network based on scattering center features has good generalization.

[0004] Existing SAR target methods based on scattering centers include two independent steps of scattering center extraction and target recognition, that is, first, a sparse reconstruction algorithm is used to extract scattering centers, and then a target recognition network is constructed to achieve SAR target recognition. This approach will lead to a mismatch between the extracted scattering center features and the target recognition network, limiting the final recognition performance; moreover, these methods use traditional sparse solution algorithms for scattering center feature extraction, and there are problems in traditional sparse solution algorithms that are difficult to determine hyperparameters. The performance of the extracted scattering centers depends on the setting of empirical parameters, and the effect of scattering center extraction is usually limited, which will affect the subsequent recognition results; at the same time, the computational complexity of traditional sparse solution algorithms is high, and the time efficiency of the scattering center extraction process is low, and the real-time performance is poor. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a SAR target recognition method based on scattering center feature extraction.

[0006] The technical problems to be solved by the present invention are realized through the following technical solutions:

[0007] A SAR target recognition method based on scattering center feature extraction, the SAR target recognition method comprising:

[0008] Step 1, obtaining a SAR image to be recognized;

[0009] Step 2, inputting the SAR image to be recognized into a trained integrated network for scattering center extraction and target recognition to obtain a recognition result;

[0010] Wherein, the integrated network for scattering center extraction and target recognition includes a scattering center extraction module, an image reconstruction module and a target recognition module. The scattering center extraction module is used to extract corresponding scattering centers from the input SAR image to output scattering coefficients. The image reconstruction module is used to perform sparse reconstruction on the SAR image using a Fourier dictionary according to the scattering coefficients output by the scattering center extraction module, and output a sparse reconstruction image, so that the loss function of the image reconstruction module generates a constraint on the scattering center extraction module during training. The target recognition module is used to construct 3D point cloud data according to the scattering coefficients output by the scattering center extraction module, and use the 3D point cloud data to implement target recognition to obtain a recognition result.

[0011] Optionally, the scattering center extraction module includes a convolutional neural network, and the convolutional neural network is used to fit a non-linear mapping function. During training, the input of the convolutional neural network is the scattering coefficient obtained in the previous iteration, and the output of the convolutional neural network is iteratively solved using a sparse reconstruction algorithm to obtain a scattering coefficient;

[0012] The convolutional neural network includes 4 sequentially connected convolutional layers, the size of the convolutional kernel is set to 3×3, and the first three convolutional layers contain an activation function ReLU, the number of output channels is 32, and the last convolutional layer does not contain an activation function, and the number of output channels is 1.

[0013] Optionally, the iterative formula of the AMP algorithm is expressed as:

[0014]

[0015] H k = α k D T D - I

[0016] R k-1 = s - Dvec(z k-1 )

[0017] Wherein, z k represents the scattering coefficient of the k-th iteration, z k-1 represents the scattering coefficient of the (k - 1)-th iteration, αk denotes an adjustable control parameter, D denotes the Fourier dictionary, T denotes the matrix transpose operation, vec denotes vectorizing a matrix, and vec -1 (·) denotes matrixizing a vector, denotes a non - linear mapping function, I denotes the identity matrix, and s denotes the input SAR image.

[0018] Optionally, the image reconstruction module is composed of a single - layer linear fully - connected layer, and the weight matrix of the single - layer linear fully - connected layer is fixed as the Fourier dictionary;

[0019] Each column of the Fourier dictionary of the SAR image is expressed as:

[0020] D = [d1(f, φ),..., d i (f, φ),..., d K0 (f, φ)]

[0021]

[0022]

[0023] where D denotes the Fourier dictionary, ||·||2 denotes the L2 norm, vec denotes vectorizing a matrix, f denotes frequency, f ∈ (f c -B / 2, f c +B / 2), f c denotes the radar carrier frequency, B denotes the bandwidth, φ denotes the azimuth angle, φ ∈ (-φ m / 2, φ m / 2), φ m denotes the rotation angle, c denotes the speed of light, denotes the imaginary part unit, (x i , y i ) denotes the two - dimensional position coordinates.

[0024] Optionally, the loss function of the image reconstruction module is expressed as:

[0025] L r = ||s - Dz|| 2 + λL1(z)

[0026] where L r denotes the loss function of the image reconstruction module, z denotes the scattering coefficient, L1(z) denotes the sparse constraint term, λ denotes an adjustable weight parameter, and s denotes the original SAR image.

