HRRP Target Recognition Method and Device Based on Azimuth Angle Geometric Constraint Classifier

Abstract

The present invention relates to a method and device for HRRP target recognition based on azimuth geometric constraint classifier. The method includes: constructing a radar target recognition model. Inputting an unbalanced HRRP data set into a feature extraction module to extract target features, and inputting the target features into an azimuth geometric constraint classifier for category learning. Training the dominant category features to obtain a dominant category azimuth frame classification sub-prototype, and using a single-layer convolution to calculate the sub-prototype displacement between the current dominant category feature and the next dominant category feature's azimuth frame classification sub-prototype. Training the marginal category features according to the sub-prototype displacement to obtain a marginal category azimuth frame classification sub-prototype. Reconstructing a loss function based on the dominant category azimuth frame classification sub-prototype and the marginal category azimuth frame classification sub-prototype, and then optimizing the radar target recognition model. Using this method can improve the radar target recognition performance under the condition of category imbalance.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target recognition, and particularly to a method and device for HRRP target recognition based on azimuth angle geometric constraint classifier. Background Art

[0002] Radar can work stably under non-ideal detection conditions and has the advantage of non-contact. Therefore, radar has become an important sensor in the field of automatic target recognition. High resolution range profile (HRRP) can reflect the distribution of scattering points of the measured target along the radar line of sight direction and contains many physical characteristics of the measured target, such as the distribution of scattering points, the position and intensity of scattering centers, the size of the target, etc. Therefore, the radar target recognition method based on HRRP has received more and more attention from researchers in recent years and has become a hot issue in the field of radar target recognition.

[0003] At present, HRRP target recognition methods can be roughly divided into two categories. One is the traditional HRRP target recognition method, and the other is the HRRP target recognition method based on deep learning. The traditional HRRP target recognition method establishes a clear signal model or physical model, extracts the separable features of HRRP targets and designs corresponding classifiers to achieve effective recognition of targets. The deep learning-based method constructs an artificial neural network, designs a clever structure, constructs an end-to-end deep neural network, trains and optimizes the parameters of the neural network by using a large number of labeled training data, so as to obtain a set of neural network parameters under the background of big data, and uses the trained parameters and the constructed neural network to achieve effective recognition of targets. Although these methods have achieved good recognition performance, most of them are trained on a class-balanced dataset, that is, each class of targets has the same number of training samples. However, in practical applications, especially for non-cooperative targets, due to the influence of factors such as changes in observation conditions, the movement trajectory of the target, and the time when the target appears, it is very difficult or even impossible to obtain the same number of training samples for each class of targets. The dataset that can be obtained for training is often class-imbalanced. In a class-imbalanced dataset, the class with sufficient sample quantity is called the dominant class, and the class with insufficient sample quantity is called the marginal class. Directly applying the above methods established under the assumption of class-balanced dataset to HRRP target recognition under class-imbalanced conditions will cause the model to overfit to the dominant class and ignore the learning of the marginal class, resulting in a decline in model performance and even facing the risk of failure. Summary of the Invention

[0004] Based on this, it is necessary to provide a method and device for HRRP target recognition based on an azimuth geometric constraint classifier, which can improve the radar's target recognition performance under the condition of class-imbalanced data, aiming at the above technical problems.

[0005] A method for HRRP target recognition based on an azimuth geometric constraint classifier, the method comprising:

[0006] Construct a radar target recognition model based on HRRP. The radar target recognition model includes: a feature extraction module, an azimuth geometric constraint classifier, and a loss function.

[0007] Input the imbalanced HRRP data set into the feature extraction module to extract target features, and input the target features into the azimuth geometric constraint classifier for class learning. The target features include: dominant class features and marginal class features.

[0008] Train the dominant class features to obtain a dominant class azimuth frame classification sub-prototype, and use one-layer convolution to calculate the sub-prototype displacement between the current dominant class feature and the next dominant class feature's azimuth frame classification sub-prototype.

[0009] Train the marginal class features according to the sub-prototype displacement to obtain a marginal class azimuth frame classification sub-prototype.

