A synthetic aperture radar image target recognition method

Abstract

This invention discloses a target recognition method for synthetic aperture radar (SAR) images, comprising: extracting features from a first SAR image using a shallow convolutional neural network to obtain mixed features of the first SAR image; decoupling the mixed features using an attention-based feature decomposition method to obtain azimuth and class features of the first SAR image; obtaining a lower bound on the mutual information between the azimuth and class features; constructing an initial target recognition model based on the lower bound estimates of the azimuth, class, and mutual information features; and optimizing the initial target recognition model using an adversarial learning algorithm based on mutual information constraints to obtain an optimized target recognition model, thereby recognizing the azimuth and class of a target in a second SAR image. This invention utilizes feature decoupling to extract class features from the image, which can remove the influence of azimuth on the accuracy of class recognition, thus alleviating the OOD (Out of Target) classification problem.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, and in particular to a method for target recognition from synthetic aperture radar images. Background Technology

[0002] Synthetic Aperture Radar (SAR) possesses unique imaging advantages such as all-day and all-weather operation capabilities, multi-polarization and multi-frequency operating modes, and omnidirectional high-precision targeting of specific targets. Therefore, automatic target identification in SAR images has wide applications in both military and civilian fields. With the development and maturation of deep learning theory and computer vision technology, significant research results have been achieved in deep learning-based automatic target identification of SAR images.

[0003] Due to their imaging mechanism, SAR images exhibit significant azimuth sensitivity, specifically in their image features. The same target displays significantly different image features under different azimuth angles, posing a challenge to automatic target identification in SAR images. Existing deep learning-based methods successfully identify targets in SAR images primarily because they possess sufficient training samples containing complete target azimuth angles, ensuring consistent distribution between training and test samples. However, in practical applications of automatic target identification in SAR images, collecting enough non-cooperative target samples at different azimuth angles is extremely difficult. Generally, only a small number of targets at certain angles can be obtained through reconnaissance, leading to the so-called out-of-distribution (OOD) classification problem, where the distributions of training and test samples differ slightly. In this case, the performance of existing automatic target identification methods for SAR images deteriorates sharply. Therefore, there is a need to propose target identification methods with higher generalization capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a synthetic aperture radar image target recognition method that uses feature decoupling to extract category features from the image, thereby removing the influence of azimuth angle on the accuracy of target category recognition and thus alleviating the OOD classification problem.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] A synthetic aperture radar image target recognition method includes:

[0007] Acquire the first synthetic aperture radar image and the second synthetic aperture radar image;

[0008] A shallow convolutional neural network is used to extract features from the first synthetic aperture radar image to obtain the hybrid features of the first synthetic aperture radar image;

[0009] An attention-based feature decomposition method is used to decouple the hybrid features to obtain the azimuth and category features of the first synthetic aperture radar image;

[0010] Obtain a lower bound for the mutual information between the azimuth feature and the category feature;

[0011] An initial target recognition model is constructed based on the azimuth features, the category features, and the lower bound estimate of the mutual information;

[0012] The initial target recognition model is optimized using an adversarial learning algorithm based on mutual information constraints to obtain an optimized target recognition model.

[0013] The second synthetic aperture radar image is input into the optimized target recognition model to identify the azimuth and category of the target in the second synthetic aperture radar image.

[0014] Optionally, the shallow convolutional neural network includes 16 convolutional layers and 4 max-pooling layers; and the convolutional kernels of the convolutional layers are 3×3 with a stride of 1; the stride of the max-pooling layers is 2.

[0015] Optionally, the step of decoupling the hybrid features using an attention-based feature decomposition method includes:

[0016] An azimuth attention mask is obtained based on the hybrid features;

[0017] The azimuth attention mask is multiplied element-wise with the hybrid feature to obtain the azimuth feature;

[0018] The azimuth feature is subtracted element by element from the mixed feature to obtain the category feature.

[0019] Optionally, the step of obtaining the azimuth attention mask based on the hybrid features includes:

[0020] Spatial attention weights and channel attention weights are obtained based on the hybrid features;

[0021] The spatial attention weights and the channel attention weights are stretched to obtain stretched spatial attention weights and stretched channel attention weights; and the number of channels, height, and width of both the stretched spatial attention weights and stretched channel attention weights are the same as the number of channels, height, and width of the hybrid feature.

[0022] The stretched spatial attention weights and stretched channel attention weights are summed and then averaged to obtain the azimuth attention mask.

