A method for synthetic aperture radar automatic target recognition

Abstract

The application discloses a kind of synthetic aperture radar automatic target identification method, belong to radar automatic target identification technical field;Including the following steps: step 1: pre-training DenseNet model, learn general image features by using ImageNet dataset;Step 2: initialization twinborn network, using the weight of pre-training DenseNet model is initialized;Step 3: pair SAR image input, obtains similarity score;Step 4: establish two-class classifier, by similarity score establishes model and exports model evaluation index;Step 5: fine-tuning model network, on SAR dataset the parameter of entire twinborn network and two-class classifier is fine-tuned;The application has less parameter amount, operation amount and higher recognition accuracy in SAR automatic target identification aspect, simultaneously compared with other convolutional neural network experiments show that the complexity and overall recognition accuracy of DenseNet model based on twinborn network and migration learning on MSTAR dataset all have superiority, in addition, also solve the gradient vanishing problem of model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar automatic target recognition, and particularly relates to a synthetic aperture radar automatic target recognition method. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing technology that uses signal processing techniques such as pulse compression to accurately obtain high-resolution images of target range and azimuth angle. By using a longer pulse width and a fast-moving or moving platform, SAR can achieve a longer synthetic aperture, which in turn enables high-resolution imaging. Synthetic aperture radar automatic target recognition is an important task in remote sensing, aiming to identify specific targets from SAR images. It has a wide range of applications, such as military reconnaissance, environmental monitoring, and disaster management. The unique features of SAR images, such as high resolution and all-weather imaging capability, make them suitable for many applications that rely on accurate object recognition.

[0003] In recent years, deep learning has achieved remarkable success in SAR automatic target recognition. With the help of convolutional neural networks (CNNs), deep learning has shown excellent results in detecting and classifying targets in SAR images. However, due to the high cost of data collection and processing, the number of SAR image datasets is often limited. This makes it challenging to build an accurate target recognition model using traditional deep learning methods, as it requires a large amount of labeled data to avoid overfitting. Twin networks are introduced in image classification tasks with limited samples to reduce overfitting and enhance generalization ability by training two identical networks with shared weights and structures. The idea is to force the two networks to learn discriminative features from the training data, reducing intra-class variability and inter-class similarity. This model has been proven to be effective in SAR automatic target recognition. In addition, transfer learning is another powerful method that uses pre-trained models on large datasets to solve new tasks with limited samples. The idea is to initialize model parameters using knowledge learned from the source domain and then fine-tune them on the target domain. Transfer learning has been successfully applied to various computer vision applications such as object detection, segmentation, and recognition.

[0004] Although the above-mentioned method for SAR recognition has certain effectiveness, due to the difficulty of obtaining SAR samples and the cost of labeled data, the sample size is often limited, and the use of convolutional neural networks can cause overfitting and gradient disappearance phenomenon, thereby affecting the recognition effect.

[0005] The disclosure of the foregoing Background Art is given solely for the purpose of aiding in the understanding of the present application by the public, and does not necessarily erect the foregoing Background Art into the state of the art as it existed before the filing date of this patent application. The foregoing Background Art should not be used to interpret or limit the novelty and inventive step of the present application, unless explicitly incorporated by reference into the description of the present application. SUMMARY

[0006] In view of the above problems, the present application attempts to fuse the twin network and transfer learning method, and proposes a method of synthetic aperture radar automatic target recognition based on SNTLDenseNet. The twin network and transfer learning are used to solve the limited sample challenge related to SAR automatic target recognition, wherein the twin network architecture is used to guide a recognition model to learn the low-dimensional part of the SAR image. In order to further solve the problem of limited samples in the SAR image data set, transfer learning is used to learn knowledge from a large-scale image data set. The absolute value function is used to represent the intra-class compactness and inter-class dispersion, and a binary classifier is established to capture these features. Experiments are carried out in the public data set MSTAR to study the method of slowing down the network gradient disappearance phenomenon.

[0007] The technical scheme adopted by the present application is:

[0008] A method of synthetic aperture radar automatic target recognition, comprising the following steps:

[0009] Step 1: pre-training the DenseNet model by using the ImageNet data set to learn general image features;

[0010] Step 2: initialize the twin network, and initialize the weight of the pre-trained DenseNet model;

[0011] Step 3: input a pair of SAR images to obtain a similarity score;

[0012] Step 4: establish a binary classifier, establish a model through the similarity score and output a model evaluation index;

[0013] Step 5: fine-tune the model network, and fine-tune the parameters of the entire twin network and binary classifier on the SAR data set.