[0010] Reconstruct the loss function according to the dominant class azimuth frame classification sub-prototype and the marginal class azimuth frame classification sub-prototype to obtain an overall loss function. Optimize the radar target recognition model according to the overall loss function, and train the imbalanced HRRP data set through the optimized radar target recognition model to output a target image.

[0011] A device for HRRP target recognition based on an azimuth geometric constraint classifier, the device comprising:

[0012] A model construction module for constructing a radar target recognition model based on HRRP. The radar target recognition model includes: a feature extraction module, an azimuth geometric constraint classifier, and a loss function.

[0013] A feature learning module for inputting the imbalanced HRRP data set into the feature extraction module to extract target features, and inputting the target features as a training sample data set into the azimuth geometric constraint classifier for class learning. The training sample data set includes: dominant class features and marginal class features.

[0014] A sub-prototype displacement acquisition module for training the dominant class features to obtain a dominant class azimuth frame classification sub-prototype, and using one-layer convolution to calculate the sub-prototype displacement between the current dominant class feature and the next dominant class feature's azimuth frame classification sub-prototype.

[0015] An edge class feature training module for training edge class features based on sub-prototype displacement to obtain an edge class azimuth frame classification sub-prototype.

[0016] A target image acquisition module for reconstructing a loss function according to the dominant class azimuth frame classification sub-prototype and the edge class azimuth frame classification sub-prototype to obtain an overall loss function, optimizing a radar target recognition model according to the overall loss function, training an unbalanced HRRP data set through the optimized radar target recognition model, and outputting a target image.

[0017] The above-mentioned HRRP target recognition method and device based on an azimuth geometric constraint classifier. First, the feature extraction module uses a deep convolutional neural network to perform multi-scale feature mining on the original HRRP data, suppresses the interference of dominant class noise through an adaptive feature selection mechanism, and retains the azimuth-sensitive features of the target. In the design of the azimuth geometric constraint classifier, an innovative dual-branch training strategy based on sub-prototype displacement is proposed: aiming at the characteristics of sufficient samples of the dominant class, a dominant class azimuth frame classification sub-prototype library is constructed, and the sub-prototype displacement vector between adjacent frames is calculated through a dynamic convolutional layer to capture the geometric change law of target movement; for the edge class, the displacement vector learned by the dominant class is used as a prior constraint, and the edge class azimuth frame classification sub-prototype is generated through geometric transformation, which improves the generalization performance of the model for edge class recognition without reducing the recognition performance of the model for the dominant class, and solves the problem of inaccurate feature distribution under the condition of small samples. By constructing a multi-task optimization system including prototype contrast loss, displacement constraint loss, and class balance loss, while maintaining the discriminative ability of the dominant class, the model is forced to learn the geometric mapping relationship between the edge class features and the dominant class displacement pattern. Thereby further improving the recognition performance of the model under the condition of class imbalance. The overall loss function of the model consists of classification loss and decoupling loss, which realizes enhancing the recognition accuracy of different class features and improving the radar target recognition performance under the condition of class imbalance data. Description of the Drawings

[0018] Figure 1 It is a schematic flowchart of the HRRP target recognition method based on an azimuth geometric constraint classifier in an embodiment;

[0019] Figure 2 It is a flowchart of the HRRP target recognition steps based on an azimuth geometric constraint classifier in an embodiment

[0020] Figure 3 It is an illustration of the azimuth geometric constraint classifier in an embodiment;

[0021] Figure 4The measured HRRP images of three different types of aircraft in one embodiment, where (a) is an An-26 aircraft, (b) is a propeller aircraft, and (c) is a Yak-42 aircraft;

[0022] Figure 5 Schematic diagram of the recognition accuracy obtained on the measured aircraft dataset in one embodiment;

[0023] Figure 6 Schematic diagram of the confusion matrix of the recognition accuracy under the condition of class imbalance on the measured aircraft dataset in one embodiment;

[0024] Figure 7 Structural block diagram of an HRRP target recognition device based on an azimuth geometric constraint classifier in one embodiment. Detailed implementation manners

[0025] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention 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 invention and are not used to limit the present invention.

[0026] In one embodiment, as Figure 1 shown, a method for HRRP target recognition based on an azimuth geometric constraint classifier is provided, including the following steps:

[0027] Step 102, constructing a radar target recognition model based on HRRP.

