Synthetic aperture radar target identification method based on multi-stream attention

Abstract

The invention discloses a synthetic aperture radar target identification method based on multi-stream attention, and belongs to the field of radar image classification. The implementation method of the invention comprises the following steps: constructing an attention-based multi-stream convolutional neural network AMS-CNN, and carrying out SAR image classification by the AMS-CNN through extracting key and diversified features; enhancing the data set by using a GAN (Generative Adversarial Network) and improving the quality of the data set; performing parameter analysis optimization on the constructed AMS-CNN model by adopting the generated high-quality synthetic SAR image to obtain a mature network model; and the obtained mature network model is adopted to carry out target identification classification on the leaf-penetrating SAR image, and the precision and efficiency of synthetic aperture radar target identification are improved according to a target identification classification result. The method also comprises the following steps: carrying out the performance evaluation and testing of the mature network model, verifying and evaluating the characteristics and precision of the mature network model, and guaranteeing the precision and effect of the mature network model when the mature network model is applied to the target recognition of the synthetic aperture radar according to the verification and evaluation result.

Description

Technical Field

[0001] The present invention belongs to the field of radar image classification, and in particular relates to a synthetic aperture radar target recognition method based on multi-stream attention. Background Art

[0002] SAR, based on electromagnetic scattering and coherence, has strong penetration capabilities, making it an advantage over optical and infrared sensors as an active image sensor for detection and surveillance. Consequently, it has numerous applications in both military and civilian fields, with classification being one of the most fundamental. Due to imaging technology, SAR images differ significantly from natural images. The resolution of SAR images can vary depending on the scene. For example, in applications involving foliage penetration, the carrier frequency of SAR is often low to allow for penetration of the tree canopy, resulting in lower bandwidth in both the range and cross-range directions. Consequently, SAR images through foliage have poor resolution, making image classification more challenging.

[0003] In recent years, neural network-based classification of high-resolution natural images has made significant progress, focusing on deep and wide networks to achieve good performance. Increasing model depth results in models with a large number of parameters, making them difficult to deploy in many real-world applications. Therefore, improving model performance with negligible computational and parameter overhead is of great significance.

[0004] SENet, proposed by Hu, J, et al. in 2018, introduced an attention module that automatically determines the weight of each channel to improve model performance. Building on SENet, CBAM, proposed by Woo, S, et al. in 2018, also introduced a spatial attention module to further improve model performance. It is worth noting that the channel attention module and the spatial attention module do not increase the width or depth of the model, so they are very effective in improving model performance. MGTS, proposed by Chen, D, et al. in 2018, uses dual streams for pedestrian search, which can be used for detection and re-identification of target feature extraction. These models significantly improve the performance of natural images, but their effectiveness in low-resolution SAR remains unverified.

[0005] Generative Adversarial Networks (GANs) have been widely used in SAR image classification to enhance data. In 2019, Zheng, C, and other scholars used a multi-discriminator (MGAN) to generate synthetic images, which were then combined with real SAR images as CNN inputs. When the dataset is limited, the accuracy of CNN classification can be improved by adding synthetic samples during training. In 2021, Mao, C, and other scholars proposed a constrained naive GAN (CN-GAN), which stably generates data by combining the characteristics of the least squares generative network (LSGAN) and image-to-image conversion (Pix2Pix). The synthesized images have high quality. The dataset of SAR images that penetrate foliage is very limited, and there has been no attempt to generate a dataset of SAR images that penetrate foliage using a generative adversarial network. Summary of the Invention

[0006] The purpose of this invention is to provide a synthetic aperture radar target recognition method based on multi-stream attention. This method is easy to deploy on devices with limited computing resources and can improve the accuracy and efficiency of synthetic aperture radar target recognition. The invention also has the advantages of good robustness and generalization.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] The present invention discloses a synthetic aperture radar target recognition method based on multi-stream attention, comprising the following steps:

[0009] Step 1: Construct an attention-based multi-stream convolutional neural network (AMS-CNN). AMS-CNN classifies SAR images by extracting key and diverse features. The AMS-CNN consists of three parts.

