SAR image vehicle target recognition method based on hybrid attention mechanism

Abstract

The application discloses a SAR image vehicle target recognition method based on a mixed attention mechanism, S1, a data set: an MSTAR data set is used; S2, a mixed attention module design: according to actual requirements and scene characteristics, a suitable mixed attention module is designed; S3, a model architecture design: an improved network MA-MobileNetV2 based on the mixed attention mechanism is designed; S4, model training: the model is trained using the MSTAR data set; S5, model evaluation and optimization: the model is evaluated and optimized through comparative experiments and performance evaluation indexes; S6, implementation and deployment: the trained model is applied to a complex SAR image vehicle target recognition task in an actual scene, and an image processing library or self-defined software can be used to realize deployment of the model. The application relates to the technical field of computer vision, and the technical scheme points are as follows: the SAR image vehicle target recognition method based on the mixed attention mechanism can improve recognition accuracy, enhance recognition stability, adapt to complex scenes and reduce computational complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, more particularly, it relates to a SAR image vehicle target recognition method based on a hybrid attention mechanism. BACKGROUND

[0002] SAR image target recognition is a follow-up task of SAR image target detection and is a key technology for realizing intelligent interpretation of remote sensing images.

[0003] Many template-based and machine learning-based SAR image target recognition algorithms have been proposed and have achieved certain results. However, these traditional SAR image target recognition algorithms have the following shortcomings: (1) insufficient feature extraction, traditional SAR image target recognition algorithms usually use hand-designed feature extraction methods, which can only extract local features of images and cannot consider the features of targets globally, resulting in insufficient extracted features. (2) Feature redundancy, since traditional SAR image target recognition algorithms use hand-designed feature extraction methods, there are often a large number of redundant features, which not only reduce the recognition accuracy, but also increase the computational complexity. (3) Difficulty in feature selection, since there are a large number of redundant features in traditional SAR image target recognition algorithms, feature selection is needed to reduce the number of features. However, due to the complex interaction relationship between features, it is often difficult to select the optimal feature subset. (4) Difficulty in feature combination, traditional SAR image target recognition algorithms usually use shallow model-based classifiers, which can only handle simple linear feature combinations and cannot handle complex nonlinear feature combinations, resulting in low recognition accuracy. These shortcomings have brought great challenges to the practical application of SAR image target recognition.

[0004] The convolutional neural network model can automatically extract important features of different targets, extract target features through a convolutional neural network model with strong feature expression capability, and significantly reduce the workload in the SAR image target recognition process. At the same time, it can avoid the limitations of manually designed target features and significantly improve the recognition ability of SAR image targets. At the same time, the application of convolutional neural networks in the field of SAR image target recognition has become a research hotspot.

[0005] In deep learning, models typically need to process high-dimensional, complex input data. This input data contains a wealth of information, but not all of it contributes to the model's output. Therefore, when processing input data, models need to focus more on task-relevant parts to improve accuracy and efficiency. Attention mechanisms are a method of weighting input data, highlighting important target features and allowing the model to better focus on target-related parts of the image, thus improving accuracy. Researchers have proposed several new attention mechanism modules, mainly categorized into three types: spatial attention, channel attention, and spatial-channel coordinated attention. Meanwhile, in the field of SAR image target recognition, some researchers are also working to apply attention mechanisms to improve network performance. Zhang et al. proposed an effective lightweight attention mechanism convolutional neural network (AM-CNN) model for automatic SAR target recognition, which shows significant advantages in performance and efficiency compared to traditional convolutional neural networks and state-of-the-art methods. Li et al. proposed a Fully Convolutional Attention Block (FCAB) that can be combined with convolutional neural networks to refine important features and suppress unnecessary features in synthetic aperture radar (SAR) images, resulting in significant performance gains for SAR recognition. Wang et al. proposed a SAR image target recognition method (GoogLeNet-APNB-ISEB) based on the GoogLeNet structure, combining an asymmetric pyramid non-local block (APNB) and SENet. SENet allows for channel dependence based on feature fusion at different scales, improving recognition accuracy. Xu et al. proposed a Multi-Scale Capsule Network with Coordinate Attention (CA-MCN) that deploys multi-scale feature extractors and coordinated attention, extracting robust features through multiple dilated convolutional layers.

