Deviation SAR target identification method based on causal decoupling characterization learning

Abstract

The invention relates to a deviation SAR target identification method based on causal decoupling representation learning. The method comprises the following steps: carrying out feature extraction, learnable binary mask unwrapping, anti-fact feature generation, batch weighted average label generation and target classification on an input SAR image to obtain a classification result, and realizing causal separation of mixed features through learnable binary mask unwrapping of improved dimension features. The causal unwrapping effect is further purified through the guidance of the anti-fact features and the batch weighted average tags, so that the model pays more attention to the causal part in the mixed features. According to the method, causal feature decoupling characterization and stable classification of the SAR target under the deviation condition can be realized, and thought support can be provided for breakthrough of a radar target recognition technology under the deviation condition.

Description

Technical Field

[0001] This application relates to the technical field of radar target recognition, and particularly to a deviation SAR target recognition method based on causal decoupling representation learning. Background Art

[0002] Synthetic Aperture Radar (SAR) automatic target recognition is a key process for converting the data captured by radar into situation information, and has wide application value in fields such as situation awareness, disaster response, and environmental monitoring. Compared with traditional manual feature extraction methods, deep learning methods have received much attention due to their powerful automatic feature learning and high-dimensional data processing capabilities.

[0003] However, the high-dimensional features extracted by deep learning are tightly coupled in the deep semantic space. Experiments have shown that deep learning often cannot disentangle causal features and non-causal features, but tends to fit all features of the training data, including background, shadows, clutter, noise, and other non-causal features that have no positive contribution to the recognition task. When the test data and the training data satisfy the independent and identically distributed assumption, deep learning shows excellent performance. Once this assumption is broken, such as when the test data has distribution shift problems such as noise interference, observation angle differences, and background changes, non-causal features will dominate the model training process, thus having an adverse impact on the generalization and robustness of the model. Existing methods for identifying problems under deviation conditions mainly have two branches. One is data and feature processing methods such as data augmentation, feature selection, and domain adaptation. The other is model optimization methods such as invariance learning, causal reasoning, and feature disentanglement. Although the above methods have made progress in eliminating spurious associations, they require a large amount of complex prior knowledge, which is difficult to meet for the SAR recognition task, and do not further disentangle causal features from the perspective of counterfactual samples. Therefore, the disentanglement is not thorough, which makes the application of deep learning in the deviation SAR target recognition problem encounter a development bottleneck. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a deviation SAR target recognition method based on causal decoupling representation learning that can be thoroughly disentangled.

[0005] A deviation SAR target recognition method based on causal decoupling representation learning, the method includes:

[0006] Based on the association relationship between various influencing factors in the deviation SAR recognition problem, construct a structural causal model, and then construct a causal decoupling representation network based on the structural causal model;

[0007] Obtain a training data set, where the training data set includes SAR sample images and corresponding target classification labels;

[0008] The causal decoupled representation network is trained using the training data to obtain a trained causal decoupled representation network. The causal decoupled representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features respectively. At the same time, the batch weighted average label generation unit generates batch average weighted labels according to a batch of SAR sample images in the training dataset, and then the classification unit makes predictions according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features respectively. After obtaining the corresponding prediction results, the causal decoupled representation network is trained using the overall loss function calculated based on the prediction results, batch average weighted labels, and target classification labels;

[0009] Obtain the SAR target image, and use the trained causal decoupled representation network to identify the SAR target image to obtain the target recognition result.

[0010] In one embodiment, the structural causal model is represented as external influencing factor nodes that affect SAR imaging, multiple observable endogenous variable nodes, and edges connecting between the nodes to represent the relationship between two nodes. Among them, the multiple observable endogenous variable nodes are input data nodes, causal sub-feature nodes, non-causal sub-feature nodes, and output data nodes.

[0011] In one embodiment, before training the causal decoupled representation network using the training dataset, the SAR sample images in the training dataset are also uniformly scaled to a preset size and normalized.