[0010] The first part uses three convolution blocks and three pooling layers to extract low-level features, where the pooling layer is configured as maximum pooling to compress information and reduce computational overhead. Each convolution block consists of two convolution layers with different numbers of channels, and the pooling layer is located after the convolution block.

[0011] At the same time, the first part also introduces the channel attention module and the spatial attention module. The channel attention module is applied after the first pooling layer, and obtains comprehensive channel information through maximum pooling and average pooling combined with the multi-layer perceptron MLP. The spatial attention module focuses on key position information by fusing spatial features. It is applied after the second pooling layer to achieve the key extraction of bottom-level and high-level features.

[0012] The second part adopts a three-stream structure to extract maximum, average and median features respectively.

[0013] The first channel is processed by group averaging after passing through the convolution block and global maximum pooling to extract the average value features of the data.

[0014] The second channel is fully connected after convolution block and global maximum pooling to extract the maximum value feature of the data.

[0015] The third channel is processed by group median after convolution block and global maximum pooling to extract the median features of the data.

[0016] The maximum flow extracts important local features, the average flow and the median flow are used to capture global and intermediate features respectively. Through the combination of the three flows, the model can extract the diversity of image features and enhance the classification ability.

[0017] The third part performs image classification by combining the maximum, average and median features extracted from the three streams.

[0018] LeakyReLU is used as the activation function in AMS-CNN, and the regularized cross entropy loss function is used as the loss function L(W):

[0019]

[0020] Where W is the parameter in the network, y i is the true label of the sample, is the predicted label of the sample, L is the total number of layers, l is the number of layers, W [l] is the parameter of the lth layer, λ is the regularization parameter; n is the number of samples.

[0021] Step 2: Use the Generative Adversarial Network (GAN) to enhance the dataset and improve its quality.

[0022] The CARABAS-Ⅱ dataset is used as a sample, and high-quality low-resolution synthetic SAR images are generated by DCGAN to increase the amount of data used for model learning, thereby significantly improving the generalization ability of the model; the IS and FID methods are used to evaluate the high-quality synthetic SAR images.

[0023] Step 3: Use the high-quality synthetic SAR image generated in step 2 to perform parameter analysis and optimization on the AMS-CNN model constructed in step 1 to obtain a mature network model.

[0024] Model parameters have a significant impact on its classification ability. The present invention uses high-quality synthetic SAR images to perform parameter analysis and optimization on the AMS-CNN model, focusing on the effects of batch size, number of iterations, and regularization parameter settings on the classification ability, and obtains a mature network model.

[0025] Effective use of large batches can significantly improve model performance. However, when the CNN model uses stochastic gradient descent (SGD) to update the model, if the batch size is too large, it will affect its generalization performance. Therefore, it is necessary to optimize the batch size suitable for the model.

[0026] At the same time, the number of times the model parameters are updated is affected by the number of iterations. If the number is too low, the model cannot fully learn the image features for classification. The present invention uses L2 regularization to reduce overfitting. The regularization loss function is an effective method to reduce the risk of model overfitting, so it is necessary to optimize the number of model iterations and regularization parameters.

[0027] Step 4: Use the mature network model obtained in step 3 to perform target recognition and classification on the SAR image through leaves, and improve the accuracy and efficiency of synthetic aperture radar target recognition based on the target recognition and classification results.

[0028] It also includes step five: performance evaluation and testing of the mature network model in step four, verifying and evaluating the characteristics and accuracy of the mature network model, and ensuring the accuracy and effect of the mature network model when applied to synthetic aperture radar target recognition based on the verification and evaluation results.

[0029] A mature network model is used to perform target recognition and classification in SAR images that penetrate foliage. Five uniform noises of different intensities are added to the SAR images that penetrate foliage. The target recognition and classification performance of the model is analyzed and compared with three classic CNNs, AlexNet, GoogleNet, and VGG16, under the same conditions. The characteristics and accuracy of the mature network model are evaluated and analyzed. The accuracy and effectiveness of the mature network model when applied to synthetic aperture radar target recognition are ensured based on the verification and evaluation results.