[0023] Optionally, the step of obtaining spatial attention weights based on the hybrid features includes:

[0024] The mixed features are subjected to global average pooling and global max pooling in the channel dimension to obtain a first feature map and a second feature map;

[0025] The first feature map and the second feature map are concatenated along the channel dimension to obtain the third feature map;

[0026] The third feature map is convolved to obtain a fourth feature map, and the fourth feature map has 1 channel.

[0027] The fourth feature map is normalized and activated by the Sigmoid function to obtain the spatial attention weights;

[0028] The steps for obtaining channel attention weights based on the hybrid features include:

[0029] The mixed features are subjected to global average pooling and global max pooling along the feature dimension to obtain a first vector and a second vector;

[0030] Add the first vector and the second vector to obtain the third vector;

[0031] The third vector is convolved to obtain a fourth vector, and the number of channels of the fourth vector is less than the number of channels of the third vector.

[0032] The fourth vector is processed by the ReLU activation function and then convolved to obtain the fifth vector, and the number of channels of the fifth vector is the same as the number of channels of the mixed feature.

[0033] The fifth vector is normalized and activated by the Sigmoid function to obtain the channel attention weights.

[0034] Optionally, the step of obtaining the lower bound of the mutual information between the azimuth feature and the category feature includes:

[0035] The azimuth feature is linearly mapped to the first normalized variable;

[0036] The category features are linearly mapped to a second normalized variable;

[0037] The lower bound of the mutual information is estimated based on the first and second normalized variables and using the MINE algorithm.

[0038] Optionally, the expression for the first normalized variable is:

[0039]

[0040] Among them, v r Let x represent the first normalized variable. r Indicates azimuth feature w r The learning parameters of the first linear mapping are represented. Indicates the transpose of wr;

[0041] The expression for the second normalized variable is:

[0042]

[0043] Among them, v f Let x represent the first normalized variable. f Indicates azimuth characteristics, w f The learning parameters represent the second linear mapping. Indicates w f The transpose of .

[0044] Optionally, since the azimuth and category of the target in the first synthetic aperture radar image are known, the steps for constructing the initial target recognition model include:

[0045] An azimuth classification model is constructed based on the azimuth features and the azimuth of the target in the first synthetic aperture radar image.

[0046] A category classification model is constructed based on the category features and the category of the target in the first synthetic aperture radar image;

[0047] The initial target recognition model is obtained by combining the azimuth classification model and the category classification model.

[0048] Optionally, the step of optimizing the initial target recognition model using an adversarial learning algorithm based on mutual information constraints includes:

[0049] The parameters of the shallow convolutional neural network and the attention-based feature decomposition method in the initial target recognition model are kept unchanged, and the learning parameters of the first linear mapping and the learning parameters of the second linear mapping are trained several times to maximize the lower bound of the mutual message and obtain the maximum value of the mutual message.

[0050] The learning parameters of the first linear mapping and the second linear mapping corresponding to the maximum value of the mutual message are kept unchanged. The parameters of the shallow convolutional neural network and the parameters of the attention-based feature decomposition method are trained several times to minimize the maximum value of the mutual message, thereby obtaining the optimized target recognition model.

[0051] Optionally, the expression for the loss function of the optimized target recognition model is:

[0052] L = L category (x f y f )+λ1L azimuth (x r y r )+λ2L MIDAL (x f x r )

[0053] Where L represents the loss function of the optimized target recognition model; L category L represents the categorical loss term. azimuth L represents the azimuth loss term. MIDAL This represents the adversarial learning loss term; y r This represents the actual azimuth label value, y f λ1 represents the actual label value of the category; λ2 represents the hyperparameter that adjusts the proportion of the azimuth loss term; λ3 represents the hyperparameter that adjusts the proportion of the adversarial learning loss term.

[0054] The expression for the adversarial learning loss term is:

[0055]

[0056] Where p represents the first synthetic aperture radar image; F, A, and C represent the feature extraction performed on the first synthetic aperture radar image, the feature decomposition performed on the mixed features, and the linear mapping performed on the azimuth feature and the category feature, respectively; I represents the lower bound of the mutual information.

[0057] Compared with the prior art, the present invention has at least one of the following advantages:

[0058] This invention provides a target recognition method for synthetic aperture radar (SAR) images. The method employs a shallow convolutional neural network to extract features from a first SAR image, obtaining hybrid features of the first SAR image. An attention-based feature decomposition method is used to decouple the hybrid features, obtaining azimuth and category features of the first SAR image. An initial target recognition model is constructed based on the lower bound of the mutual information between the obtained azimuth and category features, as well as the azimuth and category features. An adversarial learning algorithm based on mutual information constraints is used to optimize the initial target recognition model, resulting in an optimized target recognition model, thereby accurately identifying the azimuth and category of targets in new SAR images.