[0027] Optionally, the target recognition module includes a point cloud recognition network and a multi - layer MLP network, where:

[0028] A point cloud recognition network for constructing the 3D point cloud data by using M scattering centers included in the scattering coefficient, where the 3D point cloud data is represented as C i represents the i-th scattering center, C i =(x i , y i , a i ), (x i , y i ) represents the two-dimensional position coordinates, and a i represents the amplitude;

[0029] A scale normalization network for performing a normalization operation on the 3D point cloud data to obtain the normalized 3D point cloud data;

[0030] A multi-layer MLP network for obtaining a predicted probability vector through the normalized 3D point cloud data;

[0031] Among them, the multi-layer MLP network includes a first fully connected layer, a second fully connected layer, a first average pooling layer, a third fully connected layer, a fourth fully connected layer, a fifth fully connected layer, a second average pooling layer, a sixth fully connected layer, and a seventh fully connected layer connected in sequence. The output of the seventh fully connected layer is mapped to the corresponding label through softmax to obtain the predicted probability vector.

[0032] Optionally, the loss function of the target recognition module is expressed as:

[0033]

[0034] Among them, L c represents the loss function of the target recognition module, C represents the total number of target categories, y c represents the true label, represents the predicted label.

[0035] Optionally, the training method of the scattering center extraction and target recognition integrated network includes:

[0036] S1. Set the number of iterations epoch to q, the maximum number of iterations to Q, and set the learning rate ρ of the stochastic gradient algorithm;

[0037] S2. Randomly initialize the network parameters in the scattering center extraction module and the target recognition module by using the normal distribution;

[0038] S3. Select m SAR images from the training dataset to form a training sample group, and a total of T / training sample groups are obtained;

[0039] S4. Calculate the loss function \(L\) of each training sample group using the loss function of the integrated network for scattering center extraction and target recognition. m , and use the stochastic gradient descent algorithm to optimize the loss function \(L\) of each training sample group m in turn to train the network parameters in the scattering center extraction module and the target recognition module, and complete one iteration of the training process;

[0040] S5. Calculate the loss function between two consecutive iterations until the change rate of the loss function is less than 10 -3 or \(q = Q\), then terminate the iteration to obtain the trained integrated network for scattering center extraction and target recognition.

[0041] Optionally, the loss function \(L\) m is expressed as:

[0042]

[0043] where \(s\) i represents the \(i\)-th SAR image in the training sample group, \(D\) represents the Fourier dictionary, \(z\) i represents the scattering coefficient of the \(i\)-th SAR image in the training sample group, \(\lambda\) represents an adjustable weight parameter, \(L_1(z\) i ) represents the sparse constraint term of the scattering coefficient of the \(i\)-th SAR image in the training sample group, \(C\) represents the total number of target categories, represents the true label of the \(i\)-th SAR image in the training sample group, represents the predicted label of the \(i\)-th SAR image in the training sample group.

[0044] Optionally, after step S5, it further includes:

[0045] S6. Input the SAR image sample \(s\) to be predicted in the test dataset * into the scattering center extraction module to obtain the scattering coefficient \(z\) * ;

[0046] S7. Input the scattering coefficient \(z\) * into the target recognition module to obtain the predicted label vector \(y\) * ;

[0047] S8. Based on the label vector \(y\) * , determine the category to which the dimension with the maximum probability value of the SAR image sample \(s\) to be predicted belongs, and complete the category prediction of the SAR image sample \(s\) * to be predicted. * Compared with the prior art, the beneficial effects of the present invention:

[0048]

[0049] The present invention proposes a SAR target recognition method based on scattering center feature extraction, and constructs an end-to-end integrated network for scattering center extraction - target recognition. This network includes three modules: a scattering center extraction module, an image reconstruction module, and a target recognition module. The scattering center extraction module is used to extract scattering center features; the image reconstruction module combines the scattering center model of SAR images to perform image sparse reconstruction on the extracted scattering center features; the target recognition module performs target recognition on the extracted scattering center features. These three modules are within an end-to-end deep neural network framework, and using this network, the scattering center features and class prediction results of the test image can be directly inferred. Thus, by embedding the physical model of SAR images into a deep neural network, the present invention constructs a deeply interpretable network model, realizes a SAR target recognition network driven jointly by mechanism and data, and improves the interpretability of existing deep neural networks by extracting scattering center features with clear physical meanings.