[0030] Beneficial effects:

[0031] 1. The present invention discloses a synthetic aperture radar target recognition method based on multi-stream attention. By constructing a lightweight AMS-CNN model with a simple structure and small number of parameters, it can be deployed on devices with limited computing resources, thereby significantly reducing the complexity and cost of implementation.

[0032] 2. The present invention discloses a synthetic aperture radar target recognition method based on multi-stream attention, which adopts the attention mechanism and multi-stream to provide classification support for AMS-CNN. By obtaining important information and different features of the target, higher target recognition accuracy can be achieved while improving the robustness of the model.

[0033] 3. The present invention discloses a synthetic aperture radar target recognition method based on multi-stream attention, which uses DCGAN to generate high-quality low-resolution SAR data, providing the model with more diverse and representative sample data, improving the generalization ability of the model, and expanding new possibilities for the application of SAR images.

[0034] 4. The present invention discloses a synthetic aperture radar target recognition method based on multi-stream attention, which adopts a mature network model to perform target recognition and classification of SAR images that penetrate foliage. Five uniform noises of different intensities are added to the SAR images that penetrate foliage, and the target recognition and classification performance of the model is analyzed. The model is compared with three classic CNNs, AlexNet, GoogleNet and VGG16, under the same conditions, and the characteristics and accuracy of the mature network model are evaluated and analyzed. The accuracy and effect of the mature network model when applied to synthetic aperture radar target recognition are ensured based on the verification and evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a synthetic aperture radar target recognition method based on multi-stream attention of the present invention;

[0036] Figure 2 This is the structural diagram of the AMS-CNN model. DETAILED DESCRIPTION

[0037] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below with reference to the accompanying drawings and examples.

[0038] Example 1:

[0039] like Figure 1 As shown, the present embodiment discloses a synthetic aperture radar target recognition method based on multi-stream attention, and the specific implementation steps are as follows:

[0040] Step 1: Build an attention-based multi-stream convolutional neural network AMS-CNN lightweight model.

[0041] The constructed AMS-CNN model performs SAR image classification by extracting key and diverse features, such as Figure 2 As shown, the AMS-CNN consists of three parts.

[0042] The first part uses three convolution blocks and three pooling layers to extract low-level features, where the pooling layer is configured as maximum pooling to compress information and reduce computational overhead. Each convolution block consists of two convolution layers with different numbers of channels. The convolution layer is the key to extracting features. The convolution kernel of each layer is set to a step size of 1. The pooling layer is located after the convolution block and is used to compress local view information. In AMS-CNN, the pooling layer is configured as maximum pooling, where the pooling kernel of each layer is set to a step size of 2.

[0043] The activation function is located after the convolutional layer and can express the nonlinear mapping relationship between the model input and output. Traditional nonlinear activation functions include double tangent and sigmoid functions, but the gradients of these functions will vanish within a certain range, which is not conducive to network training. The ReLU activation function has good characteristics. When the input value is greater than zero, it can obtain a non-zero value, reducing the gradient vanishing. However, the network with the ReLU activation function cannot work when the value is less than zero. In order to make the activation function applicable to all values, LeakyReLU is supplemented on the basis of ReLU. The function of LeakyReLU is as follows:

[0044]

[0045] Here, x is the function input.

[0046] AMS-CNN uses LeakyReLU as the activation function to achieve nonlinearity of the network. The feature dimensions extracted by the convolution layer in the three convolution blocks are different, namely 16, 32 and 64 dimensions, respectively. [u] ,u∈{1,2,3} represents the output of the three convolutional blocks.

