SAR (Synthetic Aperture Radar) image target identification method, system and equipment based on phase-guided attention and amplitude-phase dual-branch network, and medium

Abstract

The invention provides an SAR (Synthetic Aperture Radar) image target identification method, system and device based on phase guidance attention and an amplitude-phase dual-branch network, and a medium, and the method comprises the steps: designing two enhancement strategies of phase maintenance and cooperative transformation through decoupling amplitude and phase information, and obtaining an enhanced image; designing an amplitude-phase double-branch network architecture which comprises an amplitude branch, a phase branch and a feature fusion module, and respectively acquiring an amplitude feature, a phase feature and a feature map which is used for classification and contains double information; constructing a total loss function combined with phase consistency loss and cross entropy loss to train an amplitude-phase double-branch network architecture, and obtaining trained network model parameters; and identifying the test image set based on the trained network model parameters, wherein the obtained identification result is a synthetic aperture radar image target identification result. Compared with the prior art, the method has the advantages of high recognition accuracy and the like, and can be popularized and used in the technical field of image processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a SAR image target recognition method, system, device and medium based on phase-guided attention and amplitude-phase dual-branch network. Background Art

[0002] Synthetic Aperture Radar (SAR), as an active imaging technology, has all-day, all-weather imaging capabilities and is widely used in tasks such as military reconnaissance and disaster monitoring. The complex data acquired by SAR imaging contains amplitude and phase information. Amplitude information reflects the scattering intensity of the target and plays a key role in distinguishing different types of targets with significant differences in scattering intensity. Phase information is more sensitive to subtle structural changes and geometric features of the target and plays an important role in identifying targets with complex and fine structures. However, existing research on SAR image target recognition has mostly focused on amplitude information and ignored the potential value of phase information. Therefore, it is necessary to study how to fully utilize image phase information to assist SAR image recognition.

[0003] Existing techniques for combining amplitude and phase information for image recognition often utilize phase information implicitly, lacking in-depth exploration and targeted utilization. When fusing and extracting amplitude and phase features, phase and amplitude are typically treated as homogenized, lacking specialized processing modules to fully realize their potential. Therefore, the present invention discloses a method, system, device, and medium for SAR image target recognition based on phase-guided attention and an amplitude-phase dual-branch network. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a SAR image target recognition method, system, device and medium based on phase-guided attention and amplitude-phase dual-branch network, which is used to solve the technical problem of insufficient utilization of phase information in synthetic aperture radar image target recognition.

[0005] The present invention is achieved through the following technical solutions: A SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network includes the following steps: Data enhancement: Based on the fact that complex images acquired by SAR imaging contain both amplitude and phase information, two enhancement strategies, phase preservation and collaborative transformation, are designed by decoupling amplitude and phase information to obtain enhanced images. Network construction: Design an amplitude-phase dual-branch network architecture, including an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch introduces a phase-guided attention module to extract features from amplitude information and obtain amplitude features. The phase branch extracts features from phase information and obtains phase features. The feature fusion module concatenates the extracted amplitude and phase features to generate a feature map containing dual information for classification. Training the network: Based on the enhanced image, a total loss function combining phase consistency loss and cross entropy loss is constructed to train the amplitude-phase dual-branch network architecture and obtain the trained network model parameters; Image recognition: The test image set is recognized based on the trained network model parameters, and the recognition results obtained are synthetic aperture radar image target recognition results.

[0006] Preferably, the phase-preserving enhancement strategy is to only change the amplitude spectrum while keeping the phase spectrum unchanged. The enhancement process is:

[0007] in, and are the original amplitude and the enhanced amplitude, is the exponent of the power operation; The probability of performing phase-preserving enhancement operation is to use a randomly selected The value performs a power-law transformation on the amplitude spectrum; the exponential The sampling follows a uniform distribution to control the area of ​​amplitude enhancement. When <1, the enhancement amplitude is small; When >1, the enhancement amplitude is larger.

[0008] Preferably, the collaborative transformation strategy processes amplitude and phase separately and applies a consistent spatial transformation to both. The enhancement process is:

[0009]

[0010] in, R is the rotation matrix, R -1 is its inverse transform, and are the original phase and the enhanced phase respectively; The probability of rotating the amplitude spectrum and phase spectrum at the same time, and the rotation angle is controlled in a small angle range , bilinear interpolation is used to ensure the continuity of phase information after rotation.