[0012] In one embodiment, the feature extraction unit uses a network based on a convolutional neural network architecture as the backbone network, including ResNet series, VGG series, DenseNet series, ViT series, ConvNeXt, and SwinTrainsformer neural networks.

[0013] In one embodiment, in the mask generation unit:

[0014] The mixed features are multiplied by a learnable binary mask to achieve feature decoupling, obtaining the causal sub-features and non-causal sub-features;

[0015] Among them, the initialized binary mask is randomly generated, and a threshold is set to map it to a 0-1 binary mask matrix.

[0016] In one embodiment, in the counterfactual feature generation unit:

[0017] Take the sum of the causal sub-features and randomly shuffled non-causal sub-features as the first counterfactual feature;

[0018] Take the sum of the non-causal sub-features and randomly shuffled causal sub-features as the second counterfactual feature.

[0019] In one embodiment, in the batch weighted average label generation unit:

[0020] In the batch of training data, calculate the proportion of sample data of different categories in the entire batch of training data;

[0021] Take the category proportion corresponding to the SAR sample data of one batch as the corresponding batch weighted average label.

[0022] In one embodiment, the overall loss function calculated according to the prediction result, batch average weighted label, and target classification label includes:

[0023] Respectively make predictions according to the causal sub-features, non-causal sub-features, first counterfactual feature, and second counterfactual feature to obtain corresponding first prediction result, second prediction result, third prediction result, and fourth prediction result;

[0024] Calculate the classification loss according to the first prediction result and the corresponding target classification label;

[0025] Calculate the deviation loss according to the second prediction result and the corresponding batch average weighted label;

[0026] Obtain the counterfactual loss from the sum of the MSE loss between the first counterfactual feature and the first prediction result and the MSE loss between the second counterfactual feature and the second prediction result;

[0027] Calculate the overall loss function according to the classification loss, deviation loss, counterfactual loss, and coefficient constraint of the mask generation unit.

[0028] This application also provides a deviation SAR target recognition device based on causal decoupling representation learning. The device includes:

[0029] A causal decoupling representation network construction module, which is used to construct a structural causal model based on the association relationship between various influencing factors in the deviation SAR recognition problem, and then construct a causal decoupling representation network based on the structural causal model;

[0030] A training dataset acquisition module for acquiring a training dataset, where the training dataset includes SAR sample images and corresponding target classification labels;

[0031] A causal decoupling representation network training module for training the causal decoupling representation network using the training data to obtain a trained causal decoupling representation network. The causal decoupling representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit respectively converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features. At the same time, the batch weighted average label generation unit generates a batch average weighted label according to a batch of SAR sample images in the training dataset, and then the classification unit respectively makes predictions according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupling representation network is trained using the overall loss function calculated from the prediction results, batch average weighted labels, and target classification labels;

[0032] An SAR image target recognition module for acquiring an SAR target image and using the trained causal decoupling representation network to recognize the SAR target image to obtain a target recognition result.

[0033] A computer device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the following steps are implemented:

[0034] Based on the correlation relationship between various influencing factors in the deviation SAR recognition problem, construct a structural causal model, and then construct a causal decoupling representation network based on the structural causal model;

[0035] Acquire a training dataset, where the training dataset includes SAR sample images and corresponding target classification labels;

[0036] Use the training data to train the causal decoupled representation network to obtain a trained causal decoupled representation network. The causal decoupled representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features respectively. At the same time, the batch weighted average label generation unit generates batch average weighted labels according to a batch of SAR sample images in the training dataset, and then the classification unit makes predictions respectively according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupled representation network is trained using the overall loss function calculated based on the prediction results, batch average weighted labels, and target classification labels;

[0037] Obtain a SAR target image, and use the trained causal decoupled representation network to identify the SAR target image to obtain a target recognition result.

