A radar target recognition method suitable for small sample cases
By introducing a similarity constraint method into the radar target recognition model, constructing a similarity constraint matrix and a loss function, the problem of poor generalization ability of the radar target recognition model in the case of small samples is solved, and higher recognition accuracy and data representation ability are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2024-12-31
- Publication Date
- 2026-05-12
AI Technical Summary
In small sample sizes, radar target recognition models have poor generalization ability, unstable performance, and difficulty in effectively capturing the similarities and differences between data, resulting in low recognition accuracy.
A similarity constraint matrix and loss function are constructed using a similarity constraint method. Feature mapping is performed in a high-dimensional feature space through gradient optimization. Combined with the cross-entropy loss function, a joint loss function is constructed for model training, which enhances the difference between feature distances of similar targets and feature distances of dissimilar targets.
It improves the accuracy of radar target identification, enhances the generalization ability of the model, alleviates overfitting or underfitting, and improves the data representation ability and interpretability.
Smart Images
Figure CN119780867B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar target recognition technology, and in particular to a radar target recognition method suitable for small sample situations. Background Technology
[0002] The difficulties in field acquisition of radar data stem from several factors. First, radar equipment typically needs to be deployed in specific geographical locations, and some areas may be inaccessible due to complex terrain or harsh environments, limiting data acquisition. Second, weather conditions significantly impact radar data acquisition; severe weather can cause radar equipment to malfunction or degrade data quality. Furthermore, radar data acquisition requires advanced technical equipment and professional personnel, resources that are not always readily available. Given this background, the amount of radar data currently available for research is relatively limited, restricting in-depth research and analysis of radar data.
[0003] On the other hand, aerial targets exhibit different attitudes during flight, which can lead to significant differences in the high-resolution range images of the same type of aerial target under different attitudes. Simultaneously, different types of aerial targets show high similarity under the same attitude, complicating aerial target identification and causing misjudgments. With the rapid development of deep learning technology, feature extraction plays an increasingly important role in computer vision and pattern recognition. The quality of high-dimensional feature representation is crucial for data analysis and model performance. Therefore, studying the high-dimensional feature relationships between samples and fully capturing the similarities and differences between data is of great significance for improving the accuracy of radar target identification. Summary of the Invention
[0004] This invention provides a radar target recognition method suitable for small sample conditions, which solves the problems of poor generalization ability and unstable performance of existing models due to small sample data. It realizes the adjustment of features in a high-dimensional feature space to better capture the relationship between data, better understand the similarity and difference between data, improve the data representation ability and interpretability, and thus provide stronger support for data analysis and model training.
[0005] This invention provides a radar target recognition method suitable for small sample sizes, the method comprising:
[0006] Acquire radar data to be classified and determine the radar target recognition model;
[0007] The radar target recognition model is trained using a similarity constraint method to obtain a trained radar target recognition model; wherein, the similarity constraint loss function corresponding to the similarity constraint method is constructed using the similarity constraint matrix of the training sample set, and the similarity constraint method is applied to the last high-dimensional layer after feature extraction in the radar target recognition model;
[0008] The radar data to be classified is input into the trained radar target recognition model to obtain the recognition result.
[0009] In one possible implementation, training the radar target recognition model using a similarity constraint method to obtain the trained radar target recognition model includes:
[0010] Construct the similarity constraint loss function and similarity constraint matrix corresponding to the aforementioned similarity constraint method;
[0011] Based on the training sample set, an optimized sample set is obtained using a gradient-based optimization method; wherein each sample in the training sample set is a high-dimensional feature;
[0012] The joint loss function of the radar target recognition model is constructed based on the similarity constraint function, and the joint loss function is trained using the optimized sample set to obtain the trained radar target recognition model.
[0013] In one possible implementation, the similarity constraint loss function is expressed as:
[0014]
[0015] Where Y represents the training sample set; T This indicates finding the transpose; S represents the similarity constraint matrix; Y represents the L2 norm; T Y represents the cosine distance between samples in the sample set.
