Radar HRRP target identification method based on inverse category effective feature space weighting

Abstract

The invention discloses a radar HRRP target recognition method based on inverse category effective feature space weighting. The method comprises the steps of obtaining a category-unbalanced HRRP echo data set and performing preprocessing; constructing a residual neural network, wherein the residual neural network sequentially comprises a feature extraction module and a classifier module; obtaining a category effective feature space of the HRRP echo data set; obtaining an inverse category frequency weight of each category, and further obtaining an inverse category effective feature space weight of each category; performing two-stage training on the residual neural network by using the cross entropy loss and the inverse category effective feature space weight to obtain a trained residual neural network model; and performing target identification on the to-be-identified HRRP echo data by using the trained residual neural network model. According to the method, a decoupling training strategy and an inverse class effective feature space reweighting strategy are organically combined, and the recognition accuracy gain of the model is improved on an HRRP data set with slight class imbalance and serious class imbalance.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and particularly relates to a radar HRRP target recognition method based on weighted inverse class effective feature space. Background Art

[0002] Radar target recognition refers to the determination of the target type by using the radar echo signal of the target. Compared with the traditional target detection task, the target recognition task needs to measure more in-depth target detail features, which will occupy more radar resources such as time and spectrum. As the core data form of radar target recognition, high-resolution range profile (HRRP) shows irreplaceable technical advantages under the background of complex electromagnetic confrontation and enhanced target concealment. In actual scenarios, traditional radars are vulnerable to low-altitude penetration, electronic interference, and camouflage technologies, while the HRRP-based recognition method provides rich training samples and reliable feature learning basis for the recognition model by constructing a complete multi-dimensional recognition database, that is, integrating a large number of HRRP sequence data of different targets (such as airplanes, missiles, ships) under various attitude angles, pitch angles, different environmental clutter, and electromagnetic interference, thereby significantly improving the robustness of the recognition model to actual scenarios such as low signal-to-noise ratio, partial occlusion, and decoy deception.

[0003] In actual scenarios, due to the inherent differences in the quantity distribution of targets, the extremely complex targets and environments detected by radars, and the limitations of radar system performance, it is usually difficult to obtain enough target HRRP samples for some rare non-cooperative target categories. Therefore, it is often impossible to construct a complex and complete HRRP data set. In the case of class imbalance, the recognition model often overfits the feature space of the head classes with rich samples, resulting in serious underlearning of the tail class representations. Its specific manifestations are as follows: the recognition performance of the model for the majority classes increases slightly, while the recognition performance for the minority classes decreases significantly. To address this problem, existing research usually adopts class imbalance learning methods such as data rebalancing, feature enhancement, model log probability value adjustment, and model loss function adjustment to achieve class rebalancing, increase the attention to the minority classes, and ultimately improve the recognition performance of the minority classes and the overall recognition performance of the recognition model.

[0004] The above existing class imbalance learning methods usually use the statistic of the number of samples contained in a class to characterize the size of the feature space spanned by the class. This approach ignores the logical relationship between samples. Therefore, the number of samples cannot accurately reflect the range of the feature space occupied by the class, further resulting in poor effects of the rebalancing method based on the number of samples. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a radar HRRP target recognition method based on inverse class effective feature space weighting. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] The present invention provides a radar HRRP target recognition method based on inverse class effective feature space weighting, including:

[0007] S1: Obtain a class-imbalanced HRRP echo dataset and perform preprocessing to obtain a preprocessed HRRP echo dataset;

[0008] S2: Construct a residual neural network, which sequentially includes a feature extraction module and a classifier module;

[0009] S3: Obtain the class effective feature space of the HRRP echo dataset;

[0010] S4: Based on the number of samples of each class in the HRRP echo dataset, obtain the inverse class frequency weight of each class, and then obtain the inverse class effective feature space weight of each class;

[0011] S5: Use cross-entropy loss and the inverse class effective feature space weight to train the residual neural network in two stages to obtain a trained residual neural network model;

[0012] S6: Use the trained residual neural network model to perform target recognition on the HRRP echo data to be recognized.