[0006] However, the above methods only consider spatial and channel information, or only spatial and channel coordination information, for improving the attention part of convolutional neural networks, without considering the comprehensive weighted sum of spatial information, channel information, and spatial and channel coordination information. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems and provide a SAR image vehicle target recognition method based on a hybrid attention mechanism. A SAR image vehicle target recognition network based on a hybrid attention (MA) mechanism (Mixed Attention MobileNetV2, MA-MobileNetV2) is proposed, introducing a hybrid attention mechanism into the MobileNetV2 network. This hybrid attention mechanism fully considers the calculation of spatial attention (SPA), channel attention (CHA), and coordinated attention (CA), complementarily weighting the input feature map to enhance the representation of the region of interest. The algorithm in this chapter considers the calculation of spatial attention, channel attention, and coordinated attention in deep neural network SAR image vehicle target recognition, effectively improving the accuracy of SAR image vehicle target recognition. Since the SAR image vehicle target dataset in real-world scenarios is incomplete, experiments were conducted on the MSTAR dataset. First, the target detection performance of DC-YOLOX was tested on MSTAR simulated SAR images in complex scenes. Then, the recognition performance of MA-MobileNetV2 was verified on the MSTAR slice dataset. The results show that MA-MobileNetV2 has superior recognition performance, achieving an average recognition accuracy of 99.85% for 10 target categories. The average recognition accuracy is 3.1% higher than that of the unmodified MobileNetV2 network, and also outperforms recently reported attention-related improved SAR image vehicle target recognition algorithms.

[0008] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a SAR image vehicle target recognition method based on a hybrid attention mechanism, comprising the following steps:

[0009] S1. Dataset: The MSTAR dataset is used.

[0010] S2. Hybrid Attention Module Design: Design a suitable hybrid attention module based on actual needs and scenario characteristics;

[0011] S3. Model Architecture Design: Design an improved network MA-MobileNetV2 based on a hybrid attention mechanism;

[0012] S4. Model Training: Train the model using the MSTAR dataset;

[0013] S5. Model Evaluation and Optimization: The model is evaluated and optimized by comparing experiments and performance evaluation metrics.

[0014] S6. Implementation and Deployment: Apply the trained model to complex SAR image vehicle target recognition tasks in real-world scenarios. The model can be deployed using image processing libraries or custom software.

[0015] The present invention is further configured such that: in step S3, the MA-MobileNetV2 introduces hybrid attention into the MobileNetV2 network, fully considering the comprehensive weighting of spatial attention, channel attention, and spatial and channel coordinated attention on the input feature map, thereby enhancing the representation of the features of the region of interest.

[0016] The present invention is further configured such that, in step S2, the hybrid attention module includes spatial attention, channel attention, and a spatial and channel coordinated attention mechanism.

[0017] The present invention is further configured such that: in step S5, the evaluation metrics include precision and recall, and the formulas for calculating precision and recall are as follows:

[0018]

[0019]

[0020] In the formula: TP is the number of correctly predicted targets in the recognition result, FP is the number of incorrectly predicted targets, NP is the number of true targets, P is the recognition precision, and R is the recall rate.

[0021] In summary, the present invention has the following beneficial effects:

[0022] Improving recognition accuracy: By introducing a hybrid attention mechanism, the model can automatically suppress unimportant regions in the image and focus attention on useful regions, thereby improving the accuracy of vehicle target recognition. Compared with traditional methods, this technique can significantly improve the recognition accuracy in SAR image vehicle target recognition tasks.

[0023] Enhancing Recognition Stability: The hybrid attention mechanism module comprehensively considers spatial and channel attention, as well as coordinated spatial and channel attention, to weight the input feature map, thereby improving model stability. By reducing the volatility of test loss, this technical solution can improve the model's stability in complex scenarios.