[0011] Preferably, the phase-guided attention module includes three dimensions: phase attention, channel attention, and spatial attention; the calculation process of the phase attention weight is:

[0012] in is the input phase information, For interpolation operations, Represents the Sigmoid activation function, which enables the model to capture features from the perspective of phase; The calculation process of channel attention weight is:

[0013] in, It is the global representation of the phase feature obtained after convolution, batch normalization, activation and adaptive average pooling. is the amplitude feature after adaptive average pooling, represents a linear transformation, Represents the ReLu activation function; The calculation process of spatial attention weight is:

[0014] in, is the input amplitude feature, Represents maximum pooling and average pooling respectively, Represents the convolution operation; Finally, the process of multi-dimensional attention fusion weighted adjustment of amplitude features is:

[0015]

[0016] in, is the amplitude feature of the input, is the total attention weight, Represents the amplitude feature after being processed by the phase-guided attention module.

[0017] Preferably, the process of constructing the phase consistency loss function is:

[0018] in, is the total number of pixels in the image, and The images are at positions The horizontal gradient and vertical gradient at ; is the edge weight parameter, which controls the strength of edge enhancement; is the edge mask at position The value at , the edge mask The specific definition is:

[0019]

[0020] in, and They are horizontal gradient and vertical gradient, represents the input phase image, and are the edge detection kernel functions in the horizontal and vertical directions, is the edge mask, which does not participate in back propagation and only serves as a weight factor; Preferably, the process of the total loss function is:

[0021] in, represents the total loss, is the cross entropy loss, is the phase consistency loss, parameter Used to adjust the weight of phase consistency loss in the total loss.

[0022] Preferably, the process of training the amplitude-phase dual-branch network model to obtain trained network model parameters is: The training set formed by the enhanced image is input into the network for training. The amplitude and phase of the enhanced image of the training set are respectively input into the corresponding branches and pass through all modules in sequence. The Adam optimization algorithm is used during training to set the data batch, training rounds, learning rate and weight parameter values; During training, the loss function is calculated through forward propagation, and the network weights are updated through back propagation until the loss function converges to obtain the trained network model parameters.

[0023] A SAR image target recognition system based on phase-guided attention and amplitude-phase dual-branch network, based on the SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network, comprising: The data enhancement module, based on the characteristic that complex images acquired by SAR imaging contain amplitude and phase information, decouples amplitude and phase information to design two enhancement strategies: phase preservation and collaborative transformation, to obtain enhanced images; The network construction module is used to design an amplitude-phase dual-branch network architecture, including an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch introduces a phase-guided attention module to extract features from amplitude information and obtain amplitude features; the phase branch extracts features from phase information and obtains phase features; and the feature fusion module concatenates the extracted amplitude and phase features to generate a feature map containing dual information for classification. The network training module is used to construct a total loss function combining phase consistency loss and cross entropy loss based on the enhanced image to train the amplitude-phase dual-branch network architecture and obtain the trained network model parameters; The image recognition module is used to identify the test image set based on the trained network model parameters, and the obtained recognition result is the synthetic aperture radar image target recognition result.

[0024] A computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the processor implements the steps of a SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network.

[0026] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention provides a SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network. First, based on the characteristic that complex images acquired by SAR imaging contain amplitude and phase information, a phase-preserving data enhancement method is proposed. By decoupling amplitude and phase information, two enhancement strategies, phase preservation and collaborative transformation, are designed to increase data diversity while avoiding the destruction of phase information.

[0027] Secondly, a dual-branch amplitude-phase network architecture is designed to process amplitude and phase information separately. The amplitude branch is responsible for extracting features such as strong scattering areas and target outlines from the amplitude information, while the phase branch is responsible for extracting features such as subtle structural changes and topology from the phase information. A feature fusion module is then used to combine the features extracted by each branch to generate a feature map containing both information for classification. During the feature extraction process, a phase-guided attention module is constructed within the amplitude branch, introducing an independent phase attention branch to achieve targeted mining of unique phase information. Phase information is also used to guide the processing of amplitude features and generate attention weights, highlighting key information in the amplitude features, suppressing irrelevant information, and enhancing the expressive power of the amplitude features.

[0028] During network training, a total loss function combining phase consistency loss and cross-entropy loss is constructed to guide the model to pay more attention to the consistency of phase information when processing images, thereby increasing the discriminability of phase features to assist in classification. Finally, based on the trained network model parameters, the test image set is recognized, and the recognition results obtained are the synthetic aperture radar image target recognition results.