[0028] The radar target recognition model includes: a feature extraction module, an azimuth geometric constraint classifier, and a loss function.

[0029] Step 104, inputting the unbalanced HRRP dataset into the feature extraction module to extract target features, and inputting the target features into the azimuth geometric constraint classifier for class learning.

[0030] The target features include: dominant class features and marginal class features.

[0031] Step 106, training the dominant class features to obtain a dominant class azimuth frame classification sub-prototype, and using one-layer convolution to calculate the sub-prototype displacement between the current dominant class feature and the next dominant class feature's azimuth frame classification sub-prototype.

[0032] Step 108, training the marginal class features according to the sub-prototype displacement to obtain a marginal class azimuth frame classification sub-prototype.

[0033] Step 110: Reconstruct the loss function according to the dominant class azimuth frame classification sub-prototype and the edge class azimuth frame classification sub-prototype to obtain the overall loss function. Optimize the radar target recognition model according to the overall loss function, and train the unbalanced HRRP data set through the optimized radar target recognition model to output the target image.

[0034] In the above HRRP target recognition method based on azimuth geometric constraint classifier, first, the feature extraction module uses a deep convolutional neural network to perform multi-scale feature mining on the original HRRP data, and suppresses the interference of dominant class noise through an adaptive feature selection mechanism, retaining the azimuth-sensitive features of the target. In the design of the azimuth geometric constraint classifier, an innovative dual-branch training strategy based on sub-prototype displacement is proposed: aiming at the characteristics of sufficient samples in the dominant class, a dominant class azimuth frame classification sub-prototype library is constructed, and the sub-prototype displacement vector between adjacent frames is calculated through a dynamic convolutional layer to capture the geometric change law of target motion; for the edge class, the displacement vector learned from the dominant class is used as a prior constraint, and the edge class azimuth frame classification sub-prototype is generated through geometric transformation, which improves the generalization performance of the model for edge class recognition without reducing the recognition performance of the model for the dominant class, and solves the problem of inaccurate feature distribution under the condition of small samples. By constructing a multi-task optimization system including prototype contrast loss, displacement constraint loss and class balance loss, while maintaining the discriminative ability of the dominant class, the model is forced to learn the geometric mapping relationship between the edge class features and the dominant class displacement pattern. Thereby further improving the recognition performance of the model under the condition of class imbalance. The overall loss function of the model consists of classification loss and decoupling loss, which realizes enhancing the recognition accuracy of different class features and improving the radar target recognition performance under the condition of class imbalance data.

[0035] In one embodiment, the unbalanced HRRP data set is input into the feature extraction module, and a long short-term memory network is used to extract the change features of the strong scattering centers of the targets in the unbalanced HRRP data set to obtain target features:

[0036] ;

[0037] where, is the target feature, is the unbalanced HRRP data. The target features are used as inputs to the azimuth geometric constraint classifier for class learning. The unbalanced HRRP data set includes: dominant class data and edge class data.

[0038] In one embodiment, the azimuth geometric constraint classifier is used to learn the target azimuth frame classification sub-prototype corresponding to the target features, train the dominant class features, and obtain the dominant class azimuth frame classification sub-prototype. Define the geometric constraint according to the learned dominant class azimuth frame classification sub-prototype:

[0039] ;

[0040] Among them, is the set of displacement of the dominant category azimuth frame classification sub-prototype, is the number of learned target azimuth frame classification sub-prototypes, is the number of dominant categories. And a layer of convolution is used to calculate the sub-prototype displacement between the current dominant category feature and the azimuth frame classification sub-prototype of the next dominant category:

[0041] ;

[0042] Among them, is the rd sub-prototype displacement between the th dominant category azimuth frame classification sub-prototype of the th dominant category and the th azimuth frame classification sub-prototype, is the transformation operation, is the th dominant category azimuth frame classification sub-prototype learned by the th dominant category through training, is the th azimuth frame classification sub-prototype learned by the

[0043] In one embodiment, the edge category features are trained by an azimuth geometric constraint classifier according to the sub-prototype displacement:

[0044] ;

[0045] Among them, is the newly generated edge category azimuth frame classification sub-prototype of the th edge category, is the th edge category azimuth frame classification sub-prototype learned by the th edge category through the training sample data set, is the total number of learned edge category azimuth frame classification sub-prototypes, is the th sub-prototype displacement between the th dominant category azimuth frame classification sub-prototype of the th dominant category and the rd azimuth frame classification sub-prototype, is the averaging operation, among which, .