[0014] In step 1, during the pre-training process of the DenseNet model, one-hot encoding is a method for representing data categories in a format easy to understand in machine learning algorithms. It involves creating a binary vector y (n) for a label with length equal to the number of categories k (for the ImageNet data set, k = 1000). (n)The category corresponding to the true class is set to 1, and other categories are set to 0. This encoding allows comparison and calculation between classified data points. The weight θ in the DenseNet model is denoted as f θ The input x (n) is converted into a binary vector z (n) with a length equal to the number of categories k, that is, z (n) = f θ (x (n) ). The cross-entropy loss of the model in the pre-training process is:

[0015]

[0016] Step 2: Initialize the twin network, and the pre-trained model weights are initialized;

[0017] The pre-trained DenseNet model is used with sample pair labels (x ′(i) , x ′(j) , y ′(n) ) as input.

[0018] Step 3: Input pairs of SAR images to obtain similarity scores;

[0019] The absolute value function is used to obtain the similarity score of the sample pair, that is:

[0020] z ′(n) = |f θ (x ′(i) )-f θ (x ′(j) )| (2)

[0021] Step 4: Establish a binary classifier, build a model through similarity scores and output model evaluation indicators;

[0022] A binary classifier based on a single-layer fully connected network is established, which has k input nodes and 1 output node, with z ′(n) as input. Let f θ be a binary classifier with trainable weights θ. The Sigmoid function S is applied to activate the output of the binary classifier, and the binary cross-entropy loss is established to train the binary classifier. The binary cross-entropy loss can be written as:

[0023]

[0024] Step 5: Fine-tune the model network.

[0025] The one or more technical solutions provided in the embodiments of the application have at least the following technical effects or advantages: the SNTL DenseNet model based on a twin network and transfer learning provided by the application has less parameter quantity, less operation quantity and higher recognition accuracy in SAR automatic target recognition, and compared with other convolutional neural network experiments, the SNTL DenseNet model based on a twin network and transfer learning has superiority in complexity and overall recognition accuracy on the MSTAR data set. In the case of using transfer learning, the convergence speed and convergence state of each model are significantly better than those of the non-transfer learning model, and the gradient vanishing problem of the model is solved. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0027] Figure 1 Fig. 1 shows the overall structure of the SAR target recognition method based on SNTL DenseNet in the embodiment;

[0028] Figure 2 Fig. 2 shows the content of the SAR sample set in the embodiment;

[0029] Figure 3 Fig. 3 shows a model comparison experiment data graph;

[0030] Figure 4 Fig. 4 shows a structural schematic diagram of DenseNet;

[0031] Figure 5 Fig. 5 shows a schematic diagram of the internal structure of a DenseBlock. DETAILED DESCRIPTION

[0032] Reference will now be made in detail to the specific implementations of the application. Although the application is described in conjunction with these specific implementations, it is to be understood that the application is not limited to these specific implementations. Rather, the specific implementations are intended to cover alternatives, modifications and equivalents, which can be included within the spirit and scope of the application as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the application. The application can be practiced without some or all of these specific details. In other instances, well known process operations have not been described in detail in order not to unnecessarily obscure the application.

[0033] The singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0034] The task herein is to identify whether a pair of SAR images belongs to the same class or not. Let the total number of training samples be N. Traditional deep learning methods use sample-label pairs (x (n) ,y (n) )(n∈{1,2,...,N}) to train neural networks. The sample-label pairs used herein are (x ′(i) ,x ′(j) ,y ′(n) )(i,j∈{1,2,...,N},i≠j) to train neural networks. In (x (n) ,y (n) ), x (n) represents a sample; and for the SAR target recognition task herein, x (n) is a two-dimensional matrix, y (n) represents the label corresponding to the sample x (n) , generally ranging from {1,2,...,K-1}, where K is the number of target classes. However, in (x ′(i) ,x ′(j) ,y ′(n) ), x ′(i) and x ′(j) represent different samples, and if x ′(i) and x ′(j) belong to the same class, then y ′(n) =1, otherwise y ′(n) =0. In this paper, the sample pairs of the training set and the test set are randomly divided, and the sample pairs in the training set are used to train the neural network model. Then, the trained model is tested using the sample pairs in the test set to verify whether the unknown sample pairs can be effectively identified and whether they belong to the same class. Therefore, the accuracy and loss value indicators are defined based on the sample pairs.