[0016] In one possible implementation, the step of employing a gradient-based optimization method to converge the similarity constraint loss function until the iteration condition is met includes:
[0017] The similarity constraint function is iteratively trained according to the iterative process; wherein the iterative process includes:
[0018] The first derivative of the similarity constraint loss function is obtained.
[0019] Based on the learning rate and the first derivative result, the sample estimate value corresponding to each sample in the sample set in the current training round is obtained, and the estimated sample set for the current round is obtained based on the sample estimate value corresponding to each sample.
[0020] The loss value for the current round is calculated based on the estimated sample set for the current round and the similarity loss function. It is then determined whether the loss value satisfies the iteration condition. If it does, the estimated sample set for the current round is used as the optimized sample set. If not, the iteration process is repeated based on the estimated sample set for the current round until the iteration condition is met.
[0021] In one possible implementation, the joint loss function is expressed as:
[0022] L=(1-α)L c +αL s ;
[0023] Among them, L c L represents the cross-entropy loss function; α represents the weighting factor of the loss function; L s This represents the similarity constraint loss function.
[0024] In one possible implementation, the step of inputting the radar data to be classified into a trained radar target recognition model to obtain the recognition result includes:
[0025] The radar target recognition model is used to extract features from the radar data to be classified, resulting in a high-dimensional feature set.
[0026] The radar target recognition model applies similarity constraints to the high-dimensional feature set to obtain the optimal estimated high-dimensional feature set;
[0027] The radar target recognition model identifies the radar data to be classified based on the optimal estimated high-dimensional feature set, and obtains the recognition result.
[0028] One or more technical solutions provided in this invention have at least the following technical effects or advantages:
[0029] This invention employs a similarity constraint method, constructing a similarity constraint matrix in a high-dimensional feature space to achieve strongly constrained feature mapping. This allows for manipulation of high-dimensional features, improves data representation and interpretability, fully utilizes the relationships between samples, alleviates overfitting or underfitting, enhances the model's generalization ability, and further improves radar target recognition accuracy. Attached Figure Description
[0030] Figure 1 A flowchart illustrating the steps of a radar target recognition method applicable to small sample situations, provided in an embodiment of the present invention;
[0031] Figure 2 A schematic diagram of the similarity constraint matrix provided for an embodiment of the present invention. Detailed Implementation
[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0033] This invention provides a radar target recognition method suitable for small sample sizes, such as... Figure 1 As shown, the method includes the following steps S101 to S103.
[0034] S101, acquire radar data to be classified and determine radar target recognition model;
[0035] S102, The radar target recognition model is trained using a similarity constraint method to obtain a trained radar target recognition model; wherein, the similarity constraint loss function corresponding to the similarity constraint method is constructed using the similarity constraint matrix of the training sample set, and the similarity constraint method is applied to the last high-dimensional layer after feature extraction in the radar target recognition model;
[0036] Specifically, in step S102, the radar target recognition model is trained using the similarity constraint method to obtain the trained radar target recognition model, including the following steps S1021 to S1023.
[0037] S1021, Construct the similarity constraint loss function and similarity constraint matrix corresponding to the similarity constraint method;
[0038] Here, a schematic diagram of the similarity constraint matrix is shown below. Figure 2 As shown, a value of 1 at the intersection of rows and columns of similar target data within the same batch indicates that the data within the same category are very close, while a value of 0 indicates that the data between dissimilar categories are very far apart. Specifically, in Figure 2 In the graph, both the horizontal and vertical axes represent targets. When the target pointed to by the horizontal axis is highly similar to the target pointed to by the vertical axis, it is represented by a yellow square in the graph. If the similarity is low, it is represented by a dark square.
[0039] The similarity constraint loss function is expressed as:
[0040]
[0041] Where Y represents the training sample set;T This indicates finding the transpose; S represents the similarity constraint matrix; Y represents the L2 norm; T Y represents the cosine distance between samples in the sample set. T Y is used to measure the similarity of samples in a high-dimensional feature space, and its size is batchsize*batchsize.