[0024] Adapting to Complex Scenes: SAR images in complex scenes often suffer from strong noise, interference, and occlusion, posing challenges to vehicle target recognition. A hybrid attention mechanism-based approach can automatically focus on the region of interest and weight important features to adapt to the characteristics of SAR images in complex scenes, thereby improving the ability to identify vehicle targets.

[0025] Reduced computational complexity: Hybrid attention mechanisms can automatically reduce model complexity and improve model performance. By focusing only on important regions in the image, this technique reduces computational load and improves operational efficiency. Attached Figure Description

[0026] Figure 1 This relates to the SAR vehicle target detection and recognition technology in complex scenarios in this embodiment of the invention;

[0027] Figure 2 This is a training loss diagram of the MA-MobileNetV2 network and the CHA ablation network in an embodiment of the present invention;

[0028] Figure 3 This is a statistical chart showing the recognition accuracy of the MA-MobileNetV2 network and the CHA module ablation network in this embodiment of the invention.

[0029] Figure 4 This is a statistical chart showing the identification recall rate of the MA-MobileNetV2 network and the CHA module ablation network in this embodiment of the invention;

[0030] Figure 5 This is a training loss diagram of the MA-MobileNetV2 network and the SPA module ablation network in an embodiment of the present invention;

[0031] Figure 6 This is a statistical chart showing the recognition accuracy of the MA-MobileNetV2 network and the SPA module ablation network in this embodiment of the invention;

[0032] Figure 7 This is a statistical chart showing the identification recall rate of the MA-MobileNetV2 network and the SPA module ablation network in this embodiment of the invention;

[0033] Figure 8 This is a training loss diagram of the MA-MobileNetV2 network and the CA module ablation network in an embodiment of the present invention;

[0034] Figure 9 This is a statistical chart showing the recognition accuracy of the MA-MobileNetV2 network and the CA module ablation network in this embodiment of the invention.

[0035] Figure 10 This is a statistical chart of the identification recall rate of the MA-MobileNetV2 network and the CA module ablation network in this embodiment of the invention;

[0036] Figure 11 This is a structural diagram of the Bottleneck Residual Block in an embodiment of the present invention;

[0037] Figure 12 This is a ReLU6 function graph in an embodiment of the present invention;

[0038] Figure 13 This is a structural diagram of the channel attention module in an embodiment of the present invention;

[0039] Figure 14 This is a structural diagram of the spatial attention module in an embodiment of the present invention;

[0040] Figure 15 This is a structural diagram of the spatial and channel coordination attention module in an embodiment of the present invention;

[0041] Figure 16 This is a structural diagram of the hybrid attention module in an embodiment of the present invention;

[0042] Figure 17 This is a flowchart of the MA-MobileNetV2 identification process in an embodiment of the present invention;

[0043] Figure 18 This is a training loss diagram of the MA-MobileNetV2 network and the MobileNetV2 network in an embodiment of the present invention;

[0044] Figure 19 This is a statistical chart showing the recognition accuracy of the MA-MobileNetV2 network and the MobileNetV2 network in this embodiment of the invention;

[0045] Figure 20 This is a statistical chart showing the identification recall rate of the MA-MobileNetV2 network and the MobileNetV2 network in this embodiment of the invention. Detailed Implementation

[0046] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be described in further detail below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0047] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the embodiments.

[0048] Example:

[0049] like Figures 1-20 As shown, the SAR image vehicle target recognition method based on the hybrid attention mechanism includes the following steps:

[0050] S1. Dataset: The MSTAR dataset is used.

[0051] S2. Hybrid Attention Module Design: Design a suitable hybrid attention module based on actual needs and scenario characteristics;

[0052] In this embodiment, to demonstrate the role of the CHA module, SPA module, and CA module in the hybrid attention mechanism, the present invention conducts an ablation experiment on the hybrid attention mechanism network to prove the necessity of the existence of the CHA module, SPA module, and CA module.