[0035] With reference to Figures 1-5 , the present application proposes a method for automatic target recognition of SAR based on DenseNet121(TL) of SAR, comprising the following steps:

[0036] Step 1: Pre-training the DenseNet model by learning general image features using the ImageNet dataset;

[0037] Step 2: initialize the twin network, use the weight of the pre-trained DenseNet model to initialize;

[0038] Step 3: input the pair of SAR images, get the similarity score;

[0039] Step 4: establish a binary classifier, establish a model through the similarity score and output the model evaluation index;

[0040] Step 5: fine-tune the model network, fine-tune the parameters of the entire twin network and binary classifier on the SAR data set.

[0041] Preferably, the pre-trained DenseNet model includes a dense block and a transition layer;

[0042] (1) Dense block

[0043] In the dense block module, there is a close relationship between any two layers; all layers of features are spliced with each other, that is, the output results of all layers before any layer are superimposed as the input of the layer, and then the results of the layer and the output results of the previous layer are transmitted as the input of the next layer. If a network has L layers, DenseNet has L(L+1) / 2 connections;

[0044] For an L-layer network, the input of the ith layer is obtained from formula (4):

[0045] x l =H([x0,x1,...,x l-1 ]) (4)

[0046] Wherein:

[0047] 1) x0, x1,..., x l-1 represent the feature maps corresponding to the 0th layer to the (l-1)th layer, respectively;

[0048] 2) [x0, x1,..., x l-1 ] represents cascading the feature maps corresponding to the 0th layer to the (l-1)th layer;

[0049] 3) H l (·) represents a composite function including BN, Relu and Conv three operations.

[0050] In the dense block module, each composite function H l(·) can output k feature maps, then the l-th layer will input k0+k*(l-1) feature maps, where k0 represents the number of channels in the input layer; k is a hyperparameter also known as the network growth rate; a bottleneck layer is added to the dense block, that is, a 1*1 convolution is introduced before the 3*3 convolution to reduce dimensionality. The structure of the dense block after adding the bottleneck layer is: BN+Relu+Conv(1*1)+BN+Relu+Conv(3*3);

[0051] (2) Transition layer

[0052] The transition layer connects two adjacent dense blocks, reduces the size of the feature maps through pooling, and makes the model more compact. A compression factor θ is introduced, with a value ranging from (0,1). When the dense block outputs m feature maps, the transition layer can generate θ*m features through convolution. When θ is less than 1, the number of feature maps is compressed to θ times the original value; when θ is 1, no compression occurs. The structure of the transition layer after introducing the hyperparameter θ is as follows:

[0053] BN-Relu-Conv(1*1)(filternum:θ*m)-dropout-Pooling(2*2).

[0054] Preferably, in step 3, the SAR image is obtained by using active microwave remote sensing technology and signal processing techniques such as pulse compression to accurately obtain a high-resolution image of the target range and azimuth angle; wherein, the SAR image uses the publicly available dataset MSTAR.

[0055] Preferably, the similarity score calculation function in step 3 is: z ′(n) =|f θ (x ′(i) )-f θ (x ′(j) )|;where: f θ For the weights θ in the DenseNet model, the input x (n) Convert to a binary vector z with a length equal to the number of categories k (n) .

[0056] Preferably, in step 4, the binary classifier refers to a binary classifier based on a single-layer fully connected network, which has k input nodes and 1 output node, with z... ′(n) As input; let f θ For a binary classifier with trainable weights θ, the output of the binary classifier is activated by applying the sigmoid function S, and a binary cross-entropy loss formula is established to train the binary classifier.

[0057] Preferably, the formula for the binary cross-entropy loss is:

[0058]

[0059] In particular:

[0060] Overall design of network structure:

[0061] SNTLDenseNet includes three stages: pre-training of DenseNet model, initialization of the twin network using the pre-trained model weights, and fine-tuning of the twin network for SAR target recognition. First, a DenseNet model is pre-trained on a large-scale natural image dataset, such as ImageNet, to learn general image features. Then, a twin network is established and initialized using the weights of the pre-trained DenseNet. The twin network takes pairs of SAR images as input and outputs their similarity scores. Next, a binary classifier is built on top of the similarity scores and outputs a classification score. Finally, the entire twin network and binary classifier are fine-tuned on a limited sample SAR dataset to adapt to SAR-specific features and improve their performance. The overall framework of SNTLDenseNet is shown in Figure 1 .