[0046] In one embodiment, the azimuth geometric constraint classifier uses cosine similarity as the distance metric and classifies the target feature into the category of the classification sub-prototype that is closest to the current target feature:

[0047] ;

[0048] where is the cosine similarity, is the number of target feature categories, is the total number of categories, is the total number of dominant categories, is the total number of marginal categories, is the target feature, is the th azimuth frame classification sub-prototype of the th category, and

[0049] is the distance metric.

[0050] ;

[0051] where is the probability that the target feature belongs to the target azimuth frame classification sub-prototype, is the number of learned target azimuth frame classification sub-prototypes, is the th azimuth frame classification sub-prototype of the

[0052] ;

[0053] where is the reconstructed classification loss function, is the HRRP data to be recognized, is the target category label recognized by the model, is the th azimuth frame classification sub-prototype of the th category, and

[0054] Construct a decoupled loss function according to the distance metric and the dominant category azimuth frame classification sub-prototype:

[0055] ;

[0056] where is the decoupled loss function, is the distance metric, is the th azimuth frame classification sub-prototype of the th category, is the transpose of the distance metric . The overall loss function is constructed based on the classification loss function and the decoupled loss function:

[0057] ;

[0058] wherein, is the overall loss function, is the trade-off parameter. The radar target recognition model is optimized through the overall loss function, and the unbalanced HRRP dataset is trained through the optimized radar target recognition model to output the target image.

[0059] In one embodiment, the azimuth geometric constraint classifier classifies according to an intuitive principle, that is, given an HRRP sample, it is classified into the category of the azimuth prototype closest to it. Specifically, different classification sub-prototypes are assigned to different azimuth frames of each category. Under the condition of class imbalance, the training samples of the dominant category are more sufficient than those of the marginal category. Therefore, the classification sub-prototypes of different azimuth frames of the dominant category can often learn more effectively. However, due to the lack of training samples, the learned classification sub-prototypes of the azimuth frames of the marginal category are often incomplete. To address this issue, the azimuth geometric constraint classifier transfers the geometric constraints between the classification sub-prototypes of different azimuth frames of the dominant category to those between the classification sub-prototypes of different azimuth frames of the marginal category, so that the azimuth frame classification sub-prototypes that are difficult to learn due to the lack of training samples can be learned for the marginal category. As Figure 3 shows the illustrated explanation of the azimuth geometric constraint classifier.

[0060] In the class-imbalanced dataset, assume there are dominant categories and marginal categories. For the dominant category, target azimuth frame classification sub-prototypes can be learned. For the marginal category, only target azimuth frame classification sub-prototypes can be learned from the training samples. The geometric constraint is defined as the set of displacements of all classification sub-prototypes of the dominant category, which can be expressed as . The displacement of the classification sub-prototype can be calculated by formula (1), where represents the prototype displacement between the th sub-prototype and the th sub-prototype of the th dominant category, Denote the transformation operation. In this method, a single convolutional layer is adopted for the transformation operation, which enhances the original sub-prototype displacement in the separability and class-irrelevance in the transformed space.

[0061] (1)

[0062] Using the sub-prototype displacement obtained from formula (1), the azimuth frame classification sub-prototype of the newly generated marginal class can be expressed as formula (2), where denotes the th newly generated azimuth frame classification sub-prototype of the marginal class, denotes the th azimuth frame classification sub-prototype of the marginal class learned through training samples, and denotes the averaging operation.

[0063] (2)

[0064] Using formula (1) and formula (2), all target azimuth frame classification sub-prototypes of the dominant class and the marginal class can be obtained. Based on the obtained classification sub-prototypes, the azimuth geometric constraint classifier classifies the test sample into the class of the classification sub-prototype with the closest distance to it, as shown in formula (3), where denotes the test HRRP, denotes the th azimuth frame classification sub-prototype of the th class, , denotes the total number of classes, denotes dominant classes, denotes marginal classes, denotes the distance metric. In the present invention, the cosine similarity is adopted as the distance metric, and the cosine similarity can be expressed as formula (4).