[0053] CHA module ablation

[0054] This invention trains a MobileNetV2 network with only the SPA and CA modules added, and its network structure is shown in Table 1. The recognition performance of the ablation network is then compared and analyzed with that of the hybrid attention network.

[0055] The training loss of MA-MobileNetV2 network and CHA ablation network on the MSTAR dataset is as follows: Figure 2 As shown, the test loss fluctuation of the CHA module ablation network is significantly stronger than that of the MA-MobileNetV2 network, which indicates that the CHA module plays a certain role in improving the performance and stability of the model.

[0056] Tables 2 and 9 present the recognition confusion matrices of the CHA module ablation network and the MA-MobileNetV2 network on MSTAR data, respectively. It can be observed that the misclassification rate of the CHA module ablation network is significantly higher than that of the MA-MobileNetV2 network. The statistical results of the accuracy of the MA-MobileNetV2 network and the CHA module ablation network for various targets on MSTAR are as follows: Figure 3 As shown, compared to the MA-MobileNetV2 network, the CHA module ablation network exhibits a significant decrease in the SAR image vehicle target recognition accuracy. The average recognition accuracy of the MA-MobileNetV2 network is 99.85%, while that of the CHA module ablation network is 99.56%, representing a 0.29 percentage point decrease in average recognition accuracy compared to the MA-MobileNetV2 network. The MSTAR recall rates for various target types of the MA-MobileNetV2 network and the CHA module ablation network are shown below. Figure 4As shown, it can be observed that the recall rate of vehicle target recognition in SAR images is significantly lower than that of the MA-MobileNetV2 network. The average recall rate of the MA-MobileNetV2 network is 99.85%, while that of the CHA module ablation network is 99.60%. The average recall rate of the CHA module ablation network is 0.25 percentage points lower than that of the MA-MobileNetV2 network.

[0057] In summary, the CHA module has a significant effect on improving the model's recognition accuracy, recall, and stability.

[0058] Table 1 Ablation Network Structure of CHA Module

[0059]

[0060] Table 2 Identification Confusion Matrix of CHA Module Ablation Network

[0061]

[0062] SPA module ablation

[0063] This invention trains a MobileNetV2 network with only the CHA and CA modules added, and its network structure is shown in Table 3. The recognition performance of the ablation network is then compared and analyzed with that of the hybrid attention network.

[0064] Table 3 Ablation Network Structure of SPA Module

[0065]

[0066] Table 4. Identification Confusion Matrix of SPA Module Ablation Network

[0067]

[0068]

[0069] The training loss of MA-MobileNetV2 network and SPA module ablation network on the MSTAR dataset is as follows: Figure 5 As shown, the test loss fluctuation of the SPA module ablation network is slightly stronger than that of the MA-MobileNetV2 network, which indicates that the SPA module has a certain effect on improving model stability.

[0070] Tables 4 and 9 present the recognition confusion matrices of the SPA module ablation network and the MA-MobileNetV2 network on MSTAR data, respectively. It can be observed that the misclassification rate of the SPA module ablation network is significantly higher than that of the MA-MobileNetV2 network. The MSTAR target accuracy statistics for the MA-MobileNetV2 network and the SPA module ablation network are as follows: Figure 6 As shown, compared to the MA-MobileNetV2 network, the SPA module ablation network exhibits a significant decrease in the SAR image vehicle target recognition accuracy. The average recognition accuracy of the MA-MobileNetV2 network is 99.85%, while that of the SPA module ablation network is 99.49%, representing a 0.36 percentage point decrease in average recognition accuracy compared to the MA-MobileNetV2 network. The MSTAR recall rates for various target types of the MA-MobileNetV2 network and the SPA module ablation network are shown below. Figure 7 As shown, it can be observed that the recall rate of SAR image vehicle target recognition of the SPA module ablation network is significantly lower than that of the MA-MobileNetV2 network. The average recognition recall rate of the MA-MobileNetV2 network is 99.85%, while that of the SPA module ablation network is 99.54%. The average recognition recall rate of the SPA module ablation network is 0.31 percentage points lower than that of the MA-MobileNetV2 network.