[0050] The SAR target recognition method proposed by the present invention realizes the integration of scattering center extraction - target recognition for the first time, which can effectively avoid the problem of mismatch between scattering center features and the target recognition network caused by the existing two-stage independence, thereby achieving higher recognition accuracy; at the same time, compared with the two-stage independent method, the proposed method has higher time efficiency and better practicality in actual situations.

[0051] The following will further elaborate on the present invention in conjunction with the accompanying drawings. Description of the Drawings

[0052] Figure 1 is a schematic flowchart of a SAR target recognition method based on scattering center feature extraction provided by an embodiment of the present invention;

[0053] Figure 2 is an overall block diagram of a SAR target recognition method based on scattering center feature extraction provided by an embodiment of the present invention;

[0054] Figure 3 is a schematic diagram of a scattering center extraction module provided by an embodiment of the present invention;

[0055] Figure 4 is a schematic diagram of an image reconstruction module provided by an embodiment of the present invention;

[0056] Figure 5 is a schematic diagram of a point cloud recognition module provided by an embodiment of the present invention. Detailed Embodiments

[0057] The following further describes the present invention in detail with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0058] Embodiment 1

[0059] Currently, most synthetic aperture radar (SAR) target recognition methods based on deep neural networks are pure data-driven models, which ignore the physical characteristics of SAR images, resulting in poor physical interpretability of the models. The scattering center feature is a typical physical feature of SAR images. This feature describes the electromagnetic scattering characteristics of targets from the perspective of SAR imaging, and the generalization of the scattering center feature is good, which can improve the generalization ability of the model on the test data set. Therefore, by combining the extraction of scattering center features of SAR images, a suitable network model can be designed to achieve the joint drive of physical mechanism and data, and improve the physical interpretability and generalization of deep neural networks. Therefore, the key technical problems solved by the present invention are how to integrate the physical mechanism of SAR images into deep neural networks, design a physically interpretable network to extract the scattering center features of SAR images, and how to design a suitable classification network to use the extracted scattering center features for recognition, ultimately achieving better SAR target recognition performance.

[0060] Therefore, please refer to Figure 1 and Figure 2 , Figure 1 which is a schematic flow chart of a SAR target recognition method based on the extraction of scattering center features provided by an embodiment of the present invention, Figure 2 and which is an overall block diagram of a SAR target recognition method based on the extraction of scattering center features provided by an embodiment of the present invention. An embodiment of the present invention provides a SAR target recognition method based on the extraction of scattering center features. The SAR target recognition method includes:

[0061] Step 1, obtain the SAR image to be recognized;

[0062] Step 2, input the SAR image to be recognized into the integrated network for scattering center extraction and target recognition that has been trained, and obtain the recognition result;

[0063] Among them, the integrated network for scattering center extraction and target recognition includes a scattering center extraction module, an image reconstruction module, and a target recognition module. The scattering center extraction module is used to extract the corresponding scattering centers from the input SAR image to output scattering coefficients. The image reconstruction module is used to perform sparse reconstruction of the SAR image using the Fourier dictionary according to the scattering coefficients output by the scattering center extraction module, and output the sparse reconstructed image, so that the loss function of the image reconstruction module generates constraints on the scattering center extraction module during the training process. Thus, the scattering center extraction module can extract the desired scattering center features. The target recognition module is used to construct 3D point cloud data according to the scattering coefficients output by the scattering center extraction module, and use the 3D point cloud data to achieve target recognition and obtain the recognition result.

[0064] In a specific embodiment, the scattering center extraction module includes a convolutional neural network, which is used to fit a non-linear mapping function. During the training process, the input of the convolutional neural network is the scattering coefficient obtained from the previous iteration, and the output of the convolutional neural network is iteratively solved using a sparse reconstruction algorithm to obtain the scattering coefficient;

[0065] The convolutional neural network includes 4 sequentially connected convolutional layers. The size of the convolutional kernel is set to 3×3, and the first three convolutional layers contain the activation function ReLU, with the number of output channels being 32. The last convolutional layer does not contain an activation function, and the number of output channels is 1.