[0047] At the same time, the first part also introduces the channel attention module and the spatial attention module. The channel attention module is applied after the first pooling layer. It focuses on the relationship between the channels of the features. It obtains a one-dimensional vector through maximum pooling and average pooling, and then combines it with the multi-layer perceptron MLP to obtain comprehensive channel information. The MLP consists of three layers. The middle layer serves as a bottleneck. The unit size is one-third of the size of the first layer. The two types of information obtained by the MLP are added together, and then the sigmoid function is applied. The channel attention function is:

[0048] M c (F c )=σ(MLP(AvgPool(F c )))+σ(MLP(MaxPool(F c )))

[0049] Among them, F c Is the input of the channel attention module feature map AvgPool and MaxPool are operations that take the average and maximum values of each feature map. MLP is a multi-layer perceptron, and σ represents the sigmoid function.

[0050] The spatial attention module focuses on key position information by fusing spatial features. It is applied after the second pooling layer to achieve the key extraction of low-level and high-level features. It mainly focuses on the spatial connection between features. Similar to the channel attention module, the spatial attention module uses average pooling and maximum pooling to focus on the key spatial information along the channel axis. Then, the two parts of the results are merged through the channel to obtain a dual-channel feature map, and the convolution operation is performed on it to convert it into a single-channel feature map. Finally, the sigmoid function is applied to it. The spatial attention function is:

[0051] M s (F s )=σ(f 7×7 (AvgPool(F s );MaxPool(F s )))

[0052] Among them, F s is the input of the feature map of the spatial attention module, f 7×7 is a filter with a convolution kernel of 7, and σ represents the sigmoid function.

[0053] In the first part of AMS-CNN, two domains of attention blocks are used to focus on important features, where the channel attention module is located after the first convolution block and the spatial attention module is located after the second convolution block. The output of the attention module is the product between the input feature and the attention feature.

[0054] The second part of AMS-CNN adopts a three-stream structure, which is used to extract the maximum value, average value and median features respectively. The first channel is processed by group average after passing through the convolution block and global maximum pooling to extract the average value feature of the data. The second channel is processed by full connection after passing through the convolution block and global maximum pooling to extract the maximum value feature of the data. The third channel is processed by group median after passing through the convolution block and global maximum pooling to extract the median feature of the data. j L kTo represent the output of the kth layer of the jth stream in the second part, j, k∈{1,2,3}, where j=1,2,3 represents the maximum, average, and median structures, respectively. The maximum structure includes a convolutional layer with a large stride to extract local features, used to extract 64 features, with a convolution kernel set to 3×3. The average and median structures, on the other hand, focus on the details of the convolutional layer by using a smaller stride, with a convolution kernel set to 1×1 and 512 output channels. All three structures include a global max pooling layer after the convolution layer. The output equation of the maximum structure is:

[0055]

[0056] Since the first two layers of the average and median structures have the same structure, the output of their global maximum pooling layers is:

[0057]

[0058] Where n H is the height and width of the feature map, and i is the i-th channel.

[0059] Dropout effectively reduces overfitting by randomly omitting hidden units from the network with a given probability in training samples, especially when the dataset is limited. Dropout is typically applied to fully connected layers, making the network more lightweight and robust.

[0060] The third layer of the maximum architecture uses a fully connected layer with dropout, while the third layer of the mean and median architecture uses the group mean and median, with their weights and biases remaining almost unchanged. Using the maximum, mean, and median architecture not only captures specific features but also increases feature diversity. The output of the group mean is partitioned according to the number of classes, and then averaged for each class, resulting in the following output:

[0061]

[0062] Among them, a is the number of features in the second layer, and a / f is the number of categories.

[0063] The group median output is obtained by grouping the features and calculating the median value in each class. The output is:

[0064]

[0065] Among them, a is the number of features in the second layer, and a / f is the number of categories.

[0066] Since the mean and median structure shares all weights except the last layer, the parameters of AMS-CNN can be reduced.

[0067] The third part of AMS-CNN combines the maximum, median, and average features extracted in the second part and classifies the object.