[0059] This invention utilizes feature decoupling to extract azimuth and category features from images, thereby obtaining category features independent of azimuth features. This removes the influence of azimuth on category recognition accuracy, further improving the target recognition accuracy of SAR images and alleviating the OOD classification problem.

[0060] This invention proposes an adversarial learning decorrelation algorithm based on mutual information. By alternating the training of parameters in the target recognition model, the maximum value of the mutual information between azimuth features and category features can be obtained, thereby obtaining the maximum coupling part between azimuth features and category features. In order to minimize the correlation between azimuth features and category features as much as possible by reducing the maximum value of mutual information during adversarial decoupling, it is possible to extract category features that are not sensitive to azimuth and thus alleviate the OOD classification problem. Attached Figure Description

[0061] Figure 1 This is a flowchart of a synthetic aperture radar image target recognition method provided in an embodiment of the present invention;

[0062] Figure 2 This is a logic diagram of a synthetic aperture radar image target recognition method provided in an embodiment of the present invention;

[0063] Figure 3 This is a logic diagram of obtaining an azimuth attention mask based on mixed features in a synthetic aperture radar image target recognition method provided by an embodiment of the present invention. Detailed Implementation

[0064] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the synthetic aperture radar image target recognition method proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0066] Combined with appendix Figures 1-3As shown, this embodiment provides a synthetic aperture radar (SAR) image target recognition method, including: step S1, acquiring a first SAR image (i.e., a first SAR image) and a second SAR image (i.e., a second SAR image); step S2, using a shallow convolutional neural network to extract features from the first SAR image to obtain mixed features of the first SAR image; step S3, using an attention-based feature decomposition method to decouple the mixed features to obtain azimuth and category features of the first SAR image; step S4, obtaining a lower bound of the mutual information between the azimuth and category features; step S5, constructing an initial target recognition model based on the azimuth, category, and mutual information lower bounds; step S6, optimizing the initial target recognition model using an adversarial learning algorithm based on mutual information constraints to obtain an optimized target recognition model; and step S7, inputting the second SAR image into the optimized target recognition model to identify the azimuth and category of the target in the second SAR image.

[0067] Specifically, in this embodiment, in step S1, the first synthetic aperture radar image can be used as a training set for constructing the initial target recognition model and the optimized target recognition model. The first synthetic aperture radar image can be obtained from the publicly available MSTAR (Moving and Stationary Target Acquisition and Recognition) dataset. Simultaneously, the azimuth angle and category of the target in the first synthetic aperture radar image can also be obtained from the MSTAR dataset; that is, the azimuth angle and category of the target in the first synthetic aperture radar image are both known. Furthermore, the azimuth angle of the target in the first synthetic aperture radar image obtained from the MSTAR dataset can be recorded as the actual azimuth angle value, and the category of the target in the first synthetic aperture radar image obtained from the MSTAR dataset can be recorded as the actual category value. The azimuth angle and category of the target in the second synthetic aperture radar image are both unknown, but this invention is not limited thereto.

[0068] Specifically, in some embodiments, after obtaining the actual azimuth angle and category value of the target in the first synthetic aperture radar image, to facilitate the construction of the subsequent target recognition model, the synthetic aperture radar image target recognition method further includes: dividing the 0-360° angle range into several sub-intervals according to a first preset number of tags; obtaining the actual azimuth angle tag value of the target in the first synthetic aperture radar image according to the sub-interval corresponding to the actual azimuth angle value of the target in the first synthetic aperture radar image; dividing all categories involved in the target into several major categories according to a second preset number of tags; and obtaining the actual category tag value of the target in the first synthetic aperture radar image according to the major category corresponding to the actual category value of the target in the first synthetic aperture radar image.

[0069] It is understandable that the angular range of the i-th subinterval is:

[0070]

[0071] Where, θ i N represents the angular range of the i-th subinterval; a This indicates the first preset number of labels; and the actual label value of the azimuth angle is the same as the value of i.

[0072] More specifically, the first preset number of tags N a The number of preset tags can be 36, which divides the 0-360° angle range into 36 sub-intervals. According to formula (1), the angle range of the 7th sub-interval is [60°, 70°). If the actual azimuth angle of the target in the first synthetic aperture radar image is 63°, then the actual azimuth angle value is located in the 7th sub-interval, and the actual tag value of the azimuth angle is 7. The number of preset tags can be 10, which divides all categories involved in the target into 10 major categories, such as the first major category being land, the second major category being water, the third major category being ships, the fourth major category being aircraft, etc. If the actual category of the target in the first synthetic aperture radar image is a ship, then the actual tag value of this category is 3, but the present invention is not limited to this.