[0066] That is to say, the purpose of the scattering center extraction module is to extract the corresponding scattering centers from the input SAR image. Generally, the process of extracting the scattering centers of the SAR image can be regarded as the following sparse solution problem:

[0067]

[0068] Among them, s represents the input SAR image, z represents the scattering coefficient to be solved, D represents the Fourier dictionary, λ represents an adjustable weight parameter, ||·||2 and ||·||1 represent the L2 and L1 norms respectively. To solve the scattering coefficient z, the sparse reconstruction algorithm AMP is used for iterative solution, and the iterative formula of this algorithm is expressed as:

[0069]

[0070] H k =α k D T D-I

[0071] R k-1 =s-Dvec(z k-1 )

[0072] Among them, k represents the number of iterations, vec(·) represents vectorizing a matrix, and vec -1 (·) represents matrixizing a vector, represents the non-linear mapping function, and α k represents an adjustable control parameter. To solve the scattering coefficient z using the above iterative formula, the present invention uses a deep unfolding network AMP-Net to form the scattering center extraction module.

[0073] Please refer to Figure 3 , and the scattering center extraction module constructed in the embodiment of the present invention will be further described below.

[0074] The scattering center extraction module deeply unfolds the AMP algorithm and uses a convolutional neural network to fit the The non - linear mapping process forms the deep unfolding network AMP - Net. Specifically, the input of the scattering center extraction module is the SAR image s. In this embodiment, four convolutional layers are used to fit The size of the convolutional kernel is set to 3×3, and the first three convolutional layers contain the activation function ReLU, with the number of output channels being 32; the last layer does not contain an activation function, and the number of output channels is 1. The input of the convolutional neural network is the scattering coefficient z obtained from the previous iteration k-1 , and the output of the network is substituted into the AMP algorithm iteration. Finally, the scattering center extraction module outputs the scattering coefficient z of the SAR image s.

[0075] In this embodiment, the purpose of the image reconstruction module is to perform sparse reconstruction of the original input SAR image using the Fourier dictionary D according to the scattering coefficient z obtained from the scattering center extraction module. According to the scattering center model:

[0076] s = Dz

[0077] The scattering coefficient z contains the position and amplitude information of the scattering centers, and the scattering centers have the characteristic of sparse distribution on the target. Therefore, the SAR image obtained based on the above - mentioned scattering center model represents the sparse reconstruction of the original SAR image s.

[0078] In a specific embodiment, please refer to Figure 4 , and the image reconstruction module constructed in the embodiment of the present invention will be further described below.

[0079] The input of the image reconstruction module is the scattering coefficient z obtained from the scattering center extraction module, and the output is the sparse reconstruction image of the original SAR image The image reconstruction module is composed of a single - layer linear fully - connected layer, and the weight matrix of this single - layer linear fully - connected layer is fixed as the Fourier dictionary D. For the SAR image, each column of its Fourier dictionary is expressed as:

[0080] D = [d1(f,φ),...,d i (f,φ),...,d K0 (f,φ)]

[0081] where and

[0082]

[0083] where f represents the frequency, f ∈ (f c -B / 2,f c +B / 2), f c represents the radar carrier frequency, B represents the bandwidth; φ represents the azimuth angle, φ ∈ (-φ m / 2,φm / 2), φ m represents the rotation angle, c represents the speed of light, represents the imaginary part unit, (x i , y i ) represents the two-dimensional position coordinates.

[0084] The loss function L of the image reconstruction module r is expressed as:

[0085] L r = ||s - Dz|| 2 + λL1(z)

[0086] where L1(z) represents the sparse constraint term, the present invention adopts the L1 norm to constrain the sparsity of the scattering coefficient z, and λ represents an adjustable weight parameter.