[0062] DenseNet-based transfer learning model

[0063] Feature extraction plays an important role in SAR target recognition, effectively converting high-dimensional data from the image domain to the low-dimensional feature domain, and then implementing classification. In limited sample recognition tasks, traditional supervised learning models cannot be used due to insufficient training samples. To overcome this difficulty, this paper proposes a new feature extraction model that combines the ability to learn similarity from different samples, thereby enhancing target recognition effectiveness. A feature embedding structure based on a twin network is designed in this model to detect the differences between different samples in the low-dimensional feature domain. This paper proposes a weight sharing method for the feature embedding network based on the twin structure. DenseNet is used as the network model for feature extraction, and different models are compared and analyzed in the experiment, such as EfficientNetV2, MobileNetV3, ResNet50, ResNet152, ResNeXt50, and ShuffleNetV2. Due to the tendency of the DenseNet model to overfit on small datasets, transfer learning is introduced into this framework model. First, a DenseNet model is trained using a large-scale ImageNet dataset, and then a twin network is constructed using the pre-trained DenseNet model to transfer knowledge from the source domain of ImageNet to the target domain of SAR images. The image size of ImageNet is 224x224, and the SAR image size in this paper is 100x100, as shown in Figure 1 .

[0064] During the pre-training process of DenseNet-based models, one-hot encoding is a method for representing data categories in a format that is easy to understand in machine learning algorithms. It involves creating a binary vector y with a length equal to the number of categories K (K = 1000 for the ImageNet dataset). (n) Each element in the vector corresponds to a category y. (n) The true class is set to 1, and other classes are set to 0. This encoding allows for comparison and computation between categorical data points. In the DenseNet model, the weights θ are denoted as f. θ , will input x (n) Convert to a binary vector z with a length equal to the number of categories K (n) , i.e. z (n) =f θ (x (n) The cross-entropy loss during model pre-training can be written as:

[0065]

[0066] Binary classifier

[0067] Because a pre-trained DenseNet model is used to pair samples with labels (x... ′(i) ,x ′(j) ,y ′(n) As input, to better describe intra-class compactness and inter-class dispersion, an absolute value function is used to obtain the similarity score of sample pairs, i.e.: z ′(n) =|f θ (x ′(i) )-f θ (x ′(j) )|

[0068] Construct a binary classifier based on a single-layer fully connected network. This network has K input nodes and 1 output node, with z... ′(n) As input.

[0069] Let f θ Let's consider a binary classifier with trainable weights Θ. We activate the classifier's output using the sigmoid function S and train it using a binary cross-entropy loss. The binary cross-entropy loss can be written as:

[0070]

[0071] Experiments and Results

[0072] 1. Experimental Data

[0073] To evaluate SNTLDenseNet, this paper uses the public dataset MSTAR as the evaluation benchmark, which is collected under the support of the U.S. Defense Advanced Research Projects Agency (DARPA) and the U.S. Defense Advanced Research Projects Agency (AFRL). The project is implemented by the Sandia National Laboratory to collect a large amount of target data through flight experiments, and this dataset is a subset of this project. As shown in Figure 2 Figure 1, the MSTAR dataset has ten different types of ground military vehicle targets (self-propelled artillery: 2S1; armored vehicles: BMP2, BRDM2, BTR60, and BTR70; tanks: T62, T72; self-propelled anti-aircraft gun: ZSU234; bulldozer: D7; military truck: ZIL131). The SAR image of each type of vehicle is obtained from 0° to 360°, with an azimuth angle change of 5° to 6°, and a spatial resolution of 0.3m x 0.3m, as shown in Table 1.

[0074] Table 1 Summary of benchmark test dataset

[0075]

[0076] 2. Experimental setup

[0077] Experiments were conducted on DenseNet-based models and compared with EfficientNetV2, ResNet50, ResNet152, ResNeXt50, and ShuffleNetV2, using the same dataset ImageNet for pre-training. To evaluate the fusion model under different limited sample scenarios, part of the training dataset in Table 1 was randomly selected to train the twin network, and an hyperparameter γ was introduced to represent the proportion of selected samples between training datasets. The deep learning platform used is Ubuntu 22.04.2 LTS system, Python 3.9 is selected as the language, Pytorch 2.0.0 is used as the deep learning architecture, and the computer hardware configuration is 126GB, Xeon(R) Gold 6230R CPU @ 2.10 GHz x 104 and GPU Nvidia Tesla T416GB x 8. The optimizer Adam is used for default training, the batch size is 32, and the epoch is 1000.

[0078] 3. Comparison experiment with other deep learning methods

[0079] To evaluate the model under different limited sample scenarios, the parameter γ = 0.25, 0.5, 1.0 is set. The experimental results are as follows Figure 3As shown in FIG. 6, where "TL" means "transfer learning", indicating that this CNN-based model is initialized with the weights trained by ImageNet; the CNN-based model without the label "TL" is trained from scratch, which is randomly initialized.