[0038] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, the following steps are implemented:

[0039] Based on the correlation relationship between various influencing factors in the deviation SAR recognition problem, construct a structural causal model, and then construct a causal decoupled representation network based on the structural causal model;

[0040] Obtain a training dataset, where the training dataset includes SAR sample images and corresponding target classification labels;

[0041] The causal decoupling representation network is trained using the training data to obtain a trained causal decoupling representation network. The causal decoupling representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features respectively. At the same time, the batch weighted average label generation unit generates batch average weighted labels according to a batch of SAR sample images in the training dataset. Then, the classification unit makes predictions according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features respectively. After obtaining the corresponding prediction results, the causal decoupling representation network is trained using the overall loss function calculated based on the prediction results, batch average weighted labels, and target classification labels.

[0042] An SAR target image is acquired, and the trained causal decoupling representation network is used to identify the SAR target image to obtain a target recognition result.

[0043] The above-mentioned deviation SAR target recognition method based on causal decoupling representation learning constructs a structural causal model based on the correlation relationship between various influencing factors in the deviation SAR recognition problem, and then constructs a causal decoupling representation network based on the structural causal model. The causal decoupling representation network includes a mask generation unit that decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features, a counterfactual feature generation unit that converts these two features into counterfactual features, and a batch weighted average label generation unit that generates batch average weighted labels. Inputting the SAR target image into the trained causal decoupling representation network to realize the recognition of the target, this method can realize the causal feature decoupling representation and stable classification of SAR targets under deviation conditions, and can provide an idea support for the breakthrough of radar target recognition technology under deviation conditions. Description of the Drawings

[0044] Figure 1 It is a schematic flowchart of the deviation SAR target recognition method based on causal decoupling representation learning in an embodiment;

[0045] Figure 2 It is a schematic structural diagram of a structural causal model in an embodiment;

[0046] Figure 3 It is a schematic diagram of the working principle of the causal decoupling representation network in an embodiment;

[0047] Figure 4Schematic diagram of the backbone network structure parameters of the feature extraction unit in an embodiment;

[0048] Figure 5 It is the recognition rate confusion matrix in a simulation experiment, where Figure 5 (a) is the recognition rate confusion matrix of the backbone network, Figure 5 (b) is the recognition rate confusion matrix obtained by using this method;

[0049] Figure 6 Structural block diagram of the bias SAR target recognition device based on causal decoupling representation learning in an embodiment;

[0050] Figure 7 Internal structure diagram of a computer device in an embodiment. Specific implementation manners

[0051] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0052] Aiming at the problem that in the existing automatic target recognition technology for synthetic aperture radar (SAR) using deep neural networks, under the condition of SAR feature offset, it is difficult to disentangle the features, or the disentanglement is not thorough, as Figure 1 shown, a bias SAR target recognition method based on causal decoupling representation learning is provided, including the following steps:

[0053] Step S100: Based on the correlation relationship between various influencing factors in the bias SAR recognition problem, construct a structural causal model, and then construct a causal decoupling representation network based on the structural causal model.

[0054] Step S110: Obtain a training data set, where the training data set includes SAR sample images and corresponding target classification labels.

[0055] Step S120: Use the training data to train the causal decoupling representation network to obtain the trained causal decoupling representation network. The causal decoupling representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features respectively. At the same time, the batch weighted average label generation unit generates batch average weighted labels according to a batch of SAR sample images in the training dataset, and then the classification unit makes predictions respectively according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupling representation network is trained according to the overall loss function calculated from the prediction results, batch average weighted labels, and target classification labels.

[0056] Step S130: Obtain the SAR target image, and use the trained causal decoupling representation network to identify the SAR target image to obtain the target recognition result.

[0057] In step S100, the structural causal model is represented as an external influencing factor node U that affects SAR imaging, multiple observable endogenous variable nodes, and edges connecting between the nodes to represent the relationship between two nodes. Among them, the multiple observable endogenous variable nodes are input data node X, causal sub-feature node z c , non-causal sub-feature node z b and output data node Y. The structure of the structural causal model is as Figure 2 shown.