[0029] The present invention designs a dedicated network structure for amplitude and phase, and introduces an attention mechanism into the processing flow of phase information, exploring the complementary potential of amplitude and phase information in target recognition, thereby solving the problem of insufficient utilization of phase information in SAR image target recognition. Compared with the existing technology, the method proposed in the present invention is different from the traditional homogenization processing mode. It designs a dedicated network structure for amplitude and phase, and introduces an attention mechanism into the processing flow of phase information, exploring the complementary potential of amplitude and phase information in target recognition. It has the advantages of high recognition accuracy and strong robustness, and can be promoted and used in the field of image processing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the present invention; Figure 2 This is a schematic diagram of the amplitude-phase dual-branch network structure constructed by the present invention; Figure 3 yes Figure 2 Schematic diagram of the structure of the mid-phase guided attention module; Figure 4 yes Figure 2 Schematic diagram of the detailed medium-amplitude dual-branch network. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to specific embodiments, which are intended to explain the present invention rather than to limit it.

[0032] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0033] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0034] The embodiment of the present invention provides a SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network, such as Figure 1 As shown, the following steps are included: Data Augmentation: Based on the fact that complex images acquired by SAR imaging contain both amplitude and phase information, a phase-preserving data augmentation method is proposed. By decoupling amplitude and phase information, two enhancement strategies, phase-preserving and collaborative transformation, are designed to increase data diversity while avoiding the destruction of phase information. The complex images acquired by SAR imaging are derived from three datasets: Moving and Stationary Target Acquisition and Recognition (MSTAR), including training and test sets.

[0035] The core of the phase-preserving strategy is that the enhancement operation only changes the amplitude spectrum, while the phase spectrum remains unchanged. The enhancement process is:

[0036] in, and are the original amplitude and the enhanced amplitude, Is the exponent of the power operation. The probability of performing phase-preserving enhancement operation is to use a randomly selected The value performs a power-law transformation on the amplitude spectrum. The sampling follows a uniform distribution to control the area of ​​amplitude enhancement. When <1, the enhancement amplitude is small; When >1, the enhancement amplitude is larger.

[0037] The collaborative transformation strategy processes amplitude and phase separately and applies a consistent spatial transformation to both. The enhancement process is:

[0038]

[0039] in, R is the rotation matrix, R -1 is its inverse transform, and are the original phase and the enhanced phase respectively. The probability of rotating the amplitude spectrum and phase spectrum at the same time, and the rotation angle is controlled in a small angle range , bilinear interpolation is used to ensure the continuity of phase information after rotation.

[0040] Network construction: Amplitude and phase information are processed separately, and a dual-branch amplitude-phase network architecture is designed, including an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch is responsible for extracting features such as strong scattering areas and target contours from the amplitude information; the phase branch is responsible for extracting features such as subtle structural changes and topology from the phase information. The feature fusion module is responsible for splicing the features extracted by each branch to generate a feature map containing dual information for classification. A phase-guided attention module is constructed in the amplitude branch, introducing an independent phase attention branch to achieve targeted mining of unique phase information. Phase information is used to guide the processing of amplitude features, generate attention weights, highlight key information in amplitude features, suppress irrelevant information, and enhance the expressiveness of amplitude features. The amplitude-phase dual-branch network consists of an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch includes a series of convolutional layers, pooling layers, batch normalization layers, and ReLu activation functions, as well as a phase-guided attention module; the phase branch includes a series of convolutional layers, batch normalization layers, and ReLu activation functions; the feature fusion module includes a splicing operation, a convolutional layer, a batch normalization layer, and a ReLu activation function; the amplitude and phase information of the training set image are input into the corresponding branches respectively.

[0041] It should be noted that, in this embodiment, in the amplitude-phase dual-branch network, the amplitude branch is composed of a convolution layer 1 with a convolution kernel size of 3×3, a normalization layer 1, a convolution layer 2, a normalization layer 2, a pooling layer 1, a convolution layer 3, a normalization layer 3, a convolution layer 4, a normalization layer 4, a pooling layer 2, a convolution layer 5, a normalization layer 5, a convolution layer 6, a normalization layer 6, a pooling layer 3, and a PGA module connected in series. The output of the amplitude branch is :

[0042]

[0043]

[0044]

[0045] in, represent the input amplitude and phase respectively, represents the phase-guided attention processing module, Represents the convolutional layer , normalization layer , ReLu activation function , Pooling layer and The output of a convolutional block.