[0080] As shown in FIG. 6, where "TL" means "transfer learning", indicating that this CNN-based model is initialized with the weights trained by ImageNet; the CNN-based model without the label "TL" is trained from scratch, which is randomly initialized. Figure 3 As shown in FIG. 6, where "TL" means "transfer learning", indicating that this CNN-based model is initialized with the weights trained by ImageNet; the CNN-based model without the label "TL" is trained from scratch, which is randomly initialized.

[0081] Based on the twin network SAR target recognition problem. In addition, DenseNet, ResNet50 and ResNeXt50 can still converge well without using transfer learning. This shows that the four models are all suitable for twin network-based SAR target recognition. Among them, the test set recognition accuracy of DenseNet under different γ values has a high horizontal asymptote, so in terms of SAR target recognition performance, DenseNet is the optimal model.

[0082] Table 2 Complexity and test accuracy of different CNN-based models

[0083]

[0084] In order to further comprehensively compare the complexity and recognition performance of different models, the total number of model parameters (Params), memory usage (Memory), floating point operation times (FLOPs), multiply-add operation times (MAdd), and test set accuracy with or without transfer learning (TL) are counted. As shown in Table 2, the recognition performance of DenseNet is the best, and the computational complexity is the lowest among several models.

[0085] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it; although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent ones without departing from the spirit of the technical solutions of the present application, and all of them should be covered in the technical solution range claimed by the present application.

Claims

1. A method for automatic target identification using synthetic aperture radar, characterized in that: Includes the following steps: Step 1: Pre-train the DenseNet model to learn general image features using the ImageNet dataset; Step 2: Initialize the Siamese network using the weights of the pre-trained DenseNet model; Step 3: Input paired SAR images and obtain similarity scores; Step 4: Build a binary classifier, establish a model through similarity scoring, and output model evaluation metrics; Step 5: Fine-tune the model network. Fine-tune the parameters of the entire Siamese network and the binary classifier on the SAR dataset. The pre-trained DenseNet model includes dense blocks and transition layers; (1) Dense mass In dense block modules, there is a close connection between any two layers; the features of all layers are concatenated, that is, the outputs of all layers preceding a given layer are superimposed as the input of that layer, and then the result of that layer and the outputs of the previous layers are used as the input of the next layer; if a network has DenseNet has layers, and it has a total of One connection; for For a layered network, the input of the l-th layer is obtained from formula (1): (1)； in: 1) They represent the layers from 0 to 1. The feature map corresponding to the layer; 2) This indicates the range from level 0 to level 1. The feature maps corresponding to the layers are concatenated; 3) This represents a composite function that includes BN, ReLU, and Conv operations; Each composite function in the dense block module It can output The first feature map, then the second feature map... The layer will input Feature map, where This indicates the number of channels in the input layer; It is a hyperparameter also known as the network growth rate; adding a bottleneck layer to a dense block, i.e., in Introduced before convolution Convolution is used to reduce dimensionality. The dense block structure after adding a bottleneck layer is as follows: ； (2) Transition layer The transition layer connects two adjacent dense blocks, and pooling is used to reduce the size of the feature map, making the model more compact; a compression factor is introduced. The range of values ​​is Dense block output When dealing with feature maps, the transition layer can generate [feature maps] through convolution. One characteristic; when When the number is less than 1, the number of feature maps is compressed to the original value. times; when When the value is 1, no compression is performed; a hyperparameter is introduced. The subsequent transition layer structure is as follows: 。 2. The method for automatic target identification using synthetic aperture radar according to claim 1, characterized in that: In step 3, the SAR image is obtained by using active microwave remote sensing technology and pulse compression signal processing technology to accurately obtain a high-resolution image of the target range and azimuth angle; the SAR image is obtained from the publicly available dataset MSTAR.

3. The method for automatic target identification using synthetic aperture radar according to claim 2, characterized in that: The similarity score calculation function in step 3 is: (2)； in: Weights in the DenseNet model , will input Convert to a format with a length equal to the number of categories. binary vector .

4. The method for automatic target identification using synthetic aperture radar according to claim 3, characterized in that: In step 4, the binary classifier refers to a binary classifier built based on a single-layer fully connected network, which has... One input node and one output node, with As input; let For having trainable weights A binary classifier, applied function Activate the output of the binary classifier and establish a binary cross-entropy loss formula to train the binary classifier.

5. The method for automatic target identification using synthetic aperture radar according to claim 4, characterized in that: The formula for binary cross-entropy loss is: (3)； Where: the length is equal to the number of categories k for the ImageNet dataset, The label creates a binary vector Each element in the vector corresponds to a category. .

Images