[0058] In this embodiment, the causal decoupling representation network 100 constructed based on the structural causal model includes a feature extraction unit 10, a mask generation unit 20, a counterfactual feature generation unit 30, a batch weighted average label generation unit 40, and a classification unit 50, as Figure 3 shown.

[0059] In step S110, in constructing the training dataset, it includes SAR sample images of multiple different categories of targets. Before using the training dataset to train the causal decoupling representation network 100, the SAR sample images in the training dataset are also uniformly scaled to a preset size and normalized. It should be noted here that data augmentation operations such as rotation, flipping, and cropping are no longer performed on the SAR sample images.

[0060] In one embodiment, the SAR sample images and their corresponding target classification labels in the training dataset are from the 6-class target version of the classic OpenSARShip dataset for radar target recognition. Compared with the 3-class target version of this dataset, the 6-class target version has a more obvious background deviation and greater inter-class similarity. And the original images, i.e., the SAR sample images, are uniformly scaled to images of size 224×224 to meet the input requirements of the classic backbone network selected in this embodiment.

[0061] Specifically, when training the causal decoupling representation network 100, a batch of SAR sample images and labels are selected from the training dataset as the training data. According to the actual situation, including the total number of data in the training dataset and the hardware computing power, the training data for each batch is set. For example, 64 SAR sample images are selected as a batch, and the training is iterated 50 times in total, that is, the value of max_epoch is taken as 50. Then, taking 64 SAR sample images as a batch as an example, this method will be described.

[0062] In step S120, the feature extraction unit 10 in the causal decoupling representation network 100 uses a network based on the convolutional neural network architecture as the backbone network, including ResNet series, VGG series, DenseNet series, ViT series, ConvNeXt, and Swin Trainsformer neural networks.

[0063] In one embodiment, the tiny version of win transformer is used as the feature extraction unit 10. The model of this version uses a hierarchical depth of 2 / 2 / 6 / 2, that is, it is divided into 4 stages. The number of transformer blocks included in each stage is 2, 2, 6, and 2 respectively, and the number of channels is 96, 192, 384, and 768 respectively. The data of 224×224×64 is input into the feature extraction unit, and finally a mixed feature matrix z of 64×7×7×768 is obtained.

[0064] Specifically, the feature extraction unit 10 is expressed as:

[0065] z = f extractor (X)

[0066] In the above formula, z represents the mixed feature matrix, X represents the input image, and f extractor (·) represents the feature extraction unit.

[0067] Further, as Figure 4As shown, it is a schematic diagram of the specific network structure of the feature extraction unit 10. The feature extraction unit 10, namely the Extractor, uses the swin trainsformer tiny. The input image is first divided into a feature map of 56×56 through 4×4 Patch Partition and embedded with 96-dimensional features. The model consists of 4 stages: the first stage contains 2 Transformer Blocks with a depth of 96. The second stage reduces the resolution to 28×28 through Patch Merging, increases the number of channels to 192, and contains 2 Blocks. The third stage reduces the resolution to 14×14, increases the number of channels to 384, and contains 6 Blocks. The fourth stage reduces the resolution to 7×7, increases the number of channels to 768, and contains 2 Blocks. The attention mechanism uses multi-head attention within a 7×7 window and combines the sliding window mechanism for cross-window information interaction. The overall number of parameters of the model is about 28M, which is suitable for lightweight tasks.

[0068] In this embodiment, in the mask generation unit 20: multiplying the mixed features by a learnable binary mask to achieve feature decoupling, obtaining causal sub-features and non-causal sub-features. Among them, the initialized binary mask is randomly generated, and a threshold is set to map it to a 0-1 binary mask matrix.