[0071] In summary, the SPA module has a significant effect on improving the model's recognition accuracy and recall, and also plays a certain role in improving the model's stability.

[0072] CA module ablation

[0073] This invention trains a MobileNetV2 network with only the CHA and SPA modules added, and its network structure is shown in Table 5. The recognition performance of the ablation network is then compared and analyzed with that of the hybrid attention network.

[0074] Table 5. Ablation Network Structure of CA Module

[0075]

[0076]

[0077] The training loss of MA-MobileNetV2 network and CA module ablation network on the MSTAR dataset is as follows: Figure 8As shown, the test loss fluctuation of the CA module ablation network is slightly stronger than that of the MA-MobileNetV2 network, indicating that the SPA module plays a certain role in improving model stability. Tables 6 and 9 present the recognition confusion matrices of the CA module ablation network and the MA-MobileNetV2 network on the MSTAR data, respectively. It can be found that the misclassification rate of the CA module ablation network is significantly higher than that of the MA-MobileNetV2 network.

[0078] Table 6. Identification Confusion Matrix of CA Module Ablation Network

[0079]

[0080]

[0081] Statistics on the accuracy of various targets in MSTAR for MA-MobileNetV2 network and CA module ablation network are as follows: Figure 9 As shown, compared to the MA-MobileNetV2 network, the CA module ablation network exhibits a significant decrease in the SAR image vehicle target recognition accuracy. The average recognition accuracy of the MA-MobileNetV2 network is 99.85%, while that of the CA module ablation network is 99.44%. The SPA module ablation network shows a 0.41 percentage point decrease in average recognition accuracy compared to the MA-MobileNetV2 network. The MSTAR recall rates for various target types of the MA-MobileNetV2 network and the CA module ablation network are shown below. Figure 10 As shown, it can be observed that the CA module ablation network has a significant decrease in the SAR image vehicle target recognition recall rate compared to the MA-MobileNetV2 network. The average recognition recall rate of the MA-MobileNetV2 network is 99.85%, while the average recognition recall rate of the CA module ablation network is 99.47%, which is 0.38 percentage points lower than that of the MA-MobileNetV2 network.

[0082] S3. Model Architecture Design: Design an improved network MA-MobileNetV2 based on a hybrid attention mechanism;

[0083] In this embodiment, the MobileNetV2 model is a lightweight deep neural network proposed by Google. It can significantly reduce the number of computations and memory required while maintaining the same accuracy. A very important feature is the use of the Bottleneck Residual Block (Bottleneck), which consists of inverted residuals and linear activation functions. The entire MobileNetV2 is basically composed of the Bottleneck structure. The structure of MobileNetV2 is shown in Table 7.

[0084] Table 7 MobileNetV2 Structure Table

[0085]

[0086]

[0087] The structure of a bottleneck is as follows: Figure 11 As shown, Bottleneck consists of three parts: expansion convolution, depthwise convolution, and projection convolution. The entire structure uses ReLU6 activation function, as illustrated in the diagram below. Figure 12 As shown.

[0088] Dimensionality-increasing convolution uses a 1x1 kernel to increase the number of channels in the input feature map, aiming to enhance the ability of deep convolution to extract effective information. By increasing the number of channels, the model can learn more feature representations, thereby improving the model's expressive power. Deep convolution only performs convolution on each input channel separately, using fewer parameters for computation, thus requiring fewer parameters than traditional convolution. Dimensionality-reducing convolution, on the other hand, uses a 1x1 kernel, with fewer output channels than input channels, thus limiting the model size. Dimensionality-reducing convolution ensures that the number of channels does not increase excessively, achieving the goal of reducing model parameters. The formula for calculating the ReLU6 activation function is shown below:

[0089] f(x)=ReLU(x)=min(max(x,0),6)