[0087] In this embodiment, the purpose of the target recognition module is to construct a point cloud recognition network based on the scattering coefficient z obtained by the scattering center extraction module to achieve target recognition. First, assume that the scattering coefficient z extracted from the original SAR image s by the scattering center extraction module contains M scattering centers, and each scattering center is composed of a position and an amplitude. The i-th scattering center is expressed as C i = (x i , y i , a i )(i = 1, …, M), where (x i , y i ) represents the two-dimensional position coordinates, and a i represents the amplitude. Therefore, the 3D point cloud data constructed by the M scattering centers contained in the scattering coefficient z is expressed as

[0088] In a specific embodiment, please refer to Figure 5 , the target recognition module includes a point cloud recognition network and a multi-layer MLP (fully connected layer) network, where:

[0089] The point cloud recognition network is used to construct 3D point cloud data by using the M scattering centers contained in the scattering coefficient;

[0090] The scale normalization network is used to perform a normalization operation on the 3D point cloud data to obtain the normalized 3D point cloud data;

[0091] The multi-layer MLP network is used to obtain the predicted probability vector through the normalized 3D point cloud data;

[0092] Among them, the multi-layer MLP network includes a first fully-connected layer, a second fully-connected layer, a first average pooling layer, a third fully-connected layer, a fourth fully-connected layer, a fifth fully-connected layer, a second average pooling layer, a sixth fully-connected layer, and a seventh fully-connected layer connected in sequence. The output of the seventh fully-connected layer is mapped to the corresponding label through softmax to obtain a predicted probability vector.

[0093] That is to say, in this embodiment, the constructed 3D point cloud data is first input into the scale normalization network. The affine transformation is adopted in the present invention to normalize each point cloud data C i =(x i , y i , a i ). After that, the normalized 3D point cloud data is input into the multi-layer MLP network, and the output of the seventh fully-connected layer is mapped to the corresponding label through softmax to obtain a predicted probability vector.

[0094] Here, the loss function L c of the target recognition module is expressed as:

[0095]

[0096] Among them, y c represents the true label. For example, it contains ten types of targets. Therefore, the label y c ∈{0, 1, 2, …, 9}; represents the predicted label, which is the output of the c-th node of the seventh fully-connected layer.

[0097] In a specific embodiment, the present invention also provides a training method for an integrated network of scattering center extraction and target recognition. Before training, it is necessary to generate a training data set and a test data set. Specifically, in the embodiment of the present invention, 2746 SAR images with a radar working pitch angle of 17° in the "Moving and Stationary Target Acquisition and Recognition" (MSTAR) data set containing ten types of vehicle targets are used as the training data set, and 2426 SAR images with a radar working pitch angle of 15° are used as the test data set.

[0098] In this embodiment, the training method for the integrated network of scattering center extraction and target recognition includes:

[0099] S1. Set the number of iterations epoch to q, the maximum number of iterations to Q, and set the learning rate ρ of the stochastic gradient algorithm. For example, Q = 100 and ρ = 10 -4 .

[0100] S2. Randomly initialize the network parameters in the scattering center extraction module and the target recognition module using a normal distribution.

[0101] S3. Select m SAR images from the training dataset to form a training sample group (i.e., minibatch), and a total of T / training sample groups are obtained, such as a total of 2746 / m minibatches.

[0102] S4. Calculate the loss function L of each training sample group using the loss function of the integrated network for scattering center extraction and target recognition. m and use the stochastic gradient descent algorithm to optimize the loss function L of each training sample group in turn. m to train the network parameters in the scattering center extraction module and the target recognition module, and complete one iteration of the training process.

[0103] Specifically, select m SAR images from the training dataset to form a minibatch. For example, a total of 2746 / m minibatches are obtained. Calculate the loss function L of the network constructed by the samples in each minibatch. m :

[0104]

[0105] Use the stochastic gradient descent algorithm to optimize the above objective function, train the network parameters in the scattering center extraction module and the target recognition module, and perform iterative optimization of the network parameters using 2746 / m minibatches in turn to complete one epoch of the training process. In this embodiment, m = 2, λ = 0.5, and C = 10.

[0106] S5. Calculate the loss function between two consecutive iterations until the change rate of the loss function is less than 10. -3 or q = Q, then terminate the iteration to obtain the trained integrated network for scattering center extraction and target recognition.

[0107] After obtaining the trained integrated network for scattering center extraction and target recognition, the trained integrated network for scattering center extraction and target recognition is tested.

[0108] S6. Input the SAR image sample s to be predicted in the test dataset * into the scattering center extraction module to obtain the scattering coefficient z. * .

[0109] S7. Input the scattering coefficient z * into the target recognition module to obtain the predicted label vector y. * .