[0069] Specifically, input the mixed feature matrix into the mask generation unit 20, and multiply it by the generated binary mask to obtain the corresponding causal sub-feature matrix z c and non-causal sub-feature matrix z b .

[0070] Furthermore, first randomly generate a mask matrix of the same size as the mixed feature matrix z, then set the threshold to 0.5, map the mask matrix to a 0-1 binary mask matrix, and multiply this binary mask matrix by the mixed feature matrix to achieve the separation of the mixed feature matrix, obtaining the causal sub-feature matrix z c and non-causal sub-feature matrix z b .

[0071] In one embodiment, the mask generation unit 20 is expressed as:

[0072] {z c , z b} = f LBMD (z)

[0073] In the above formula, f LBMD (·) represents the learnable mask generation unit 20. Specifically, denote the binary mask as M, and its inverse matrix is denoted as Then the mixed feature matrix z realizes decoupling z c = M⊙z and z b = Among them I is the identity matrix.

[0074] In this embodiment, in the counterfactual feature generation unit 30: the sum of the causal sub-features and the randomly shuffled non-causal sub-features is used as the first counterfactual feature, and the sum of the non-causal sub-features and the randomly shuffled causal sub-features is used as the second counterfactual feature.

[0075] Specifically, the counterfactual feature generation unit 30 includes a feature random shuffler and a feature combiner, which are used to decouple the causal sub-feature matrix z c and the non-causal sub-feature matrix z b into the first counterfactual feature z cb and the first counterfactual feature z bc . The counterfactual feature generation unit is expressed as:

[0076] {z cb , z bc} = f CSG (z c , z b )

[0077] In the above formula, z cb = z c + S(z b ), z bc = z b + S(z c ), and S(·) represents randomly shuffling and rearranging the feature vector.

[0078] In this embodiment, the batch weighted average label generation unit 40 calculates the batch weighted average label y a according to the number of samples of each category in each batch of data. Specifically, in this batch of training data, calculate the proportion of the sample data of different categories in the entire batch of training data, and use the proportion of the number of samples of each category in this batch to the number of samples in this batch as the corresponding batch weighted average label.

[0079] Taking the OpenSARShip dataset as an example, there are 6 types, so the batch weighted average label is the proportion of the number of samples of each category in this batch to the batch size of 64.

[0080] Furthermore, the batch weighted average label generation module 40 includes a batch weighted average label generator, which is used to generate the average label of each batch, expressed as:

[0081] y a = f BWALG (X)

[0082] In the above formula, f BWALG (·) represents the batch weighted average label generation unit 40, ya Represents the batch weighted average label, which is a one-dimensional vector consistent with the number of sample types in the batch. The element values in the vector are the ratios of the number of samples of the corresponding category in the batch to the total number of samples in the batch, that is:

[0083] In this embodiment, the classification unit 50 includes a fully connected network f. It is used to input the causal sub-feature matrix z c , the non-causal sub-feature matrix z b , the first counterfactual feature matrix z cb and the second counterfactual feature matrix z bc into the classifier f to obtain the predicted label and , that is, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result.

[0084] Specifically, the classification unit 50 outputs using a fully connected layer to obtain a predicted label with a dimension of 64×4, and uses the softmax activation function, which is expressed as:

[0085]

[0086] In this embodiment, when calculating the overall loss function, according to the first prediction result and the corresponding target classification label y c calculate the classification loss, according to the second prediction result and the corresponding batch average weighted label y a calculate the deviation loss, according to the MSE loss between the first counterfactual feature and the first prediction result , and the sum of the MSE losses between the second counterfactual feature and the second prediction result to obtain the counterfactual loss. Finally, according to the classification loss, deviation loss, counterfactual loss, and the coefficient constraint of the mask generation unit, calculate the overall loss function, which is expressed as:

[0087] Loss comp = Loss cls + λ1·Loss ctf + λ2·Loss bia + λ3·Loss sps ,

[0088] In the above formula, the classification loss Loss cls is the cross-entropy loss between the causal sub-feature and the label, the deviation loss Loss bia is the cross-entropy loss between the non-causal sub-feature and the batch weighted average label, the counterfactual loss is the counterfactual feature and the causal feature prediction label the MSE loss between, counterfactual features and non-causal feature prediction labels The sum of the MSE losses between, Loss sps is the sparse constraint of the learnable binary mask, and λ1, λ2, and λ3 are the corresponding weight hyperparameters.