[0090] Channel attention module

[0091] Channel attention mechanisms can improve model recognition accuracy and reduce interference from channels that are less critical to the recognition task by selecting channels that are more important to the task. The structure of a channel attention module is as follows: Figure 13As shown, the channel attention mechanism first compresses the input feature map in the spatial dimension using mean pooling and max pooling, resulting in two different vectors of size C×1×1. and These represent the average pooling feature and the max pooling feature, respectively. Then... and Input into a shared network to obtain channel attention mapping and The shared network consists of a multilayer perceptron (MLP) with one hidden layer. When applying the shared network... and Next, the output feature vectors are merged using element-wise summation, and then the final channel attention weights M are obtained by passing them through the sigmoid function. c (F). The calculation formula is as follows:

[0092]

[0093]

[0094]

[0095] Where F represents the input feature, M c (F) represents the channel attention weight, AvgPool c MaxPool indicates that average pooling is performed in the channel dimension. c This indicates that max pooling is performed in the channel dimension.

[0096] Spatial attention module

[0097] Spatial attention mechanisms can control the attention weights at different locations, thus focusing more effort on important regions and locations with higher information content, thereby helping the model better capture important information in the image. The structure of the channel attention module is as follows: Figure 14 As shown, to compute spatial attention, average pooling and max pooling operations are first applied along the channel axis to obtain two different feature descriptions of size 1×H×W. and These features are then concatenated to generate a 2×H×W feature description. This 2×H×W feature description is then convolved through a standard convolutional layer, and the final spatial attention weights M are obtained after passing through a sigmoid function. s (F). The calculation formula is as follows:

[0098]

[0099]

[0100]

[0101] Where F represents the input feature, M s (F) represents the spatial attention weights, f represents the standard convolution operation, and AvgPool s MaxPool represents the average pooling operation performed in the spatial dimension. c This indicates that max pooling is performed in the spatial dimension.

[0102] Spatial and Channel Coordination Attention Module

[0103] Spatial and channel coordinated attention considers not only channel information but also orientation-related positional information. It learns the dependencies between features based on spatial relationships, thus better capturing the relationships between different regions in an image. The structure of the spatial and channel coordinated attention module is as follows: Figure 15 As shown, to compute spatial and channel coordinated attention, the input feature map of size C×H×W is first average pooled along the X-axis and Y-axis directions of the space, respectively, to obtain a feature map of size C×1×W. and the size is C×H×1 These features are then concatenated to generate a feature description of size C×1×(H+W). This C×1×(H+W) feature description is convolved through a standard convolutional layer. The result is then decomposed into C×1×W and C×H×1 feature descriptions, which are then passed through a sigmoid function to obtain the final spatial attention weights M. h (F) and M w (F). The calculation formula is as follows:

[0104]

[0105]

[0106]

[0107] Where F represents the input feature, M h (F) represents the attention weight in the spatial X-axis direction, M w (F) represents the attention weights along the Y-axis in space, f represents the standard convolution operation, and AvgPool h AvgPool represents the average pooling operation performed along the X-axis in space. w This indicates that average pooling is performed in the Y-axis direction of the space.

[0108] Hybrid Attention Convolutional Neural Network (MA-MobileNetV2)

[0109] Channel attention can automatically weight the channel dimension, spatial attention can automatically weight the spatial dimension, and the spatial and channel coordinated attention module can weight the edge distribution of space on the channel, such as... Figure 13 , Figure 14 , Figure 15 As shown, these three weighting methods are complementary in different dimensions, and can complementarily weight the feature maps to improve the recognition accuracy of the network.

[0110] The structure of the hybrid attention module is as follows: Figure 16 As shown, to calculate the hybrid attention, the input features are weighted using CHA, SPA, and CA methods respectively, ultimately yielding a hybrid attention-weighted output feature map. The calculation of features and weights follows a broadcast mechanism. The calculation formula is shown below:

[0111] F'=F*M c (F)*M s (F)*(M h (F)*M w (F))

[0112] Where F represents the input feature, M c (F) represents the channel attention weight, M s (F) represents the spatial attention weights, M h (F) represents the attention weight in the spatial X-axis direction, M w (F) represents the attention weight in the Y-axis direction of space, and F' represents the feature after being weighted by the hybrid attention module.