[0073] Please continue to refer to this. Figure 1 In step S2, the shallow convolutional neural network includes 16 convolutional layers (Conv) and 4 max pooling layers; the convolutional kernels of the convolutional layers are 3×3 and the stride is 1; the stride of the max pooling layers is 2.

[0074] Specifically, in this embodiment, as shown in Table 1, the layers in the shallow convolutional neural network are ordered as follows: 2 first convolutional layers, 1 first max pooling layer, 2 second convolutional layers, 1 second max pooling layer, 4 third convolutional layers, 1 third max pooling layer, 4 fourth convolutional layers, 1 fourth max pooling layer, and 4 fifth convolutional layers; wherein, the kernel of the first convolutional layer is 3×3 and the number of channels is 64; the kernel of the first max pooling layer is 2×2 and the number of channels is 6. 4; The second convolutional layer has a 3×3 kernel and 128 channels; the second max pooling layer has a 2×2 kernel and 128 channels; the third convolutional layer has a 3×3 kernel and 256 channels; the third max pooling layer has a 2×2 kernel and 256 channels; the fourth convolutional layer has a 3×3 kernel and 512 channels; the fourth max pooling layer has a 2×2 kernel and 512 channels; the fifth convolutional layer has a 3×3 kernel and 512 channels. More specifically, in step S2, after using the shallow convolutional neural network to extract features from the first synthetic aperture radar image, the number of channels of the resulting hybrid features is the same as the number of channels in the fifth convolutional layer, which is 512, but this invention is not limited to this.

[0075] Table 1. Structure of shallow convolutional neural networks

[0076] name kernel size Number of channels number of floors First convolutional layer Conv1 3×3 64 2 First max pooling layer Max pool 1 2×2 64 1 Second convolutional layer Conv2 3×3 128 2 Second max pooling layer Max pool 2 2×2 128 1 Third convolutional layer Conv3 3×3 256 4 Third max pooling layer (Max pool 3) 2×2 256 1 Fourth convolutional layer Conv4 3×3 512 4 Fourth max pooling layer 2×2 512 1 Fifth convolutional layer Conv5 3×3 512 4

[0077] Please continue to refer to this. Figure 1 , Figure 2 and Figure 3 Step S3 includes: Step S31, obtaining an azimuth attention mask based on the mixed features; Step S32, multiplying the azimuth attention mask element-wise with the mixed features to obtain the azimuth features; Step S33, subtracting the azimuth features element-wise from the mixed features to obtain the category features.

[0078] It is understood that step S31 includes: step S311, obtaining spatial attention weights based on the hybrid features; step S312, obtaining channel attention weights based on the hybrid features; step S313, copying and stretching the spatial attention weights and the channel attention weights to obtain stretched spatial attention weights and stretched channel attention weights; and the number of channels, height, and width of both the stretched spatial attention weights and the stretched channel attention weights are the same as the number of channels, height, and width of the hybrid features; step S314, summing and averaging the stretched spatial attention weights and the stretched channel attention weights to obtain the azimuth attention mask.

[0079] Specifically, step S311 includes: step S3111, performing global average pooling and global max pooling on the mixed features along the channel dimension to obtain a first feature map and a second feature map; step S3112, concatenating the first feature map and the second feature map along the channel dimension to obtain a third feature map; step S3113, performing convolution operation on the third feature map to obtain a fourth feature map, wherein the fourth feature map has 1 channel; step S3114, normalizing the fourth feature map and applying a sigmoid activation function to obtain the spatial attention weights.

[0080] Step S312 includes: Step S3121, performing global average pooling and global max pooling on the hybrid features along the feature dimension to obtain a first vector and a second vector; Step S3122, adding the first vector and the second vector to obtain a third vector; Step S3123, performing a convolution operation on the third vector to obtain a fourth vector, wherein the number of channels of the fourth vector is less than the number of channels of the third vector; Step S3124, performing a convolution operation on the fourth vector after passing it through a ReLU activation function to obtain a fifth vector, wherein the number of channels of the fifth vector is the same as the number of channels of the hybrid features; Step S3125, performing normalization processing and a Sigmoid activation function on the fifth vector to obtain the channel attention weights;

[0081] In this embodiment, in step S3111, the size of the input mixed feature is C×H'×W', where C represents the number of channels of the image (i.e., channel dimension), H' represents the number of pixels in the vertical direction (i.e., height), W' represents the number of pixels in the horizontal direction (i.e., width), and H'×W' represents the feature dimension of the image. The size of the first feature map obtained by performing global average pooling on the mixed feature based on the channel dimension is 1×H'×W'. The size of the second feature map obtained by performing global max pooling on the mixed feature based on the channel dimension is also 1×H'×W'. In step S3112, the size of the third feature map is 2×H'×W'. In step S3113, a convolutional layer with a kernel of 7×7 and a channel of 1 can be used to perform convolution operation on the third feature map, and the size of the resulting fourth feature map is 1×H'×W', but the present invention is not limited thereto.