[0089] In this embodiment, the L1 norm is used to impose a sparse constraint on the binary mask.

[0090] In this embodiment, when training the causal decoupling representation network, according to the overall loss value, the gradient information is backpropagated to update the parameters of units including the feature extractor, learnable binary mask, and classifier. The Adam optimizer is used to update the parameters. Among them, the binary mask is first mapped to a continuous value matrix through the sigmoid function and then the parameter update learning is performed. When implementing segmentation, it is then mapped to a 0-1 binary mask according to the threshold. Until the training termination condition is reached, the trained causal decoupling representation network parameters are saved as the causal decoupling representation network. In some embodiments, the training termination condition is to reach the preset maximum number of iterations, that is, 50 times.

[0091] In step S130, when using the trained causal decoupling representation network for target recognition, a batch of SAR images is input into the trained network to obtain a target recognition matrix.

[0092] In this article, the effectiveness of the method in this article is also demonstrated through simulation experiments. In the simulation experiment, the test set data of the OpenSARShip dataset is used as the test data to perform performance testing on the deviation SAR recognition method based on causal decoupling representation. The accuracy Accuracy is selected as the evaluation index, and the test results are as Figure 5 shown.

[0093] Figure 5 is the recognition rate confusion matrix. The abscissa represents the predicted category of the sample, and the ordinate represents the true label of the sample. Figure 5 (a) is the recognition rate confusion matrix of the backbone network. Figure 5 (b) is the recognition rate confusion matrix obtained by using this method. The numerical value of the element in the matrix represents the corresponding percentage, and the diagonal element represents the proportion of each class of samples that are correctly classified. The experimental results in this embodiment show that the recognition rate of this method is higher than that of the baseline method. Especially in the classification of ContainerShip, General Cargo, and Tanker categories, the recognition rate of this method is more than 10% higher than that of the comparative method.

[0094] In the above-mentioned bias SAR target recognition method based on causal decoupling representation learning, causal separation of mixed features is achieved by unwrapping the learnable binary mask of the enhanced-dimensional features. Guided by counterfactual features and batch-weighted average labels, the causal decoupling effect is further purified, enabling the model to focus more on the causal part of the mixed features. This method can achieve causal feature decoupling representation and stable classification of SAR targets under bias conditions, providing an idea support for the breakthrough of radar target recognition technology under bias conditions.

[0095] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover,

[0096] In one embodiment, as Figure 6 shown, a bias SAR target recognition device based on causal decoupling representation learning is provided, including: a causal decoupling representation network construction module 200, a training data set acquisition module 210, a causal decoupling representation network training module 220, and a SAR image target recognition module 230, where:

[0097] The causal decoupling representation network construction module 200 is used to construct a structural causal model based on the association relationship between various influencing factors in the bias SAR recognition problem, and then construct a causal decoupling representation network based on the structural causal model;

[0098] The training data set acquisition module 210 is used to acquire a training data set, and the training data set includes SAR sample images and corresponding target classification labels;

[0099] The causal decoupling representation network training module 220 is used to train the causal decoupling representation network by using the training data to obtain a trained causal decoupling representation network. The causal decoupling representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit respectively converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features. At the same time, the batch weighted average label generation unit generates a batch average weighted label according to a batch of SAR sample images in the training dataset, and then the classification unit respectively makes predictions according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupling representation network is trained according to the overall loss function calculated based on the prediction results, batch average weighted labels, and target classification labels;

[0100] The SAR image target recognition module 230 is used to obtain a SAR target image and recognize the SAR target image by using the trained causal decoupling representation network to obtain a target recognition result.