[0113] Table 8 Structure of the MA-MobileNetV2 Algorithm

[0114]

[0115]

[0116] The MA-MobileNetV2 network introduces hybrid attention into the MobileNetV2 network, fully considering the combined weighting of spatial attention, channel attention, and spatial-channel coordinated attention on the input feature map to enhance the representation of features in the region of interest and improve network performance. The structure of MA-MobileNetV2 is shown in Table 8, with the bolded sections indicating the locations where the hybrid attention module is introduced. The MA-MobileNetV2 recognition process is as follows: Figure 17 As shown, the input image is first processed through a backbone network to obtain a feature map. Then, a hybrid attention module is used to weight the obtained feature map. The weighted feature map is then input into a pooling layer and a fully connected layer to obtain the final recognition result.

[0117] S4. Model Training: Train the model using the MSTAR dataset;

[0118] S5. Model Evaluation and Optimization: The model is evaluated and optimized by comparing experiments and performance evaluation metrics.

[0119] In this embodiment, to quantitatively compare the recognition performance of the models, precision and recall are used as performance evaluation metrics for the recognition results. The formulas for calculating precision and recall are as follows:

[0120]

[0121]

[0122] In the formula: TP is the number of correctly predicted targets in the recognition result, FP is the number of incorrectly predicted targets, NP is the number of true targets, P is the recognition precision, and R is the recall rate.

[0123] S6. Implementation and Deployment: Apply the trained model to complex SAR image vehicle target recognition tasks in real-world scenarios. The model can be deployed using image processing libraries or custom software.

[0124] Performance comparison between MA-MobileNetV2 network and MobileNetV2 network

[0125] The MA-MobileNetV2 network is an improved network that introduces the Hybrid Attention Module (MA) on top of the MobileNetV2 network.

[0126] The training loss of MA-MobileNetV2 network and MobileNetV2 network on the MSTAR dataset is as follows: Figure 18 As shown, the model was trained with the backbone network frozen for the first 20 epochs. The MA-MobileNetV2 network primarily adjusted the parameters of the hybrid attention module and fully connected layers during the first 20 epochs, while the MobileNetV2 network primarily adjusted the parameters of the fully connected layers. It can be observed that the test loss of the MobileNetV2 network fluctuated significantly during the first 20 epochs of training, while the test loss fluctuation of the MA-MobileNetV2 network was significantly reduced compared to the MobileNetV2 network. Adding a hybrid attention module to the MobileNetV2 network can automatically suppress unimportant regions in the image and focus attention only on useful regions, thereby reducing model complexity and improving model performance and stability.

[0127] Tables 9 and 10 present the recognition confusion matrices of the MA-MobileNetV2 network and the MobileNetV2 network on MSTAR data, respectively. It can be observed that the MA-MobileNetV2 network significantly improves recognition accuracy compared to the MobileNetV2 network.

[0128] Table 9. Identification and Confusion Matrix of MA-MobileNetV2 Network

[0129]

[0130]

[0131] Table 10. Identification and Confusion Matrix of MobileNetV2 Network

[0132]

[0133] Statistics on the accuracy of various targets in MA-MobileNetV2 networks and MobileNetV2 networks' MSTAR targets are as follows: Figure 19 As shown, compared to the MobileNetV2 network, the MA-MobileNetV2 network significantly improves the accuracy of vehicle target recognition in SAR images. The average recognition accuracy of the MA-MobileNetV2 network is 99.85%, while that of the MobileNetV2 network is 96.75%, representing a 3.1 percentage point improvement in average recognition accuracy. The MSTAR recall rates for various target types for both the MA-MobileNetV2 and MobileNetV2 networks are shown below. Figure 20 As shown, it can be found that the MA-MobileNetV2 network has a significantly improved recall rate for SAR image vehicle target recognition compared to the MobileNetV2 network. The average recall rate of the MA-MobileNetV2 network is 99.85%, while the average recall rate of the MobileNetV2 network is 96.92%, representing an improvement of 2.93 percentage points in the average recall rate of the MA-MobileNetV2 network.