[0046] In the amplitude-phase dual-branch network, the phase branch is composed of a convolution layer 7 with a convolution kernel size of 3×3, a normalization layer 7, a pooling layer 4, a convolution layer 8, a normalization layer 8, a pooling layer 5, a convolution layer 9, a normalization layer 9, and a pooling layer 6 connected in series. The output of the phase branch is :

[0047] in, represents the phase of the input, Represents the convolutional layer , normalization layer , ReLu activation function and pooling layers .

[0048] At the same time, in the construction of the amplitude-phase dual-branch network, the feature fusion module consists of a convolution layer 10, a normalization layer 10, a pooling layer 7, and a fully connected layer 1. The output of the recognition module is :

[0049] in Represent the outputs of the amplitude branch and the phase branch respectively, represents the concatenation of tensors, is the fully connected layer, Represents the Softmax activation function.

[0050] The phase-guided attention module includes three dimensions: phase attention, channel attention, and spatial attention. The calculation process of phase attention weight is as follows:

[0051] in is the input phase information, For interpolation operations, Represents the Sigmoid activation function, which enables the model to capture features from the perspective of phase.

[0052] The calculation process of channel attention weight is:

[0053] in, It is the global representation of the phase feature obtained after convolution, batch normalization, activation and adaptive average pooling. is the amplitude feature after adaptive average pooling, represents a linear transformation, Represents the ReLu activation function.

[0054] The calculation process of spatial attention weight is:

[0055] in, is the amplitude feature of the input, Represents maximum pooling and average pooling respectively, Represents a convolution operation.

[0056] Finally, the process of multi-dimensional attention fusion weighted adjustment of amplitude features is:

[0057]

[0058] in, is the amplitude feature of the input, is the total attention weight, Represents the amplitude feature after being processed by the phase-guided attention module.

[0059] Training the network: Based on the enhanced image, a total loss function combining phase consistency loss and cross entropy loss is constructed to train the amplitude-phase dual-branch network architecture to obtain the trained network model parameters. Among them, the combination of phase consistency loss and cross entropy loss guides the model to pay more attention to the consistency of phase information when processing images, increasing the discriminability of phase features to assist classification.

[0060] The process of training the amplitude-phase dual-branch network model and obtaining the trained network model parameters is as follows: The enhanced image dataset is input into the network for training, and the amplitude and phase of the training set image are input into the corresponding branches respectively and pass through all modules in sequence.

[0061] The Adam optimization algorithm is used in the training process to set the data batch, training rounds, learning rate and weight parameter values. Specifically, in this embodiment, the data batch can be set to 32, the training rounds can be set to 1000 rounds, the learning rate can be set to 0.0001, and the edge weight Can be set to 0.5, loss weight Can be set to 0.2.

[0062] During training, the loss function is calculated through forward propagation, and the network weights are updated through back propagation until the loss function converges to obtain the trained network model parameters.

[0063] The loss function combines phase consistency loss and cross entropy loss. The process of constructing the phase consistency loss function is:

[0064] in, is the total number of pixels in the image, and The images are at positions The horizontal and vertical gradients at . is the edge weight parameter that controls the strength of edge enhancement. is the edge mask at position The value at , the edge mask The specific definition is:

[0065]

[0066] in, and They are horizontal gradient and vertical gradient, represents the input phase image, and are the edge detection kernel functions in the horizontal and vertical directions, is the edge mask, which does not participate in back propagation and only serves as a weight factor.

[0067] The basic loss term of the phase consistency loss function is the sum of squares of phase gradients, which encourages the smoothness of the overall phase. The sum of squared gradients multiplied by the edge mask is used as an edge enhancement term to impose greater weight on gradient changes in edge areas.

[0068] The process of constructing the total loss function is:

[0069] in, represents the total loss, is the cross entropy loss, is the phase consistency loss, parameter It is used to adjust the weight of phase consistency loss in the total loss. The total loss function is Ensure that the consistency constraint of phase information is satisfied; on the other hand, use Make the model prediction results closer to the true label.

[0070] The test image set is recognized based on the trained network model parameters, and the recognition results obtained are synthetic aperture radar image target recognition results.

[0071] Example 1 A SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network in this embodiment comprises the following steps: Perform data augmentation: The MSTAR dataset (three categories) was used as training and test data, with an image size of 128×128. The original training set consisted of observations with a pitch angle of 17°, and the original test set consisted of observations with a pitch angle of 15°. A phase-preserving data augmentation method was designed, including two augmentation strategies: phase-preserving and co-transformation. The core of the phase-preserving strategy is that the augmentation operation only changes the amplitude spectrum, while the phase spectrum remains unchanged. The augmentation process is as follows:

[0072] in, and are the original amplitude and the enhanced amplitude, Is the exponent of the power operation. The probability of performing phase-preserving enhancement operation is to use a randomly selected The value performs a power-law transformation on the amplitude spectrum. The sampling follows a uniform distribution to control the area of ​​amplitude enhancement. When <1, the enhancement amplitude is small. When >1, the enhancement amplitude is larger.