[0101] For the specific limitations of the bias SAR target recognition device based on causal decoupling representation learning, reference can be made to the limitations of the bias SAR target recognition method based on causal decoupling representation learning in the above text, which will not be elaborated here. Each module in the above bias SAR target recognition device based on causal decoupling representation learning can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0102] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 7As shown in the figure. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it realizes a deviation SAR target recognition method based on causal decoupling representation learning. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0103] Those skilled in the art can understand that Figure 7 the structure shown in the figure is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0104] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0105] Based on the association relationship between the influencing factors in the deviation SAR recognition problem, construct a structural causal model, and then construct a causal decoupling representation network based on the structural causal model;

[0106] Obtain a training data set, where the training data set includes SAR sample images and corresponding target classification labels;

[0107] Training the causal decoupled representation network using the training data to obtain a trained causal decoupled representation network, where the causal decoupled representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit respectively converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features. At the same time, the batch weighted average label generation unit generates batch average weighted labels according to a batch of SAR sample images in the training dataset, and then the classification unit respectively makes predictions according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupled representation network is trained according to the overall loss function calculated from the prediction results, batch average weighted labels, and target classification labels;

[0108] Obtain a SAR target image, and use the trained causal decoupled representation network to identify the SAR target image to obtain a target recognition result.

[0109] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0110] Based on the association relationship between various influencing factors in the deviation SAR recognition problem, construct a structural causal model, and then construct a causal decoupled representation network based on the structural causal model;

[0111] Obtain a training dataset, where the training dataset includes SAR sample images and corresponding target classification labels;

[0112] The causal decoupled representation network is trained using the training data to obtain a trained causal decoupled representation network. The causal decoupled representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit respectively converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features. At the same time, the batch weighted average label generation unit generates a batch average weighted label according to a batch of SAR sample images in the training dataset. Then, the classification unit makes predictions respectively according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupled representation network is trained using the overall loss function calculated based on the prediction results, batch average weighted labels, and target classification labels;

[0113] Obtain a SAR target image, and use the trained causal decoupled representation network to identify the SAR target image to obtain a target recognition result.

[0114] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0115] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0116] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several variations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A deviation SAR target recognition method based on causal decoupling representation learning, characterized in that, The method includes: Based on the correlation relationships among various influencing factors in the deviation SAR recognition problem, constructing a structural causal model, and then constructing a causal decoupling representation network based on the structural causal model; Obtaining a training data set, where the training data set includes SAR sample images and corresponding target classification labels; Using the training data to train the causal decoupling representation network to obtain a trained causal decoupling representation network. The causal decoupling representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit respectively converts the causal sub-features and non-causal sub-features into corresponding first counterfactual features and second counterfactual features. At the same time, the batch weighted average label generation unit generates a batch average weighted label according to a batch of SAR sample images in the training data set, and then uses the classification unit to make predictions respectively according to the causal sub-features, non-causal sub-features, first counterfactual features, and second counterfactual features. After obtaining the corresponding prediction results, the causal decoupling representation network is trained according to the overall loss function calculated from the prediction results, batch average weighted labels, and target classification labels; Obtaining a SAR target image, and using the trained causal decoupling representation network to identify the SAR target image to obtain a target recognition result.

2. The deviation SAR target recognition method based on causal decoupling representation learning according to claim 1, wherein The structural causal model is represented as external influencing factor nodes affecting SAR imaging, multiple observable endogenous variable nodes, and edges connecting between the nodes to represent the relationship between two nodes. Among them, the multiple observable endogenous variable nodes are input data nodes, causal sub-feature nodes, non-causal sub-feature nodes, and output data nodes.