[0042] The goal is to make the similarity between samples approximate a set similarity constraint matrix S, and to obtain the estimated sample set Y* for the current round through an optimization method based on gradient-based alternating projection.
[0043] S1022, Based on the training sample set, a gradient-based optimization method is used to make the similarity constraint loss function converge until the iteration condition is met, thus obtaining an optimized sample set; wherein, each sample in the training sample set is a high-dimensional feature;
[0044] Here, a gradient-based optimization approach is used to converge the similarity constraint loss function until the iteration condition is met, including:
[0045] (1) Iteratively train the similarity constraint function according to the iterative process; wherein, the iterative process includes:
[0046] (1.1) Take the first derivative of the similarity constraint loss function to obtain the first derivative result;
[0047] (1.2) Based on the learning rate and the first derivative result, obtain the sample estimate value corresponding to each sample in the sample set in the current training round, and obtain the estimated sample set for the current round based on the sample estimate value corresponding to each sample;
[0048] (1.3) Calculate the loss value of the current round based on the estimated sample set of the current round and the similarity loss function, and determine whether the loss value satisfies the iteration condition. If it does, use the estimated sample set of the current round as the optimized sample set; otherwise, repeat the iteration process based on the estimated sample set of the current round until the iteration condition is met.
[0049] S1023, construct the joint loss function of the radar target recognition model according to the similarity constraint function, and train the joint loss function using the optimized sample set to obtain the trained radar target recognition model.
[0050] Here, the joint loss function is expressed as:
[0051] L=(1-α)L c +αL s ;
[0052] Among them, Lc L represents the cross-entropy loss function; α represents the weighting factor of the loss function; L s This represents the similarity constraint loss function.
[0053] After incorporating the similarity constraint method into the neural network model, a joint loss function is constructed. The weight factor α of the joint loss function ranges from [0,1]. When α is 0, it indicates that the current training only uses an end-to-end deep neural network; when α is 1, it indicates that the current training only uses the similarity constraint method; when α is (0,1), it indicates that the similarity constraint method is integrated into the end-to-end deep neural network.
[0054] This invention provides a radar target recognition method suitable for small sample situations. It maintains a strongly constrained feature mapping in a high-dimensional space through a similarity constraint matrix, that is, it reduces the feature distance of the same type of target at different angles and increases the feature distance of dissimilar targets at the same angle, thereby achieving the manipulation of high-dimensional features of the target.
[0055] S103, input the radar data to be classified into the trained radar target recognition model to obtain the recognition result.
[0056] Specifically, in step S103, the radar data to be classified is input into the trained radar target recognition model to obtain the recognition result, including the following steps S1031 to S1033.
[0057] S1031, The radar target recognition model is used to extract features from the radar data to be classified, and a high-dimensional feature set is obtained;
[0058] S1032, The radar target recognition model applies similarity constraints to the high-dimensional feature set to obtain the optimal estimated high-dimensional feature set;
[0059] S1033, the radar target recognition model identifies the radar data to be classified based on the optimal estimated high-dimensional feature set, and obtains the recognition result.
[0060] In a specific embodiment of the present invention, the HRRP data used in this example is the measured echo data of three types of aerial targets recorded by a certain ISAR system. The dataset is first preprocessed, and the training sets of "An-26", "Citation" and "Yak-42" are divided into 52, 52 and 44 frames respectively, with 1,000 samples in each frame. 1%, 10%, 20% and 30% of the samples are randomly selected from each frame to form the training set for the following small sample experiments, and 15,000 samples are randomly selected from the test set to form the test set for the following small sample experiments.
[0061] To avoid the uncertainty caused by random sampling, this example conducts five repeated experiments and averages the results to obtain a more convincing conclusion.
[0062] The following analysis examines the impact of similarity constraint methods on HRRP recognition performance using the CNN (Convolutional Neural Network) approach.