[0073] The collaborative transformation strategy processes amplitude and phase separately and applies a consistent spatial transformation to both. The enhancement process is:

[0074]

[0075] in, R is the rotation matrix, R -1 is its inverse transform, and are the original phase and the enhanced phase respectively. The probability of rotating the amplitude spectrum and phase spectrum at the same time, and the rotation angle is controlled in a small angle range , bilinear interpolation is used to ensure the continuity of phase information after rotation.

[0076] In this embodiment, The enhanced image dataset is obtained after being processed by the phase-preserving data enhancement method.

[0077] Construct an amplitude-phase two-branch network for augmenting data: The amplitude-phase dual-branch network consists of an amplitude branch, a phase branch, a feature fusion module, and a loss function, wherein the amplitude branch includes a series of convolutional layers, pooling layers, batch normalization layers, and ReLu activation functions, as well as a phase-guided attention module; the phase branch includes a series of convolutional layers, batch normalization layers, and ReLu activation functions; the feature fusion module includes a splicing operation, a convolutional layer, a batch normalization layer, and a ReLu activation function; the amplitude and phase information of the training set image are input into the corresponding branches respectively.

[0078] like Figure 2 As shown, the amplitude branch and the phase branch of this embodiment are connected in parallel, and the extracted features are fused for recognition.

[0079] In the amplitude-phase dual-branch network, the amplitude branch is composed of convolution layer 1 with a convolution kernel size of 3×3, normalization layer 1, convolution layer 2, normalization layer 2, pooling layer 1, convolution layer 3, normalization layer 3, convolution layer 4, normalization layer 4, pooling layer 2, convolution layer 5, normalization layer 5, convolution layer 6, normalization layer 6, pooling layer 3, and PGA module connected in series. The output of the amplitude branch is :

[0080]

[0081]

[0082]

[0083] in, represent the input amplitude and phase respectively, represents the phase-guided attention processing module, Represents the convolutional layer , normalization layer , ReLu activation function , pooling layer and The output of the convolutional block; In the amplitude-phase dual-branch network, the phase branch is composed of a convolution layer 7 with a convolution kernel size of 3×3, a normalization layer 7, a pooling layer 4, a convolution layer 8, a normalization layer 8, a pooling layer 5, a convolution layer 9, a normalization layer 9, and a pooling layer 6 connected in series. The output of the phase branch is :

[0084] in, represents the phase of the input, Represents the convolutional layer , normalization layer , ReLu activation function and pooling layers ; At the same time, in the construction of the amplitude-phase dual-branch network, the feature fusion module and the output module are composed of a convolution layer 10, a normalization layer 10, a pooling layer 7 and a fully connected layer 1, and the output of the recognition module is :

[0085] in Represent the outputs of the amplitude branch and the phase branch respectively, represents the concatenation of tensors, is the fully connected layer, Represents the Softmax activation function.

[0086] like Figure 3 As shown, the phase-guided attention module of this embodiment is composed of three branches in parallel: phase attention, channel attention, and spatial attention; The calculation process of phase attention weight is:

[0087] in is the input phase information, For interpolation operations, Represents the Sigmoid activation function, which enables the model to capture features from the perspective of phase.

[0088] The calculation process of channel attention weight is:

[0089] in, It is the global representation of the phase feature obtained after convolution, batch normalization, activation and adaptive average pooling. is the amplitude feature after adaptive average pooling, represents a linear transformation, Represents the ReLu activation function.

[0090] The calculation process of spatial attention weight is:

[0091] in, is the amplitude feature of the input, Represents maximum pooling and average pooling respectively, Represents a convolution operation.

[0092] Finally, the process of multi-dimensional attention fusion weighted adjustment of amplitude features is:

[0093]

[0094] in, is the amplitude feature of the input, is the total attention weight, Represents the amplitude feature after being processed by the phase-guided attention module.

[0095] Construct the loss function: The total loss function is determined as follows :

[0096]

[0097]

[0098]

[0099] In this embodiment, the edge detection kernel function may adopt the Sobel operator kernel function, that is, .