[0065] (3)

[0066] (4)

[0067] where denotes the test HRRP belongs to the target azimuth frame classification sub-prototype with probability. To satisfy the property that the sum of probabilities is 1 and the values are non-negative, the probability can be expressed as formula (5).

[0068] (5)

[0069] Based on probability , the probability can be expressed as formula (6). The classification loss function generated by the azimuth geometric constraint classifier can be expressed as formula (7), where represents the classification loss generated by the azimuth geometric constraint classifier, represents the number of samples.

[0070] (6)

[0071] (7)

[0072] In one embodiment, in order to avoid the problem of confusion of category-related information that may occur when generating a new azimuth frame classification sub-prototype, a decoupled loss function is further proposed to ensure that the geometric constraint does not contain category-related information of the dominant category. The decoupled loss function aims to separate the geometric constraint and the category-related information of the dominant category. Therefore, in order to maximize the separation degree between the geometric constraint and the azimuth frame classification sub-prototype of the dominant category, an orthogonal constraint is adopted to strengthen the decoupled learning of the geometric constraint and the azimuth frame classification sub-prototype of the dominant category. Specifically, the cosine similarity is used to measure the coherence between the geometric constraint and the azimuth frame classification sub-prototype of the dominant category . The decoupled loss can be expressed as formula (8), where represents the decoupled loss.

[0073] (8)

[0074] Furthermore, the overall loss function of the model consists of two parts, namely the classification loss function generated by the azimuth geometric constraint classifier and the decoupled loss function. The classification loss function generated by the azimuth geometric constraint classifier is as shown in formula (7), the decoupled loss function is as shown in formula (8), and the overall loss function of the model is as shown in formula (10), represents the trade-off parameter.

[0075] (9)

[0076] It should be noted that an azimuth geometric constraint classifier is proposed. Under the condition of class imbalance, the azimuth geometric constraint classifier learns a classification sub-prototype for each azimuth frame of the dominant class, and uses the geometric constraints of the classification sub-prototype of each azimuth frame of the dominant class to generate azimuth frame classification sub-prototypes that are not learned due to insufficient training samples for the marginal class. The azimuth geometric constraint classifier realizes the classification and recognition of the target by using the distance between the HRRP feature and the azimuth frame classification sub-prototype in the feature space. On this basis, in order to remove the class-related information of the dominant class that may be included in the geometric constraint, a decoupled loss function is further proposed. By using the decoupled loss function, the class-irrelevance of the geometric constraint is enhanced. The overall loss function of the model consists of the classification loss function generated by the azimuth geometric constraint classifier and the decoupled loss function. HRRP target recognition experiments under the condition of class imbalance are carried out on a three-class measured aircraft dataset. The experimental results prove the effectiveness of the proposed method under the condition of class imbalance and have important engineering value. As Figure 4 shown, the HRRP images of three different types of aircraft (An-26, propeller aircraft, Yak-42) in the measured aircraft dataset are given. Among them, Figure 4 (a) is the HRRP image measured in the HRRP target recognition experiment under the condition of class imbalance on the An-26 aircraft dataset, Figure 4 (b) is the HRRP image measured in the HRRP target recognition experiment under the condition of class imbalance on the propeller aircraft dataset, Figure 4 (c) is the HRRP image measured in the HRRP target recognition experiment under the condition of class imbalance on the Yak-42 aircraft dataset. Under the condition of class imbalance, the recognition accuracies obtained by this method and other comparative recognition methods on the measured aircraft dataset are as Figure 5 shown. Under the condition of class imbalance, the confusion matrix of the recognition accuracies of each class target by the method of the present invention on the measured aircraft dataset is as Figure 6 shown.

[0077] It should be understood that although Figures 1 - 2 the steps in the flowchart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, Figures 1 - 2 at least a part of the steps in

[0078] In one embodiment, as Figure 7 shown, a HRRP target recognition device based on azimuth geometric constraint classifier is provided, including: a model construction module 702, a feature learning module 704, a sub-prototype displacement acquisition module 706, an edge class feature training module 708, and a target image acquisition module 710, where:

[0079] The model construction module 702 is used to construct a radar target recognition model based on HRRP. The radar target recognition model includes: a feature extraction module, an azimuth geometric constraint classifier, and a loss function.