[0082] In this embodiment, in step S3121, multiple sets of convolutional kernels of different sizes can be used to perform global average pooling and global max pooling on the feature dimension of the input mixed features of size C×H'×W', and the pooled features of each set of convolutional kernels are unfolded into vectors and concatenated together to obtain two vectors of size C×1×1, namely the first vector and the second vector; in step S3123, a convolutional layer with a convolutional kernel of 1×1 and a channel number of 1 / r can be used to perform convolution operation on the third vector to obtain the fourth vector with a channel number of C / r; in step S3124, the fourth vector can be passed through the ReLU activation function and then through a convolutional layer with a convolutional kernel of 1×1 and a channel number of r to obtain the fifth vector with the channel number restored to C, but the present invention is not limited thereto.

[0083] In this embodiment, in step S313, the extracted spatial attention weights of size 1×H'×W' and channel attention weights of size C×1×1 are copied and stretched to obtain stretched spatial attention weights and stretched channel attention weights of size C×H'×W'. Then, in step S314, the stretched spatial attention weights and stretched channel attention weights are summed and averaged to obtain the azimuth attention mask.

[0084] More specifically, in this embodiment, the azimuth feature and the category feature obtained according to step S3 can be represented as follows:

[0085]

[0086] x f =Xx r (3)

[0087] Where, x r Indicates azimuth characteristics; x f X represents categorical features; σ represents mixed features; c (X) represents the stretched channel attention weights, σ s (X) represents the stretched spatial attention weights. This represents the azimuth attention mask. Furthermore, the number of channels for both the azimuth feature and the category feature is the same as the number of channels for the hybrid feature.

[0088] Please continue to refer to this. Figure 1 and Figure 2Step S4 includes: Step S41, linearly mapping the azimuth feature to a first normalized variable; Step S42, linearly mapping the category feature to a second normalized variable; Step S43, estimating the lower bound of the mutual information based on the first normalized variable and the second normalized variable, and using the MINE algorithm.

[0089] It is understandable that the expression for the first normalized variable is:

[0090]

[0091] Among them, v r Let x represent the first normalized variable. r Indicates azimuth characteristics, w r The learning parameters of the first linear mapping are represented. Indicates w r Transpose of;

[0092] The expression for the second normalized variable is:

[0093]

[0094] Among them, v f Let x represent the first normalized variable. f Indicates azimuth characteristics, w f The learning parameters represent the second linear mapping. Indicates w f The transpose of .

[0095] Specifically, in this embodiment, in step S41, a first linear normalization mapping module including a first linear layer, a second linear layer, and a third linear layer can be used to map the azimuth feature to the first normalized variable. As shown in Table 2, the number of channels in the first linear layer is the same as the number of channels in the azimuth feature, which is 512; the number of channels in the second linear layer is the same as the number of channels in the first linear layer, which is 512; and the number of channels in the third linear layer is the same as the number of the first preset labels, which is 36. However, this invention is not limited to this.

[0096] Table 2 First Linear Normalization Mapping Module

[0097] name Number of channels First linear layer Linear1 512 Second linear layer Linear2 512 Third linear layer Linear3 36

[0098] Specifically, in this embodiment, in step S42, a second linear normalization mapping module, including a fourth linear layer, a fifth linear layer, and a sixth linear layer, can be used to map the category features to the second normalized variable. As shown in Table 3, the second linear normalization mapping module includes a fourth linear layer, a fifth linear layer, and a sixth linear layer. The number of channels in the fourth linear layer is the same as the number of channels in the category features, which is 512. The number of channels in the fifth linear layer is the same as the number of channels in the fourth linear layer, which is 512. The number of channels in the sixth linear layer is the same as the number of the second preset labels, which is 10. However, this invention is not limited to this.

[0099] Table 3 First Linear Normalization Mapping Module

[0100] name Number of channels Linear4, the fourth linear layer 512 Linear5 (Fifth Linear Layer) 512 Linear6, the sixth linear layer 10

[0101] Specifically, the calculation process for estimating the lower bound I of the mutual information using the MINE (mutual information neural estimation) algorithm based on the first normalized variable and the second normalized variable can be found in existing methods and will not be elaborated here.