[0100] Train the network: The enhanced image dataset is input into the network for training. The amplitude and phase of the training set image are input into the corresponding branches and pass through all modules in sequence. The Adam optimization algorithm is used in the training process, and the data batch is set to 32, the training round is 1000 rounds, the learning rate is 0.0001, and the edge weight is 0. is 0.5, loss weight is 0.2; During training, the loss function is calculated through forward propagation, and the network weights are updated through backpropagation until the loss function converges; Target Identification: Input the test data set into the trained network to obtain the recognition results through all modules; A SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network was completed.

[0101] To verify the effectiveness of our method for assisting target recognition using SAR image phase information, we used the MSTAR three-category dataset. The dataset used an observation pitch angle of 17° as the original training set, and an observation pitch angle of 15° as the original test set. The number of samples for each type in the dataset is summarized in Table 1.

[0102] Table 1 Number of training and testing samples in the MSTAR dataset

[0103] In order to verify the influence of each key component in the model proposed in this invention on SAR recognition performance, the effectiveness of this invention will be verified through ablation experiments, from the perspectives of single component effectiveness and multi-component synergy effect. ) and the Phase Guided Attention (PGA) module, as follows: Table 2 Single component validity verification results

[0104] From the experimental results, it can be seen that removing PGA, The model of the present invention, which consists of three components, PPDA, and PPDA, serves as the baseline model. Its overall recognition rate is 95.05%. The recognition rates for the BMP2, T72, and BTR70 models are 95.73%, 91.72%, and 97.70%, respectively. This shows that the T72 is relatively difficult to recognize, followed by the BMP2, and the BTR70 is relatively easy. As each component is gradually added, the overall recognition rate improves significantly compared to the baseline model. The details are as follows: Specifically, after adding the PGA module, the model's recognition rate for the T72 model increased from 91.72% to 96.34%, while the recognition rates for the BMP2 and BTR70 models remained largely unchanged. This indicates that the PGA module effectively enhances the model's feature extraction capabilities for models with less distinct features, such as the T72, but has limited impact on the BMP2 and BTR70 models, which have more distinct features.

[0105] Introduction When the loss function is applied, the recognition rate for the BMP2 model increases to 97.69%, the T72 to 93.69%, and the BTR70 maintains 97.70%. Experimental results show that the BMP2 and T72 models are more sensitive to phase changes, demonstrating that the loss function can improve the model's sensitivity to the detailed features of different models by constraining phase consistency. When PPDA was applied, the recognition rates for the three models reached 99.31%, 97.46%, and 99.15%, respectively, for an overall recognition rate of 98.64%, representing the most significant improvement among single components. This indicates that PPDA, by preserving phase information while expanding the data, enabled the model to learn more robust feature representations for the BMP2, T72, and BTR70 models, significantly contributing to improved model performance.

[0106] To further analyze the interaction between components, the following four groups of experiments were designed for comparison.

[0107] Table 3 Verification results of multi-component synergistic effect

[0108] From the experimental results, when PGA and The overall performance is the worst when both are enabled at the same time, with an overall recognition rate of 97.79%. The T72 recognition rate further improved to 97.45%. While higher than when each module was used alone, it was significantly lower than other combinations. This suggests a certain synergy between the two modules, with a significant effect on models like the T72 that rely on local features for distinction. However, the improvement was relatively limited.

[0109] When used in combination with PPDA, the overall recognition rate reached 99.15%, demonstrating excellent performance. BMP2 and BTR70 recognition rates both exceeded 99%, maintaining the performance of PPDA alone. The T72 recognition rate increased to 98.92%, demonstrating that this combination, based on data augmentation, further optimizes model learning through loss constraints, enhancing the differential expression of features between models and further improving the T72 recognition rate.

[0110] When PGA and PPDA are used together, the overall recognition rate is 99.32%, the highest among all multi-component combinations. The recognition rates for BMP2 and BTR70 remain comparable to those achieved using PPDA alone, while the recognition rate for T72 rises to 99.18%, a significant improvement compared to using either PGA or PPDA alone. This demonstrates that the two components working together not only enhance model generalization through data augmentation but also focus on key features through the attention mechanism, significantly improving model performance.

[0111] The complete model integrates PGA, The three major components of PPDA, BMP2 and BTR70, have recognition rates close to 100%, T72 reaches 99.62%, and the overall recognition rate is 99.83%, achieving the overall optimal recognition rate for a single model.

[0112] In summary, a single component can improve the recognition rate of a specific model in a targeted manner (e.g. PGA focuses on optimizing T72, The synergistic effect of multiple components enables systematic optimization of single-model recognition rates through complementary mechanisms. The complete model, integrating these three components, achieved significant improvements in recognition rates across all models, validating the combined effectiveness of phase-preserving data augmentation, phase consistency loss, and phase-guided attention.