[0063] In the small-sample experiments based on the CNN model, as shown in Table 1, adding the similarity constraint method to the CNN model improves the recognition accuracy of HRRP to a certain extent. The improvement is most significant when the training set is 1% of the samples, with an increase of 2.45%. This indicates that the similarity constraint method is a radar target recognition method suitable for small-sample situations.
[0064] Table 1. Small-sample ablation experiments based on CNN model
[0065] Training set size Using only CNN models Add similarity constraint method 1％ 82.70％ 85.15％ 10％ 92.08％ 93.11％ 20％ 92.82％ 93.42％ 30％ 91.93％ 93.51％
[0066] This invention aims to manipulate high-dimensional features through similarity constraints. From the perspective of manifold learning, there exists a manifold spatial structure relationship in high dimensions. This invention utilizes this information through similarity constraints to adjust features in the high-dimensional feature space, better capturing the relationships between data. This allows for a better understanding of the similarities and differences between data, improving the data's representational power and interpretability, thus providing stronger support for data analysis and model training. Even with small sample datasets, this method, combined with end-to-end deep neural networks, can achieve good radar target recognition accuracy.
[0067] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.
[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.
Claims
1. A radar target recognition method suitable for small sample sizes, characterized in that, include: Acquire radar data to be classified and determine the radar target recognition model; The radar target recognition model is trained using a similarity constraint method to obtain a trained radar target recognition model. Specifically, a similarity constraint loss function corresponding to the similarity constraint method is constructed using the similarity constraint matrix of the training sample set, and the similarity constraint method is applied to the last high-dimensional layer after feature extraction in the radar target recognition model. The step of training the radar target recognition model using a similarity constraint method to obtain the trained radar target recognition model includes: Construct the similarity constraint loss function and similarity constraint matrix corresponding to the aforementioned similarity constraint method; wherein, the similarity constraint loss function is expressed as: ;in, Represents the training sample set; This indicates the transpose; Represents the similarity constraint matrix; This indicates a search for the L2 norm; Represents the cosine distance between samples in the sample set; Based on the training sample set, an optimized sample set is obtained using a gradient-based optimization method; wherein each sample in the training sample set is a high-dimensional feature; The joint loss function of the radar target recognition model is constructed based on the similarity constraint function, and the joint loss function is trained using the optimized sample set to obtain the trained radar target recognition model; wherein, the joint loss function is expressed as: ； in, Represents the cross-entropy loss function; This represents the weighting factor of the loss function; Represents the similarity constraint loss function; The radar data to be classified is input into the trained radar target recognition model to obtain the recognition result.
2. The radar target recognition method for small sample situations according to claim 1, characterized in that, The step of employing a gradient-based optimization method to converge the similarity constraint loss function until the iteration condition is met includes: The similarity constraint function is iteratively trained according to the iterative process; wherein the iterative process includes: The first derivative of the similarity constraint loss function is obtained. Based on the learning rate and the first derivative result, the sample estimate value corresponding to each sample in the sample set in the current training round is obtained, and the estimated sample set for the current round is obtained based on the sample estimate value corresponding to each sample. The loss value for the current round is calculated based on the estimated sample set for the current round and the similarity loss function. It is then determined whether the loss value satisfies the iteration condition. If it does, the estimated sample set for the current round is used as the optimized sample set. If not, the iteration process is repeated based on the estimated sample set for the current round until the iteration condition is met.
3. The radar target recognition method for small sample situations according to claim 1, characterized in that, The step of inputting the radar data to be classified into the trained radar target recognition model to obtain the recognition result includes: The radar target recognition model is used to extract features from the radar data to be classified, resulting in a high-dimensional feature set. The radar target recognition model applies similarity constraints to the high-dimensional feature set to obtain the optimal estimated high-dimensional feature set; The radar target recognition model identifies the radar data to be classified based on the optimal estimated high-dimensional feature set, and obtains the recognition result.