[0102] In this embodiment, the mutual information between the azimuth feature and the category feature can represent the correlation between the azimuth feature and the category feature; and the greater the correlation between the azimuth feature and the category feature, the greater the mutual information; the smaller the correlation between the azimuth feature and the category feature, the smaller the mutual information; if the azimuth feature and the category feature are independent, the mutual information is 0, but the present invention is not limited thereto.

[0103] Please continue to refer to this. Figure 1 and Figure 2 Step S5 includes: Step S51, constructing an azimuth classification model based on the azimuth features and the azimuth of the target in the first synthetic aperture radar image; Step S52, constructing a category classification model based on the category features and the category of the target in the first synthetic aperture radar image; Step S53, combining the azimuth classification model and the category classification model to obtain the initial target recognition model.

[0104] Specifically, in step S51, the azimuth feature x can be... r As input data, the azimuth predicted label value of the target in the first synthetic aperture radar image is obtained by combining a convolutional layer with a kernel of 1×1 and the same number of channels as the first preset label amount with a softmax function; then, the cross-entropy function is used as the loss function L of the azimuth classification model. azimuthThe azimuth classification model is constructed based on the predicted azimuth label value of the target in the first synthetic aperture radar image and the actual azimuth label value of the target in the first synthetic aperture radar image. That is, the structure of the azimuth classification model consists of a convolutional layer with a 1×1 kernel and the same number of channels as the first preset label quantity, and a softmax function. The loss function of the azimuth classification model is the cross-entropy function. Furthermore, in this embodiment, a classification model is preferred over a regression model for the azimuth recognition task because regression models have high requirements for the assumptions regarding the sample space distribution. If the assumptions are incorrect, the obtained distribution parameters become meaningless. Moreover, observations show that when the azimuth angle changes slightly, the target's appearance model in the image does not change significantly. Therefore, a classification model is more conducive to correctly extracting azimuth information, but this invention is not limited to this.

[0105] Specifically, in step S52, the category feature x can be... f As input data, a combination of a 1×1 convolutional layer with the same number of channels as the second preset label size and softmax is used to obtain the predicted category label value of the target in the first synthetic aperture radar image; then, the cross-entropy function is used as the loss function L of the category classification model. category The category classification model is constructed based on the predicted label value of the target category in the first synthetic aperture radar image and the actual label value of the target category in the first synthetic aperture radar image. That is, the structure of the category classification model consists of a convolutional layer with a convolutional kernel of 1×1 and the same number of channels as the second preset label amount and a softmax function. The loss function of the category classification model is the cross-entropy function, but the present invention is not limited thereto.

[0106] Please continue to refer to this. Figure 1 and Figure 2 Step S6 includes: Step S61, keeping the parameters of the shallow convolutional neural network and the parameters of the attention-based feature decomposition method in the initial target recognition model unchanged, and training the learning parameters of the first linear mapping and the second linear mapping several times to maximize the lower bound of the mutual message, thereby obtaining the maximum value of the mutual message; Step S62, keeping the learning parameters of the first linear mapping and the second linear mapping corresponding to the maximum value of the mutual message unchanged, and training the parameters of the shallow convolutional neural network and the parameters of the attention-based feature decomposition method several times to minimize the maximum value of the mutual message, thereby obtaining the optimized target recognition model.

[0107] Specifically, in this embodiment, in step S61, by maximizing the lower bound of the mutual information, the maximum coupling between the azimuth feature and the category feature can be obtained. This allows the correlation between the azimuth feature and the category feature to be minimized as much as possible by reducing the maximum value of the mutual information during anti-decoupling in the subsequent step S62. This enables the extraction of the category feature that is not sensitive to the azimuth, thereby alleviating the OOD classification problem. However, this invention is not limited thereto.

[0108] Specifically, in this embodiment, in step S61, the learning parameters of the first linear mapping and the learning parameters of the second linear mapping can be trained 10 times, i.e., the 1st to 10th training rounds; in step S62, the parameters of the shallow convolutional neural network and the parameters of the attention-based feature decomposition method can be trained 20 times, i.e., the 11th to 30th training rounds. Optionally, the parameters of the shallow convolutional neural network include the weights and biases of each layer (such as the convolutional layers and max pooling layers in Table 1); the parameters of the attention-based feature decomposition method include the weights and biases of the global average pooling layer, the global max pooling layer, and the convolutional layer involved in steps S311 and S312. Optionally, in some embodiments, step S6 further includes: step S63, cyclically executing steps S61 and S62 until the total number of training iterations reaches a preset value, but the present invention is not limited thereto.