[0080] The feature learning module 704 is used to input an unbalanced HRRP data set into the feature extraction module to extract target features, and input the target features into the azimuth geometric constraint classifier for class learning. The target features include: dominant class features and edge class features.

[0081] The sub-prototype displacement acquisition module 706 is used to train the dominant class features to obtain a dominant class azimuth frame classification sub-prototype, and use one-layer convolution to calculate the sub-prototype displacement between the current dominant class feature and the next dominant class feature's azimuth frame classification sub-prototype.

[0082] The edge class feature training module 708 is used to train the edge class features according to the sub-prototype displacement to obtain an edge class azimuth frame classification sub-prototype.

[0083] The target image acquisition module 710 is used to reconstruct the loss function according to the dominant class azimuth frame classification sub-prototype and the edge class azimuth frame classification sub-prototype to obtain an overall loss function, optimize the radar target recognition model according to the overall loss function, train the unbalanced HRRP data set through the optimized radar target recognition model, and output a target image.

[0084] For the specific limitations of the HRRP target recognition device based on the azimuth geometric constraint classifier, reference can be made to the limitations of the HRRP target recognition method based on the azimuth geometric constraint classifier in the above text, which will not be elaborated here. Each module in the above HRRP target recognition device based on the azimuth geometric constraint classifier 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.

[0085] Those skilled in the art can understand, Figure 7The structure shown is only a block diagram of some of the structures related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0086] Those of ordinary skill in the art can understand that all or part of the processes in the methods of 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 embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory 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 DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0087] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of 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 described in this specification.

[0088] The above-described embodiments only represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A HRRP target recognition method based on azimuth geometric constraint classifier, characterized in that: The method comprises: Constructing a radar target recognition model based on HRRP; the radar target recognition model includes: a feature extraction module, an azimuth geometric constraint classifier and a loss function; Inputting the unbalanced HRRP data set into the feature extraction module to extract target features, and inputting the target features into the azimuth geometric constraint classifier for category learning; the target features include: dominant category features and marginal category features; Training the dominant category feature to obtain a dominant category azimuth frame classification sub-prototype, and using a layer of convolution to calculate a sub-prototype displacement between the current dominant category feature and the azimuth frame classification sub-prototype of the next dominant category feature; The edge category feature is trained according to the sub-prototype displacement to obtain an edge category azimuth frame classification sub-prototype; The loss function is reconstructed according to the dominant category azimuth frame classification sub-prototype and the marginal category azimuth frame classification sub-prototype to obtain an overall loss function, the radar target recognition model is optimized according to the overall loss function, the unbalanced HRRP dataset is trained by the optimized radar target recognition model, and a target image is output.

2. The method according to claim 1, characterized in that Inputting the unbalanced HRRP data set into the feature extraction module to extract target features, and inputting the target features into the azimuth geometric constraint classifier for category learning, including: The unbalanced HRRP dataset is input into the feature extraction module, and the long short-term memory network is used to extract the change characteristics of the target strong scattering center in the unbalanced HRRP dataset to obtain the target features: in, is the target feature, For unbalanced HRRP data; Inputting the target features into the azimuth geometry classifier for category learning; The unbalanced HRRP dataset includes: dominant category data and marginal category data.

3. The method according to claim 2, characterized in that The dominant category feature is trained to obtain a dominant category azimuth frame classification sub-prototype, and a layer of convolution is used to calculate the sub-prototype displacement between the current dominant category feature and the azimuth frame classification sub-prototype of the next dominant category feature, including: The target azimuth frame classification sub-prototype corresponding to the target feature is learned by the azimuth geometric constraint classifier, and the dominant category feature is trained to obtain the dominant category azimuth frame classification sub-prototype; the geometric constraint is defined according to the learned dominant category azimuth frame classification sub-prototype: in, is the set of sub-prototype displacements of the dominant category azimuth frame classification, is the number of learned target azimuth frame classification sub-prototypes, is the number of dominant categories; A layer of convolution is used to calculate the sub-prototype displacement between the azimuth frame classification sub-prototype of the current dominant category feature and the next dominant category feature: in, For the The leading category The dominant category azimuth frame classification sub-prototype and the The sub-prototype displacement between the azimuth frame classification sub-prototypes, For the transformation operation, For the The dominant category is learned through training dominant category azimuth frame classification sub-prototypes, For the The dominant category is learned through training Azimuth frame classification sub-prototypes.