[0113] In order to verify the performance of the present invention. The experiments of SAR image target recognition using Example 1 of the present invention and only using the amplitude branch of the network model of the present invention (without any phase information processing) are used as comparative experiment 1; the existing related art "X. Huang, Q. Yang and H. Qiao, "Lightweight Two-Stream Convolutional Neural Network for SAR Target Recognition," in IEEE Geoscience and Remote Sensing Letters, vol. 18, no. 4, pp. 667-671, April 2021." is used as comparative experiment 2; the existing related art "Y. Zhang, J. Xia, X. Gao, L. Xue, X. Zhang and X. Li, "SM-CNN: Separability Measure-Based CNN for SAR Target Recognition," in IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1-5, 2023, Artno.4005605." is used as comparative experiment 3; the existing related art "Y. Guo, Z. Zeng, M. Jin, J. Sun, Z. Meng and W. Hong, "Multilevel Attention Networks for Synthetic ApertureRadar Automatic Target Recognition," in IEEE Geoscience and Remote Sensing Letters, vol. 21, pp. 1-5, 2024, Art no. 4012505.” As comparative experiment 4, a comparative experiment was conducted and evaluated by the recognition accuracy. The parameter settings of all networks were the same.

[0114] As shown in Table 4, for the BMP2 target, the recognition rates in Comparative Experiment 1 were 96.43%, in Comparative Experiment 2 was 99.18%, and in Comparative Experiment 3 was 98.87%. Both Comparative Experiment 4 and Example 1 achieved recognition rates above 99.90%. For the T72 and BTR70 targets, Example 1 achieved the highest recognition rates, at 99.64% and 99.94%, respectively. In terms of overall recognition rate, Example 1 achieved the highest, at 99.83%, followed by Comparative Experiment 4 at 99.48%, Comparative Experiment 3 at 99.31%, and Comparative Experiment 2 at 98.80%. Comparative Experiment 1 had the lowest overall recognition rate, at 97.61%. This demonstrates that Example 1, through its unique dual-branch structure, fully utilizes both amplitude and phase information. In the SAR target recognition task, both for individual models and overall recognition rate, it demonstrates superior performance compared to traditional amplitude-based CNN methods, validating the effectiveness of this network architecture.

[0115] Table 4 Comparative experiments with other algorithms

[0116] The present invention provides a method for SAR image target recognition based on phase-guided attention and amplitude-phase dual-branch network, comprising: The data enhancement module is used to obtain enhanced images based on the characteristics of complex images obtained by SAR imaging, which contain amplitude and phase information. By decoupling the amplitude and phase information, two enhancement strategies, phase preservation and collaborative transformation, are designed. The network construction module is used to design an amplitude-phase dual-branch network architecture, including an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch introduces a phase-guided attention module to extract features from amplitude information and obtain amplitude features; the phase branch extracts features from phase information and obtains phase features; and the feature fusion module concatenates the extracted amplitude and phase features to generate a feature map containing dual information for classification. The network training module is used to construct a total loss function combining phase consistency loss and cross entropy loss based on the enhanced image to train the amplitude-phase dual-branch network architecture and obtain the trained network model parameters; The image recognition module is used to identify the test image set based on the trained network model parameters, and the obtained recognition result is the synthetic aperture radar image target recognition result.

[0117] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory, wherein the memory is configured to store a computer program, the computer program including program instructions, and the processor is configured to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function. The processor described in the embodiment of the present invention can implement the operation of the SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network.

[0118] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device, used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space, which stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the SAR image target recognition method based on phase-guided attention and an amplitude-phase dual-branch network in the above-mentioned embodiment.

[0119] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0121] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network, characterized in that: The following steps are involved: Data enhancement: Based on the fact that complex images acquired by SAR imaging contain both amplitude and phase information, two enhancement strategies, phase preservation and collaborative transformation, are designed by decoupling amplitude and phase information to obtain enhanced images. Network construction: Design an amplitude-phase dual-branch network architecture, including an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch introduces a phase-guided attention module to extract features from amplitude information and obtain amplitude features. The phase branch extracts features from phase information and obtains phase features. The feature fusion module concatenates the extracted amplitude and phase features to generate a feature map containing dual information for classification. Training the network: Based on the enhanced image, a total loss function combining phase consistency loss and cross entropy loss is constructed to train the amplitude-phase dual-branch network architecture and obtain the trained network model parameters; Image recognition: The test image set is recognized based on the trained network model parameters, and the recognition results obtained are synthetic aperture radar image target recognition results.