[0109] Specifically, in this embodiment, the loss function expression of the optimized target recognition model is:

[0110] L = L category (x f y f )+λ1L azimuth (x r y r )+λ2L MIDAL (x f x r (6)

[0111] Where L represents the loss function of the optimized target recognition model; L category L represents the categorical loss term. azimuth L represents the azimuth loss term. MIDAL This represents the adversarial learning loss term; y r This represents the actual azimuth label value, y f λ represents the actual label value of the category; λ1 represents the hyperparameter that adjusts the proportion of the azimuth loss term, and λ2 represents the hyperparameter that adjusts the proportion of the adversarial learning loss term. Optionally, λ1 can be 0.25 and λ2 can be 0.30.

[0112] The expression for the adversarial learning loss term, i.e., the adversarial learning algorithm based on mutual information constraints, is as follows:

[0113]

[0114] Where p represents the first synthetic aperture radar image; F, A, and C represent the feature extraction performed on the first synthetic aperture radar image, the feature decomposition performed on the mixed features, and the linear mapping performed on the azimuth feature and the category feature, respectively; I represents the lower bound of the mutual information.

[0115] Furthermore, in this embodiment, in order to reduce the impact of background speckle noise on the experiment, the input size of the first synthetic aperture radar image can be cropped to 88×88, the network batch size during training is set to 8, the initial learning rate is set to 0.001, the learning rate is multiplied by 0.1 every 30 training rounds, the attenuation coefficient is 4e-8, the momentum is 0.9, for a total of 150 training rounds, and the SGD optimizer is used to optimize the relevant parameters in the target recognition model including the initial one, but the present invention is not limited thereto.

[0116] Furthermore, experiments were conducted using the MSTAR public dataset to verify the effectiveness of the target method provided in this embodiment. Accuracy was used as the evaluation metric; a higher accuracy indicates better classification performance. To simulate the coexistence of cooperative and non-cooperative targets in real-world applications, three different training and test set partitioning methods were implemented based on the number of non-cooperative targets included. Table 4 lists the quantitative evaluation of the classification results obtained using each method on the MSTAR dataset.

[0117] Table 4 shows the recognition performance of each method on the MSTAR dataset.

[0118]

[0119]

[0120] As shown in Table 4, the target recognition method proposed in this embodiment has high classification accuracy. In cases with 9 and 10 non-cooperative classes, the proposed method significantly outperforms the compared methods. Furthermore, as shown in Table 5, the target recognition method proposed in this embodiment significantly outperforms other compared methods under training at different azimuth angles.

[0121] Table 5 shows the recognition performance of each method trained using different azimuth angles.

[0122]

[0123] Furthermore, this embodiment also provides a synthetic aperture radar image target recognition system, including a computer device, which includes at least a microprocessor and a memory. The microprocessor is programmed or configured to execute the steps of the aforementioned synthetic aperture radar image target recognition method, or the memory stores a computer program programmed or configured to execute the aforementioned synthetic aperture radar image target recognition method.

[0124] Furthermore, this embodiment also provides a computer-readable storage medium storing a computer program programmed or configured to perform the aforementioned synthetic aperture radar image target recognition method and computer program.

[0125] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for target recognition in synthetic aperture radar images, characterized in that, include: Acquire the first synthetic aperture radar image and the second synthetic aperture radar image; A shallow convolutional neural network is used to extract features from the first synthetic aperture radar image to obtain the hybrid features of the first synthetic aperture radar image; An attention-based feature decomposition method is used to decouple the hybrid features to obtain the azimuth and category features of the first synthetic aperture radar image; Obtain a lower bound for the mutual information between the azimuth feature and the category feature; An initial target recognition model is constructed based on the azimuth features, the category features, and the lower bound estimate of the mutual information; The initial target recognition model is optimized using an adversarial learning algorithm based on mutual information constraints to obtain an optimized target recognition model. The second synthetic aperture radar image is input into the optimized target recognition model to identify the azimuth and category of the target in the second synthetic aperture radar image.

2. The synthetic aperture radar image target recognition method as described in claim 1, characterized in that, The shallow convolutional neural network includes 16 convolutional layers and 4 max pooling layers; the convolutional kernels of the convolutional layers are 3×3 with a stride of 1; and the stride of the max pooling layers is 2.

3. The synthetic aperture radar image target recognition method as described in claim 1, characterized in that, The steps for decoupling the hybrid features using an attention-based feature decomposition method include: An azimuth attention mask is obtained based on the hybrid features; The azimuth attention mask is multiplied element-wise with the hybrid feature to obtain the azimuth feature; The azimuth feature is subtracted element by element from the mixed feature to obtain the category feature.