[0068] In the forward propagation of the network, the parameters in the network need to be updated according to a specific strategy, which requires the use of a loss function. Commonly used loss functions are mean square error and cross entropy. The cross entropy loss function used in AMS-CNN can more accurately express the relationship between the network and the true distribution, thereby reflecting the difference between the true label and the predicted label. For n samples x1, x2, ..., x n , the cross entropy loss function is:

[0069]

[0070] Where W is the parameter in the network, y i For sample x i The true label, is the sample x i The predicted label of . By adding parameter regularization (i.e. dropout) to the loss function, the model overfitting can be reduced. The loss function after regularization is L(W):

[0071]

[0072] Among them, L is the total number of layers, l is the current layer number, and W [l] is the parameter of the lth layer, and λ is the regularization parameter.

[0073] In AMS-CNN, the optimization function Adam is used to update the training parameters in the loss function. The parameters in Adam are set to β1 = 0.9 and β2 = 0.999. The loss function in AMS-CNN is:

[0074]

[0075] where EB1(W), EB2(W) and EB3(W) are the cross entropy loss functions in each stream, and λ is set to 0.5.

[0076] Step 2: Use the Generative Adversarial Network (GAN) to enhance the dataset and improve its quality.

[0077] The GAN model is a generative model method that can be trained on a small amount of data. It consists of two components: a generator and a discriminator. The generator uses random Gaussian noise to generate images, while the discriminator determines whether the images created by the generator are real images. GAN focuses on using the generator to produce images that can deceive the discriminator.

[0078] In a typical GAN, both the generator G and the discriminator D are fully connected layers. G is a generative network that takes random noise data of a given dimension as input and outputs generated image data. D is a discriminator network (i.e., a classifier) that determines whether the generated image data represents real data. By inputting the image data generated by the generator and the real data with labels into the discriminator, the parameters of the generator and discriminator are iterated until the image generated by the generator can successfully deceive the discriminator. The loss function of the GAN is:

[0079]

[0080] Where x is a real sample, z is a noise sample, G(·) is the generator network, D(·) is the discriminator network, G(z) is the data generated by the noise through the generator network, D(x) is the possibility that the discriminator will judge x as a real sample data, and D(G(z)) is the probability that the discriminator will judge the generated data as real data.

[0081] DCGAN is similar in principle to GAN, but uses convolutional networks instead of the multilayer perceptron networks in the generator and discriminator. While GAN's sophisticated network structure enables self-supervised learning of data distributions, network training is unstable, resulting in some meaningless results. DCGAN is based on convolutional networks, and by adjusting the details of the network structure, it can accelerate network convergence and improve sample quality. Specifically, DCGAN's structural adjustments include adding data normalization to the generator and discriminator networks, using the tanh activation function in the final layer of the G network, and using the rectified linear unit (ReLU) activation function (LeakyReLU) as the D network function in other layers.

[0082] IS and FID are two widely used metrics to evaluate the quality of generated images. IS evaluates the generated image based on the JS divergence of two distributions, while FID evaluates the image by the mean and variance of the data distribution.

[0083] IS evaluates GAN networks from two perspectives: the quality and diversity of generated images. Higher quality images are associated with lower conditional entropy, while more diverse images are associated with higher marginal probability entropy. IS performance is measured using the Kullback-Leibler (KL) divergence. A higher KL divergence indicates higher diversity and quality of generated images. The IS formula is:

[0084]

[0085] Among them, p g (x) is the data distribution of the generated image, x~p g For x from p g The sampled image, DKL (p||q) is the KL divergence of distributions p and q, p(y|x) is the distribution of label y under the condition of x, is the distribution of labels y.

[0086] FID is based on the Inception network with 2048-dimensional features. Its principle is to make the generated data distribution as similar as possible to the real data distribution. The distance between the two distributions is calculated using the mean and covariance. The formula for FID is as follows:

[0087]

[0088] Among them, μ x is the mean of the distribution of the real data, μ y is the mean of the distribution of generated data, ∑ x is the distribution variance of the real data, ∑ y is the variance of the generated data, T r is the trace of the matrix. The lower the FID value, the closer the two distributions are, and the higher the quality and diversity of the generated images.