[0110] S8. Based on the label vector y* , the SAR image sample s to be predicted * is judged as the category to which the dimension with the maximum probability value belongs, and the category prediction of the SAR image sample s to be predicted is completed. *

[0111] First, the present invention embeds the physical model of SAR images into a deep neural network, constructs a scattering center extraction module, and accurately extracts the scattering center features of SAR images based on the deep learning mechanism. Existing methods all use traditional sparse reconstruction algorithms for scattering center extraction. The traditional algorithms have high computational complexity and difficult-to-determine hyperparameters, resulting in inaccurate extraction of scattering center features and affecting the target recognition performance. The method proposed in the present invention constructs a deep network to learn the non-linear mapping relationship between SAR images and scattering center features based on the stochastic gradient descent algorithm using training data, and can learn scattering center features more accurately, thus being beneficial to the improvement of target recognition performance.

[0112] Second, the method proposed in the present invention is an end-to-end integrated deep framework for scattering center extraction - target recognition, including a scattering center feature extraction module, an image reconstruction module, and a target recognition module, and jointly optimizes the network parameters of each module using the stochastic gradient algorithm. In most current methods, scattering center extraction and target recognition are two independent steps, which will lead to the mismatch between the extracted scattering center features and the target recognition network, affecting the final recognition result. In addition, after the training of each module in the method proposed in the present invention is completed, the scattering center features and recognition results of the test image can be obtained through forward propagation in the test stage. Compared with the two-stage independent method that needs to re-use the sparse reconstruction algorithm to extract scattering centers, the method proposed in the present invention has higher time efficiency, demonstrating the time efficiency of the proposed method in practical applications.

[0113] The following further illustrates the effect of the present invention through experimental data of actual measurements:

[0114] 1. Experimental Conditions and Experimental Contents

[0115] The software platform used in this experiment is: Ubuntu 18.04 Linux operating system, python 3.6, pytorch.

[0116] The hardware platform used in this experiment is: Dell T7910 workstation, CPU: Intel Core(TM)i7-4770, GPU: NVIDIA GeForce RTX 3080Ti.

[0117] The data used in this experiment is the MSTAR measured data set used in SAR target recognition experiments, which is measured by an X-bandwidth SAR sensor. The SAR sensor parameters of this data set are shown in Table 1.1. In this data set, the form of each data is a complex-valued SAR image, and the resolution of each SAR image is, and the image size is 0.3m×0.3m. The azimuth range of all SAR images in the data set is from 0° to 360°, and the pitch angles include 17° and 15°. This data set contains ten types of vehicle targets, namely: BMP2, BTR70, T72, T62, BRDM2, BTR60, ZSU23 / 4, D7, ZIL131 and 2S1. The training data set is 2746 SAR images at a pitch angle of 17°, and the test data is 2426 SAR images at a pitch angle of 15°. The specific number of samples for each class is shown in Table 1.2.

[0118] Table 1.1 SAR Sensor Parameters of MSTAR Data Set

[0119]

[0120] Table 1.2 SAR Sensor Parameters of MSTAR Data Set

[0121]

[0122] 2. Experimental Results and Analysis

[0123] Evaluation index for the recognition results of the test data set: Recognition accuracy = number of correctly judged samples / total number of samples. In order to verify the recognition performance of the method proposed in the present invention, Table 1.3 gives the confusion matrix of the proposed method on ten types of test samples. Each row represents the recognition result of each type of target. The experimental results show that the proposed method can accurately recognize each type of target, with a low misrecognition rate on other types, and finally achieves a relatively good average recognition accuracy.

[0124] Table 1.3 Confusion Matrix of the Proposed Method on Ten Types of Test Samples

[0125]

[0126]

[0127] To compare with some other comparison methods, the comparison methods include the traditional method SVM, the convolutional neural network VGG16, and the existing scattering center-based recognition method FGL. The comparison experiment results are shown in Table 1.4. According to the experimental results, the proposed method achieves better recognition accuracy compared with the comparison methods. Compared with the traditional method SVM, the proposed method is a deep neural network method with powerful feature mining ability and can achieve better recognition performance. Compared with the convolutional neural network VGG16, the proposed method combines the physical characteristics of SAR images, mines the scattering center features that can represent the structural characteristics of SAR targets, and thus achieves better performance. Compared with the existing scattering center-based recognition method FGL, the proposed method is an end-to-end deep neural network, avoiding the problem of mismatch between feature extraction and target recognition, and thus has higher recognition accuracy performance.