2. The SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to claim 1 is characterized in that: The phase-preserving enhancement strategy is to only change the amplitude spectrum while keeping the phase spectrum unchanged. The enhancement process is: in, and are the original amplitude and the enhanced amplitude, is the exponent of the power operation; The probability of performing phase-preserving enhancement operation is to use a randomly selected The value performs a power-law transformation on the amplitude spectrum; the exponential The sampling follows a uniform distribution to control the area of ​​amplitude enhancement. When < 1, the enhancement amplitude is small; > 1, the area with larger enhancement amplitude.

3. The SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to claim 1 is characterized in that: The collaborative transformation strategy processes amplitude and phase separately and applies a consistent spatial transformation to both. The enhancement process is: in, R is the rotation matrix, R -1 is its inverse transform, and are the original phase and the enhanced phase respectively; The probability of rotating the amplitude spectrum and phase spectrum at the same time, and the rotation angle is controlled in a small angle range , bilinear interpolation is used to ensure the continuity of phase information after rotation.

4. The SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to claim 1 is characterized in that: The phase-guided attention module includes three dimensions: phase attention, channel attention, and spatial attention. The calculation process of the phase attention weight is: in is the input phase information, For interpolation operations, Represents the Sigmoid activation function, which enables the model to capture features from the perspective of phase; The calculation process of channel attention weight is: in, It is the global representation of the phase feature obtained after convolution, batch normalization, activation and adaptive average pooling. is the amplitude feature after adaptive average pooling, represents a linear transformation, Represents the ReLu activation function; The calculation process of spatial attention weight is: in, is the amplitude feature of the input, Represents maximum pooling and average pooling respectively, Represents the convolution operation; Finally, the process of multi-dimensional attention fusion weighted adjustment of amplitude features is: in, is the amplitude feature of the input, is the total attention weight, Represents the amplitude feature after being processed by the phase-guided attention module.

5. The SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to claim 1, characterized in that: The process of constructing the phase consistency loss function is: in, is the total number of pixels in the image, and The images are at positions The horizontal gradient and vertical gradient at ; is the edge weight parameter, which controls the strength of edge enhancement; is the edge mask at position The value at , the edge mask The specific definition is: in, and They are horizontal gradient and vertical gradient, represents the input phase image, and are the edge detection kernel functions in the horizontal and vertical directions, is the edge mask, which does not participate in back propagation and only serves as a weight factor.

6. The SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to claim 1, characterized in that: The expression of the total loss function is: in, represents the total loss, is the cross entropy loss, is the phase consistency loss, parameter Used to adjust the weight of phase consistency loss in the total loss.

7. The SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to claim 1, characterized in that: The process of training the amplitude-phase dual-branch network model to obtain the trained network model parameters is as follows: The training set formed by the enhanced image is input into the network for training. The amplitude and phase of the enhanced image of the training set are respectively input into the corresponding branches and pass through all modules in sequence. The Adam optimization algorithm is used during training to set the data batch, training rounds, learning rate and weight parameter values; During training, the loss function is calculated through forward propagation, and the network weights are updated through back propagation until the loss function converges to obtain the trained network model parameters.

8. A SAR image target recognition system based on phase-guided attention and amplitude-phase dual-branch network, characterized in that: A SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network according to any one of claims 1 to 7, comprising: The data enhancement module is used to obtain enhanced images based on the characteristics of complex images obtained by SAR imaging, which contain amplitude and phase information. By decoupling the amplitude and phase information, two enhancement strategies, phase preservation and collaborative transformation, are designed. The network construction module is used to design an amplitude-phase dual-branch network architecture, including an amplitude branch, a phase branch, and a feature fusion module. The amplitude branch introduces a phase-guided attention module to extract features from amplitude information and obtain amplitude features; the phase branch extracts features from phase information and obtains phase features; and the feature fusion module concatenates the extracted amplitude and phase features to generate a feature map containing dual information for classification. The network training module is used to construct a total loss function combining phase consistency loss and cross entropy loss based on the enhanced image to train the amplitude-phase dual-branch network architecture and obtain the trained network model parameters; The image recognition module is used to identify the test image set based on the trained network model parameters, and the obtained recognition result is the synthetic aperture radar image target recognition result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a SAR image target recognition method based on phase-guided attention and amplitude-phase dual-branch network are implemented as described in any one of claims 1 to 7.

Images