[0089] In this example, the CARABAS-II dataset was used as a sample. CARABAS-II is the only open-source dataset for foliage-penetrating SAR images, but it contains only 24 images. Therefore, data augmentation was performed on this dataset using DCGAN. The dataset was balanced by selecting the same number of vehicles for each class in the real data. Since TGB40 has the least data, the same number of samples was selected for TGB11 and TGB30, resulting in 168 samples per class, for a total of 504 samples. 396 of these samples were selected for training, and 108 for testing. DCGAN performed the same amount of data augmentation on each vehicle class in the dataset, generating 393 synthetic SAR images. Therefore, the dataset for foliage-penetrating SAR image classification in this example contains 897 samples. Table 1 provides the sample sizes used for training and testing.

[0090] Table 1 Sample size of real data and synthetic data in the dataset

[0091]

[0092] Step 3: Use the high-quality synthetic SAR image generated in step 2 to perform parameter analysis and optimization on the AMS-CNN model constructed in step 1 to obtain a mature network model.

[0093] This example uses the high-quality synthetic SAR images in step 2 to perform parameter analysis and optimization on the AMS-CNN model. Model parameters have a significant impact on its classification ability. This example uses high-quality synthetic SAR images to perform parameter analysis and optimization on the AMS-CNN model, focusing on the effects of batch size, number of iterations, and regularization parameter settings on classification ability, to obtain a mature network model.

[0094] Generally speaking, effectively utilizing large batches can significantly reduce model parameter updates and improve performance. The larger the batch size, the more accurate the descent direction and the smaller the oscillation. However, when CNN models use stochastic gradient descent (SGD) and its variants to update the model, using larger batches of parameters can actually reduce the quality of the generated performance. This is because large batches of CNN models often translate into sharp minima, which greatly affects the generalization performance of the model, while small batches are more likely to produce flat minima. The number of model parameter updates is affected by the number of iterations. Although the value of the loss function decreases with the increase in the number of model parameter updates, there is still a risk of model overfitting. When the model dataset is limited, CNN models are more prone to overfitting. Specifically, the model will show higher accuracy on the training data, but lower accuracy on the test data. Given that the dataset in our experiment is very limited, L2 regularization is used in the embodiment to reduce overfitting. L2 regularization can reduce the value of model parameters and reduce network complexity. Based on this, in this embodiment, we analyze the impact of batch size, iterations, and regularization parameters on the model.

[0095] Table 2 shows the classification accuracy results of the AMS-CNN model for different batch sizes and iterations. In the embodiment, the batch size is set to 16, 32, and 48 respectively, and these three different batch sizes are further used to obtain the classification accuracy. The results show that 32 is the best batch size, and when the batch size increases or decreases, the model classification accuracy will deteriorate. For the number of iterations, five different iteration optimization settings are performed in the embodiment, namely 400, 800, 1200, 1600 and 2000, among which when the iteration is set to 1600, the model has the highest classification accuracy. Afterwards, we also optimized the regularization parameters with different settings, setting them to 0, 0.1, 0.5 and 1 respectively. The results show that the regularization parameters in the model significantly reduce model overfitting, and when λ is set to 0.5, the model performs best, as shown in Table 3. Therefore, the regularization parameter is set to 0.5 in the embodiment.

[0096] Table 2 Recognition accuracy of different iterations and batch sizes

[0097]

[0098] Table 3 Recognition accuracy under different regularization parameters

[0099]

[0100] Step 4: Use the mature network model obtained in step 3 to perform target recognition and classification on the SAR image through leaves, and improve the accuracy and efficiency of synthetic aperture radar target recognition based on the target recognition and classification results.

[0101] This embodiment uses a mature network model to perform target recognition and classification in SAR images through foliage. The accuracy of target recognition and classification for the same images is compared between related networks and classic CNNs, including classic CNN models such as AlexNet, VGG16, and GoogleNet. AlexNet was the first model to use CNN for classification and has higher accuracy than traditional machine learning methods. VGG16 uses a 16-layer deep network to extract high-level features of the target, significantly improving CNN classification performance. GoogleNet increases depth and width to improve classification performance. For the classic CNN, this embodiment uses the same parameter settings as AMS-CNN. Table 4 shows the classification accuracy and parameter count of the model, verifying the effectiveness of the proposed model.