[0134] Performance comparison of MA-MobileNetV2 network and related methods

[0135] Table 11 shows a performance comparison between MA-MobileNetV2 and the latest attention-correlation-based improved SAR image vehicle target recognition algorithm. It can be concluded that the average recognition accuracy of the MA-MobileNetV2 network is significantly improved compared to the latest attention-correlation-based improved SAR image vehicle target recognition algorithm. This demonstrates the superior target recognition performance of the MA-MobileNetV2 network.

[0136] Table 11 Performance Comparison of MA-MobileNetV2 and State-of-the-Art Attention-Related Improved Recognition Algorithms

[0137]

[0138] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A SAR image vehicle target recognition method based on a hybrid attention mechanism, characterized in that, The method comprises the following steps: S1, dataset: using MSTAR dataset; S2, hybrid attention module design: according to the actual demand and scene characteristics, design appropriate hybrid attention module; To calculate the spatial and channel coordination attention, firstly, the input feature map with size CxHxW is respectively averaged-pooled along the X-axis direction and the Y-axis direction of the space to obtain the feature description with size Cx1xW and CxHx1 , and then they are connected to generate a feature description with size Cx1x(H+W); the Cx1x(H+W) feature description is convolved through a standard convolution layer, and then the obtained result is decomposed into Cx1xW and CxHx1 feature descriptions, and then the feature descriptions are respectively subjected to sigmoid functions to obtain the final spatial attention weights and ; the calculation formula is as follows: ； ； ； wherein, denotes an input feature, denotes an attention weight in the X-axis direction of the space, denotes an attention weight in the Y-axis direction of the space, denotes a standard convolution operation, denotes an average pooling operation in the X-axis direction of the space, denotes an average pooling operation in the Y-axis direction of the space S3, model architecture design: design improved network MA-MobileNetV2 based on hybrid attention mechanism; Channel attention can automatically weight the channel dimension, spatial attention can automatically weight the spatial dimension, and spatial and channel coordination attention module can weight the edge distribution of space in channel. The three weighting methods have complementarity in different dimensions, which can complementarily weight the feature map and improve the recognition accuracy of the network; In order to calculate the hybrid attention, the input features are respectively weighted by CHA, SPA and CA, and finally the output feature map weighted by hybrid attention is obtained, wherein the calculation of features and weights follows the broadcast mechanism; its calculation formula is as follows: ； wherein, represents an input feature, represents a channel attention weight, represents a spatial attention weight, represents a spatial X-axis direction attention weight, represents a spatial Y-axis direction attention weight, represents a feature weighted by the mixed attention module; S4, model training: using MSTAR dataset for model training; S5, model evaluation and optimization: through comparative experiment and performance evaluation index, the model is evaluated and optimized; S6, implementation and deployment: the trained model is applied to the complex SAR image vehicle target recognition task in the actual scene, and the deployment of the model can be realized by using image processing library or custom software.

2. The SAR image vehicle target recognition method based on the hybrid attention mechanism according to claim 1, characterized in that: In step S3, the MA-MobileNetV2 is introduced into the MobileNetV2 network, fully considering the comprehensive weighting of input feature map by spatial attention, channel attention and spatial and channel coordination attention, and enhancing the representation of region of interest features.

3. The SAR image vehicle target recognition method based on the hybrid attention mechanism according to claim 1, characterized in that: In step S2, the hybrid attention module includes spatial attention, channel attention and spatial and channel coordination attention mechanism.

4. The SAR image vehicle target identification method based on the hybrid attention mechanism according to claim 1, characterized in that: In step S5, the evaluation index includes accuracy and recall rate, and the calculation formula of the accuracy and recall rate is: ； ； In the formula: is the number of correctly predicted targets in the recognition result, is the number of incorrectly predicted targets, is the number of real targets, is the recognition accuracy, is the recall rate.

Images