[0128] Table 1.4 Comparison results of recognition rates between the proposed method and comparison methods

[0129] Different methods Recognition accuracy rate (%) SVM 96.78 VGG16 98.85 FGL 99.08 The proposed method 99.51

[0130] As deep neural networks gradually show powerful feature mining ability and non-linear fitting ability, SAR target methods based on deep neural networks have received increasing attention in the field of radar sensing. Currently, most SAR target methods based on deep neural networks are pure data-driven methods, that is, using a large amount of SAR image training data to learn separable features related to classification, ignoring the physical characteristics of SAR images themselves, such as electromagnetic scattering characteristics. This pure data-driven neural network method has a "black box" structure, and the extracted features are relatively abstract, with certain decision-making risks in practical applications. With the in-depth research on deep neural networks, in addition to ensuring the performance of the model, it is also hoped to combine physical mechanisms with deep neural networks, improve the transparency of deep neural networks, learn interpretable features, and reduce decision-making risks in practice. Therefore, the present invention embeds the physical model of SAR images into a deep neural network, designs a convolutional network structure to extract the scattering center features of SAR images, and this feature has a clear physical meaning, representing the position and amplitude information of strong scattering points in SAR images, effectively improving the interpretability of deep neural networks. Therefore, in the actual application scenario of SAR target recognition, the method proposed by the present invention has strong decision-making reliability.

[0131] The model proposed in the present invention can decouple radar parameters and the model. Specifically, the method proposed in the present invention constructs an image reconstruction module based on the physical model of SAR images, uses the Fourier dictionary as the weight matrix of the network, and the radar imaging parameters are reflected in the Fourier dictionary. For SAR images obtained under different radar imaging parameters, by adjusting the corresponding Fourier dictionary, the scattering center features under the corresponding radar imaging parameters can be obtained, and the model has good generalization. The traditional SAR target methods based on deep neural networks only rely on structural features such as target contours and lines in training images. Therefore, these methods are more sensitive to changes in the structural characteristics of the images. When the radar imaging parameters change, the scattering characteristics of the SAR images change greatly, and the structural characteristics of the targets change. The recognition performance of these methods will decrease significantly, and the generalization of the model is poor. Therefore, the method proposed in the present invention can extract the target scattering center features in the SAR image data obtained under different radar parameters, has good generalization performance among different radar parameters, and has strong application prospects in actual scenarios.

[0132] It should be noted that the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more features. In the description of the present invention, "a plurality" means two or more, unless otherwise specifically defined.

[0133] In the description of this specification, the description with reference to the terms "an embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0134] Although the present invention has been described in conjunction with various embodiments herein, however, in the process of implementing the claimed invention, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings and the disclosure. In the specification, the word "including" does not exclude other components or steps, and "a" or "one" does not exclude a plurality of cases. Certain measures are described in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0135] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as falling within the protection scope of the present invention.

Claims

1. A SAR target recognition method based on scattering center feature extraction, characterized in that, The SAR target recognition method includes: Step 1, obtaining a SAR image to be recognized; Step 2, inputting the SAR image to be recognized into a trained integrated network for scattering center extraction and target recognition to obtain a recognition result; Among them, the integrated network for scattering center extraction and target recognition includes a scattering center extraction module, an image reconstruction module, and a target recognition module. The scattering center extraction module is used to extract corresponding scattering centers from the input SAR image to output scattering coefficients. The image reconstruction module is used to perform sparse reconstruction on the SAR image using a Fourier dictionary according to the scattering coefficients output by the scattering center extraction module, and output a sparse reconstructed image, so that the loss function of the image reconstruction module generates constraints on the scattering center extraction module during training. The target recognition module is used to construct 3D point cloud data based on the scattering coefficients output by the scattering center extraction module, and use the 3D point cloud data to achieve target recognition to obtain a recognition result; The scattering center extraction module includes a convolutional neural network, which is used to fit a non-linear mapping function. During training, the input of the convolutional neural network is the scattering coefficient obtained in the previous iteration, and the output of the convolutional neural network is iteratively solved using a sparse reconstruction algorithm to obtain a scattering coefficient; The convolutional neural network includes 4 sequentially connected convolutional layers. The size of the convolutional kernel is set to 3×3, and the first three convolutional layers contain the activation function ReLU, and the number of output channels is 32. The last convolutional layer does not contain an activation function, and the number of output channels is 1; The iterative formula of the sparse reconstruction algorithm is expressed as: Among them, represents the scattering coefficient of the th iteration, represents the scattering coefficient of the th iteration, represents an adjustable control parameter, represents the Fourier dictionary, represents the matrix transpose operation, represents vectorizing the matrix, represents matrixing the vector, represents the non - linear mapping function, represents the identity matrix, represents the input SAR image; The image reconstruction module is composed of a single-layer linear fully connected layer, and the weight matrix of the single-layer linear fully connected layer is fixed as a Fourier dictionary; Each column of the Fourier dictionary of the SAR image is expressed as: Among them, represents the Fourier dictionary, represents the L2 norm, represents vectorizing the matrix, represents frequency, , represents the radar carrier frequency, represents the bandwidth, represents the azimuth angle, , represents the rotation angle, represents the speed of light, represents the imaginary part unit, represents the two-dimensional position coordinates; The loss function of the image reconstruction module is expressed as: Among them, represents the loss function of the image reconstruction module, represents the scattering coefficient, represents the sparse constraint term, represents the adjustable weight parameter, represents the original SAR image.