[0102] Table 4 Recognition accuracy and parameter count of different models

[0103]

[0104]

[0105] As shown in Table 4, AMS-CNN outperformed other networks in target recognition and classification accuracy, achieving a classification accuracy of 96.8%. Among classic CNNs, VGG16 achieved the highest target recognition and classification accuracy, while GoogleNet performed the worst. The classification results demonstrate that when CNNs focus solely on depth and width, deep networks are more effective than wide networks trained on a low-resolution SAR image dataset. ROC-VGG, based on VGG11, first used a CNN to classify SAR images through foliage, achieving higher accuracy than traditional machine learning methods. Compared to classic CNNs, ROC-VGG performed better, achieving target recognition and classification accuracy slightly higher than that of VGG16, demonstrating that the use of ROC can effectively improve accuracy. Unlike ROC-VGG, which is based on VGG and uses the receiver operating characteristic (ROC) to determine final accuracy, our invention proposes a different framework that provides classification support to CNNs through an attention mechanism and multiple streams. This improves the accuracy and efficiency of synthetic aperture radar target recognition based on the target recognition and classification results.

[0106] In addition, comparing the number of parameters of all CNN models, it can be seen that the number of parameters of VGG16 and GoogleNet reached 15.2×10 6 and 9.2×10 6 , because both models improve classification accuracy by increasing the number of parameters. ROC-VGG is based on VGG11, belongs to the VGG series, and also has many parameters. The AMS-CNN proposed in this paper has fewer parameters than ROC-VGG, only 0.17×10 6 , which shows that AMS-CNN is lightweight and easy to deploy on devices. This is because the constructed model adopts multi-stream and attention modules to extract important and different information, which can improve the classification accuracy while reducing the number of model parameters.

[0107] Step 5: Perform performance evaluation and testing on the mature network model in step 4, and ensure the accuracy and effectiveness of the mature network model when applied to synthetic aperture radar target recognition based on the verification and evaluation results.

[0108] To evaluate the robustness of the proposed AMS-CNN network model, this example applied five types of uniform noise of varying severity to SAR images penetrating foliage. The model's target recognition and classification performance under these conditions was analyzed and compared with three classic CNN models: AlexNet, GoogleNet, and VGG16. Table 5 shows the target recognition and classification accuracy of the different models under varying noise intensities.

[0109] Table 5 Model classification accuracy with different noise intensities

[0110]

[0111] The results show that model classification accuracy decreases as noise levels increase, but the degree of decrease varies across the four models, with AMS-CNN experiencing the least decrease and VGG16 experiencing the most. VGG16 exhibits poor robustness, primarily due to its deep network structure, which aims to achieve better classification accuracy in clean images. AMS-CNN, on the other hand, achieves better classification accuracy in both clean and noisy images. This is because the proposed AMS-CNN is designed to extract key features and multiple features for image classification, making it more robust to noisy images. These validation and evaluation results ensure the accuracy and effectiveness of mature network models when applied to synthetic aperture radar target recognition.

[0112] This paper proposes a lightweight AMS-CNN enhanced by an attention module and multi-stream. It extends the successful application of lightweight CNN in natural images to low-resolution SAR images. It also uses DCGAN to generate a dataset of SAR images through foliage. The model is optimized by analyzing the classification accuracy of the model under different parameters. Finally, the robustness of the model is tested by comparing the AMS-CNN with the previous model and using noise of varying sizes. The results show that the AMS-CNN performs well with fewer parameters and has better robustness. This paper lays a solid foundation for future research on SAR image classification.