4. The synthetic aperture radar image target recognition method as described in claim 3, characterized in that, The steps for obtaining the azimuth attention mask based on the hybrid features include: Spatial attention weights and channel attention weights are obtained based on the hybrid features; The spatial attention weights and the channel attention weights are stretched to obtain stretched spatial attention weights and stretched channel attention weights; and the number of channels, height, and width of both the stretched spatial attention weights and stretched channel attention weights are the same as the number of channels, height, and width of the hybrid feature. The stretched spatial attention weights and stretched channel attention weights are summed and then averaged to obtain the azimuth attention mask.

5. The synthetic aperture radar image target recognition method as described in claim 4, characterized in that, The steps for obtaining spatial attention weights based on the hybrid features include: The mixed features are subjected to global average pooling and global max pooling in the channel dimension to obtain a first feature map and a second feature map; The first feature map and the second feature map are concatenated along the channel dimension to obtain the third feature map; The third feature map is convolved to obtain a fourth feature map, and the fourth feature map has 1 channel. The fourth feature map is normalized and activated by the Sigmoid function to obtain the spatial attention weights; The steps for obtaining channel attention weights based on the hybrid features include: The mixed features are subjected to global average pooling and global max pooling along the feature dimension to obtain a first vector and a second vector; Add the first vector and the second vector to obtain the third vector; The third vector is convolved to obtain a fourth vector, and the number of channels of the fourth vector is less than the number of channels of the third vector. The fourth vector is processed by the ReLU activation function and then convolved to obtain the fifth vector, and the number of channels of the fifth vector is the same as the number of channels of the mixed feature. The fifth vector is normalized and activated by the Sigmoid function to obtain the channel attention weights.

6. The synthetic aperture radar image target recognition method as described in claim 1, characterized in that, The steps for obtaining the lower bound of the mutual information between the azimuth feature and the category feature include: The azimuth feature is linearly mapped to the first normalized variable; The category features are linearly mapped to a second normalized variable; The lower bound of the mutual information is estimated based on the first and second normalized variables and using the MINE algorithm.

7. The synthetic aperture radar image target recognition method as described in claim 6, characterized in that, The expression for the first normalized variable is: Among them, v r Let x represent the first normalized variable. r Indicates azimuth characteristics, w r The learning parameters of the first linear mapping are represented. Indicates w r Transpose of; The expression for the second normalized variable is: Among them, v f Let x represent the first normalized variable. f Indicates azimuth characteristics, w f The learning parameters represent the second linear mapping. Indicates w f The transpose of .

8. The synthetic aperture radar image target recognition method as described in claim 7, characterized in that, The azimuth and category of the target in the first synthetic aperture radar image are known. The steps for constructing the initial target recognition model include: An azimuth classification model is constructed based on the azimuth features and the azimuth of the target in the first synthetic aperture radar image. A category classification model is constructed based on the category features and the category of the target in the first synthetic aperture radar image; The initial target recognition model is obtained by combining the azimuth classification model and the category classification model.

9. The synthetic aperture radar image target recognition method as described in claim 8, characterized in that, The steps for optimizing the initial target recognition model using an adversarial learning algorithm based on mutual information constraints include: The parameters of the shallow convolutional neural network and the attention-based feature decomposition method in the initial target recognition model are kept unchanged, and the learning parameters of the first linear mapping and the learning parameters of the second linear mapping are trained several times to maximize the lower bound of the mutual message and obtain the maximum value of the mutual message. The learning parameters of the first linear mapping and the second linear mapping corresponding to the maximum value of the mutual message are kept unchanged. The parameters of the shallow convolutional neural network and the parameters of the attention-based feature decomposition method are trained several times to minimize the maximum value of the mutual message, thereby obtaining the optimized target recognition model.

10. The synthetic aperture radar image target recognition method as described in claim 9, characterized in that, The expression for the loss function of the optimized target recognition model is as follows: L=L category (x f ,y f )+λ1L azimuth (x r ,y r )+λ2L MIDAL (x f ,x r ) Where L represents the loss function of the optimized target recognition model; L category L represents the categorical loss term. azimuth L represents the azimuth loss term. MIDAL This represents the adversarial learning loss term; y r This represents the actual azimuth label value, y f λ1 represents the actual label value of the category; λ2 represents the hyperparameter that adjusts the proportion of the azimuth loss term; λ1 represents the hyperparameter that adjusts the proportion of the adversarial learning loss term. The expression for the adversarial learning loss term is: Where p represents the first synthetic aperture radar image; F, A, and C represent the feature extraction performed on the first synthetic aperture radar image, the feature decomposition performed on the mixed features, and the linear mapping performed on the azimuth feature and the category feature, respectively; I represents the lower bound of the mutual information.

Images