2. The SAR target recognition method according to claim 1, characterized in that, The target recognition module includes a point cloud recognition network and a multi-layer MLP network, where: A point cloud recognition network for constructing the 3D point cloud data by using the scattering centers included in the scattering coefficient, where the 3D point cloud data is represented as , denotes the -th scattering center, , denotes the two-dimensional position coordinates, denotes the amplitude; A scale normalization network, which is used to perform a normalization operation on the 3D point cloud data to obtain a normalized 3D point cloud data; A multi-layer MLP network, which is used to obtain a predicted probability vector through the normalized 3D point cloud data; Among them, the multi-layer MLP network includes a first fully connected layer, a second fully connected layer, a first average pooling layer, a third fully connected layer, a fourth fully connected layer, a fifth fully connected layer, a second average pooling layer, a sixth fully connected layer, and a seventh fully connected layer connected in sequence. The output of the seventh fully connected layer is mapped to the corresponding label through softmax to obtain a predicted probability vector.

3. The SAR target recognition method according to claim 2, wherein, The loss function of the target recognition module is expressed as: Among them, represents the loss function of the target recognition module, represents the total number of target categories, represents the true label, represents the predicted label.

4. The SAR target recognition method according to claim 1, wherein The training method of the integrated network for scattering center extraction and target recognition includes: S1. Set the number of iterations epoch to , and the maximum number of iterations to . Set the learning rate of the stochastic gradient algorithm to ; S2, randomly initializing the network parameters in the scattering center extraction module and the target recognition module using a normal distribution; S3. Select SAR images from the training dataset to form a training sample group, and a total of T / training sample groups are obtained; S4. Calculate the loss function of each training sample group using the loss function of the integrated network for scattering center extraction and target recognition , and use the stochastic gradient descent algorithm to optimize the loss function of each of the training sample groups in turn, so as to train the network parameters in the scattering center extraction module and the target recognition module, and complete one iteration of the training process; S5. Calculate the loss function between two consecutive iterations until the change rate of the loss function is less than or , then terminate the iteration to obtain the trained integrated network for scattering center extraction and target recognition.

5. The SAR target recognition method according to claim 4, wherein The loss function is expressed as: Among them, represents the th SAR image in the training sample group, represents the Fourier dictionary, represents the scattering coefficient of the th SAR image in the training sample group, represents the adjustable weight parameter, represents the sparse constraint term of the scattering coefficient of the th SAR image in the training sample group, represents the total number of target categories, represents the true label of the th SAR image in the training sample group, represents the predicted label of the th SAR image in the training sample group.

6. The SAR target recognition method according to claim 4, wherein After step S5, it further includes: S6. Input the SAR image sample to be predicted in the test dataset into the scattering center extraction module to obtain the scattering coefficients ; S7. Input the scattering coefficient into the target recognition module to obtain a predicted label vector ; S8. Based on the label vector , determine the category to which the dimension with the maximum probability value belongs among the SAR image samples to be predicted , and complete the category prediction of the SAR image samples to be predicted .

Images