3. The deviation SAR target recognition method based on causal decoupling representation learning according to claim 1, characterized in that Before training the causal decoupling representation network with the training data set, the SAR sample images in the training data set are also uniformly scaled to a preset size and normalized.

4. The deviation SAR target recognition method based on causal decoupling representation learning according to claim 2, wherein The feature extraction unit uses a network based on a convolutional neural network architecture as the backbone network, including ResNet series, VGG series, DenseNet series, ViT series, ConvNeXt, and Swin Trainsformer neural networks.

5. The deviation SAR target recognition method based on causal decoupling representation learning according to claim 4, characterized in that, In the mask generation unit: Multiplying the mixed features by a learnable binary mask to achieve feature decoupling, and obtaining the causal sub-features and non-causal sub-features; Among them, the initialized binary mask is randomly generated, and a threshold is set to map it to a 0-1 binary mask matrix.

6. The deviation SAR target recognition method based on causal decoupling representation learning according to claim 5, characterized in that, In the counterfactual feature generation unit: Taking the sum of the causal sub-features and randomly shuffled non-causal sub-features as the first counterfactual feature; Taking the sum of the non-causal sub-features and randomly shuffled causal sub-features as the second counterfactual feature.

7. The deviation SAR target recognition method based on causal decoupling representation learning according to claim 6, wherein, In the batch weighted average label generation unit: In the batch of training data, calculate the proportion of sample data of different classes in the entire batch of training data; Use the class ratio corresponding to the batch of SAR sample data as the corresponding batch weighted average label.

8. A deviation SAR target recognition method based on causal decoupling representation learning according to any one of claims 1-7, characterized in that, The overall loss function calculated based on the prediction result, batch average weighted label, and target classification label includes: Respectively make predictions based on the causal sub-features, non-causal sub-features, first counterfactual feature, and second counterfactual feature to obtain corresponding first prediction result, second prediction result, third prediction result, and fourth prediction result; Calculate the classification loss based on the first prediction result and the corresponding target classification label; Calculate the deviation loss based on the second prediction result and the corresponding batch average weighted label; Obtain the counterfactual loss based on the sum of the MSE loss between the first counterfactual feature and the first prediction result and the MSE loss between the second counterfactual feature and the second prediction result; Calculate the overall loss function based on the classification loss, deviation loss, counterfactual loss, and coefficient constraint of the mask generation unit.

9. A bias SAR target recognition device based on causal decoupling representation learning, characterized in that The device includes: A causal decoupling representation network construction module, which is used to construct a structural causal model based on the association relationship between various influencing factors in the deviation SAR recognition problem, and then construct a causal decoupling representation network based on the structural causal model; A training dataset acquisition module, which is used to acquire a training dataset, and the training dataset includes SAR sample images and corresponding target classification labels; A causal decoupling representation network training module, which is used to train the causal decoupling representation network with the training data to obtain a trained causal decoupling representation network. The causal decoupling representation network includes a feature extraction unit, a mask generation unit, a counterfactual feature generation unit, a batch weighted average label generation unit, and a classification unit. Among them, the mask generation unit decouples the mixed features output by the feature extraction unit to obtain causal sub-features and non-causal sub-features. The counterfactual feature generation unit respectively converts the causal sub-features and non-causal sub-features into corresponding first counterfactual feature and second counterfactual feature. At the same time, the batch weighted average label generation unit uses a batch of SAR sample images in the training dataset to generate a batch average weighted label, and then the classification unit respectively makes predictions based on the causal sub-features, non-causal sub-features, first counterfactual feature, and second counterfactual feature. After obtaining the corresponding prediction results, train the causal decoupling representation network with the overall loss function calculated based on the prediction results, batch average weighted labels, and target classification labels; An SAR image target recognition module, which is used to acquire an SAR target image and use the trained causal decoupling representation network to recognize the SAR target image to obtain a target recognition result.

10. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images