4. The method according to claim 3, characterized in that The edge category feature is trained according to the sub-prototype displacement to obtain an edge category azimuth frame classification sub-prototype, including: The edge category features are trained by the azimuth geometric constraint classifier according to the sub-prototype displacement: in, For the The edge categories newly generated edge category azimuth frame classification sub-prototype, For the The edge categories are learned through the target features. edge category azimuth frame classification sub-prototypes, is the total number of learned edge category azimuth frame classification sub-prototypes, For the The leading category The dominant category azimuth frame classification sub-prototype and the The sub-prototype displacement between the azimuth frame classification sub-prototypes, For averaging operation, , .

5. The method according to claim 4, characterized in that After the step of training the edge category feature according to the sub-prototype displacement to obtain the edge category azimuth frame classification sub-prototype, the method further includes: The azimuth geometric constraint classifier uses cosine similarity as a distance metric to classify the target feature into the category of the classification sub-prototype that is closest to the current target feature: in, is the cosine similarity, is the number of target feature categories, is the total number of categories, is the total number of dominant categories, is the total number of edge categories, is the target feature, For the Category Azimuth frame classification sub-prototypes, is the distance measure.

6. The method according to claim 5, characterized in that Reconstructing the loss function according to the dominant category azimuth frame classification sub-prototype and the marginal category azimuth frame classification sub-prototype to obtain an overall loss function includes: The probability that the target feature belongs to the target azimuth frame classification sub-prototype is generated according to the dominant category azimuth frame classification sub-prototype, the marginal category azimuth frame classification sub-prototype and the current target feature: in, is the probability that the target feature belongs to the target azimuth frame classification sub-prototype, is the number of learned target azimuth frame classification sub-prototypes, For the Category Azimuth frame classification sub-prototypes; Reconstruct the classification loss generated by the azimuth geometric constraint classifier according to the probability: in, is the reconstructed classification loss function, is the HRRP data to be identified, The target category label identified by the model, For the Category Azimuth frame classification sub-prototypes, is the sample size; A decoupling loss function is constructed based on the distance metric and the dominant category azimuth frame classification sub-prototype: in, To decouple the loss function, is the distance metric, For the Category Azimuth frame classification sub-prototypes, is the distance metric The transpose of Construct an overall loss function based on the classification loss function and the decoupling loss function: in, is the overall loss function, For trade-off parameters; The radar target recognition model is optimized by the overall loss function, the unbalanced HRRP data set is trained by the optimized radar target recognition model, and a target image is output.

7. A HRRP target recognition device based on azimuth geometric constraint classifier, characterized in that: The device comprises: A model building module, used to build a radar target recognition model based on HRRP; the radar target recognition model includes: a feature extraction module, an azimuth geometric constraint classifier and a loss function; A feature learning module is used to input the unbalanced HRRP data set into the feature extraction module to extract target features, and input the target features into the azimuth geometric constraint classifier for category learning; the target features include: dominant category features and marginal category features; A sub-prototype displacement acquisition module is used to train the dominant category feature to obtain a dominant category azimuth frame classification sub-prototype, and use a layer of convolution to calculate the sub-prototype displacement between the current dominant category feature and the azimuth frame classification sub-prototype of the next dominant category feature; An edge category feature training module, used for training the edge category feature according to the sub-prototype displacement to obtain an edge category azimuth frame classification sub-prototype; The target image acquisition module is used to reconstruct the loss function according to the dominant category azimuth frame classification sub-prototype and the marginal category azimuth frame classification sub-prototype to obtain an overall loss function, optimize the radar target recognition model according to the overall loss function, train the unbalanced HRRP data set through the optimized radar target recognition model, and output a target image.

Images