[0113] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A synthetic aperture radar target recognition method based on multi-stream attention, characterized by: The following steps are included: Step 1: Construct an attention-based multi-stream convolutional neural network (AMS-CNN). AMS-CNN extracts key and diverse features to classify SAR images. The AMS-CNN consists of three parts: The first part uses three convolution blocks and three pooling layers to extract low-level features. The pooling layer is configured as maximum pooling to compress information and reduce computational overhead. Each convolution block consists of two convolution layers with different numbers of channels, and the pooling layer is located after the convolution block. At the same time, the first part also introduces the channel attention module and the spatial attention module. The channel attention module is applied after the first pooling layer and obtains comprehensive channel information through maximum pooling and average pooling combined with multi-layer perceptron MLP. The spatial attention module focuses on key position information by fusing spatial features and is applied after the second pooling layer to achieve the key extraction of low-level and high-level features. The second part adopts a three-stream structure to extract the maximum, average and median features respectively; The first channel is processed by group averaging after convolution blocks and global maximum pooling to extract the average features of the data; The second channel is fully connected after convolution block and global maximum pooling to extract the maximum value feature of the data; The third channel is processed by group median after convolution block and global maximum pooling to extract the median feature of the data; The maximum flow extracts important local features, the average flow and the median flow are used to capture global and intermediate features respectively. Through the combination of these three flows, the model can extract the diversity of image features and enhance classification capabilities. The third part performs image classification by combining the maximum, average and median features extracted from the three streams; LeakyReLU is used as the activation function in AMS-CNN, and the regularized cross entropy loss function is used as the loss function L(W): Where W is the parameter in the network, y i is the true label of the sample, is the predicted label of the sample, L is the total number of layers, l is the number of layers, W [l] is the parameter of the lth layer, λ is the regularization parameter; n is the number of samples; Step 2: Use Generative Adversarial Network (GAN) to enhance the dataset and improve its quality; The CARABAS-Ⅱ dataset was used as a sample, and high-quality low-resolution synthetic SAR images were generated through DCGAN to increase the amount of data used for model learning, thereby significantly improving the generalization ability of the model. The IS and FID methods were used to evaluate the high-quality synthetic SAR images. Step 3: Using the high-quality synthetic SAR image generated in step 2, perform parameter analysis and optimization on the AMS-CNN model constructed in step 1 to obtain a mature network model; Step 4: Use the mature network model obtained in step 3 to perform target recognition and classification on the SAR image through leaves, and improve the accuracy and efficiency of synthetic aperture radar target recognition based on the target recognition and classification results.

2. The method for synthetic aperture radar target recognition based on multi-stream attention according to claim 1, wherein: In step 3, we used high-quality synthetic SAR images to analyze and optimize the parameters of the AMS-CNN model, focusing on the effects of batch size, number of iterations, and regularization parameter settings on classification performance, and obtained a mature network model. Effective use of large batches can significantly improve model performance. However, when the CNN model uses stochastic gradient descent (SGD) to update the model, if the batch size is too large, it will affect its generalization performance. Therefore, it is necessary to optimize the batch size suitable for the model. The number of times the model parameters are updated is affected by the number of iterations. If the number is too low, the model cannot fully learn the image features for classification. L2 regularization is used to reduce overfitting. The regularized loss function is an effective way to reduce the risk of model overfitting, so the number of model iterations and regularization parameters need to be optimized.

3. The method for synthetic aperture radar target recognition based on multi-stream attention according to claim 1 or 2, characterized in that: It also includes step five: performing performance evaluation and testing on the mature network model obtained in step four, verifying and evaluating the characteristics and accuracy of the mature network model, and ensuring the accuracy and effect of the mature network model when applied to synthetic aperture radar target recognition based on the verification and evaluation results.

4. The method for synthetic aperture radar target recognition based on multi-stream attention according to claim 3, wherein: In step five, A mature network model is used to perform target recognition and classification in SAR images that penetrate foliage. Five uniform noises of different intensities are added to the SAR images that penetrate foliage. The target recognition and classification performance of the model is analyzed and compared with three classic CNNs, AlexNet, GoogleNet, and VGG16, under the same conditions. The characteristics and accuracy of the mature network model are evaluated and analyzed. The accuracy and effectiveness of the mature network model when applied to synthetic aperture radar target recognition are ensured based on the verification and evaluation results.

Images