[0013] In an embodiment of the present invention, the feature extraction module includes a convolutional layer, a max pooling layer, a first residual unit, a second residual unit, a third residual unit, and a fourth residual unit connected in sequence. The classifier module includes a global average pooling layer, a fully connected layer, and a Softmax classification layer connected in sequence, where

[0014] The first residual unit includes three residual blocks connected in series. The second residual unit includes four residual blocks connected in series. The third residual unit includes six residual blocks connected in series. The fourth residual unit includes three residual blocks connected in series. And each residual block includes two convolutional layers with 3×3 convolutional kernels, and the output end of the previous residual block is connected to the output end of the next residual block.

[0015] In an embodiment of the present invention, the S3 includes:

[0016] Obtain the sampling probability of data samples with class y and perform normalization to obtain the normalized sample sampling probability;

[0017] Obtain multiple batches of HRRP echo data from the preprocessed HRRP echo dataset based on the normalized sample sampling probability, and obtain the effective feature space of category y in each batch;

[0018] Obtain the effective feature space of category y in the complete HRRP echo dataset composed of all batches based on the effective feature space of category y in each batch;

[0019] Obtain the effective feature space of each category in the complete HRRP echo dataset.

[0020] In an embodiment of the present invention, the sampling probability is expressed as:

[0021]

[0022] where ρ y represents the sampling probability of category y, that is, the probability of data samples containing category y in each batch, and β y is a hyperparameter reflecting the effectiveness of different category samples, with a value in [0, 1], and N y represents the number of data samples containing category y in the collected dataset;

[0023] The normalized sample sampling probability is:

[0024]

[0025] where ρ' y represents the normalized sampling probability of category y, and C represents the number of categories in the HRRP echo dataset.

[0026] In an embodiment of the present invention, the effective feature space of category y in each batch is expressed as:

[0027]

[0028] where represents the effective feature space of category y in the b-th batch in the HRRP echo dataset, B≥b≥1, a yb is a vector of size 1*N yb R yb represents the correlation matrix of {f1, f2,... f Nyb}, N yb is the number of samples belonging to category y within the b-th batch, and the i-th element value of a yb is i = 1, 2,..., N yb , and B represents the number of batches.

[0029] In one embodiment of the present invention, the correlation matrix R yb has the following expression:

[0030]

[0031] where represents the cosine similarity between the feature f i corresponding to the data samples i and j in the HRRP echo dataset, μ j is the prototype of the class y, and E represents taking the expectation. y

[0032] In one embodiment of the present invention, obtaining the inverse class frequency weight for each class based on the number of samples of each class in the HRRP echo dataset includes:

[0033] Obtaining the class probability of class y in the HRRP echo dataset:

[0034]

[0035] where C is the total number of classes in the HRRP echo dataset, and N(y, D) is the total number of data samples belonging to class y in the HRRP echo dataset;

[0036] Obtaining the inverse class frequency weight of class y:

[0037]

[0038] In one embodiment of the present invention, the expression of the inverse class effective feature space weight is:

[0039]

[0040] where represents the effective feature space of class y in the complete HRRP echo dataset;

[0041] Or

[0042] Or

[0043] Or where Φ -1 represents the inverse cumulative distribution function of the standard normal distribution.

[0044] In one embodiment of the present invention, the S6 includes:

[0045] ​The first - stage training of the residual neural network is carried out using cross - entropy loss, and the training strategy is as follows: instance - balanced sampling is adopted for the complete HRRP echo data set, with a total of 200 rounds. In the later stage of training, a learning - rate decay strategy is adopted. At the 160th round, the learning rate is multiplied by a first decay factor, and at the 180th round, the learning rate is multiplied by a second decay factor smaller than the first decay factor until the training ends;

[0046] The second - stage training of the classifier module is carried out using the inverse - class sample - space weighting strategy. The training strategy is as follows: freeze the feature - extraction module, and only make the parameters of the classifier module participate in gradient update. ICEFS weighting is performed on the log - odds values output by the fully - connected layer, and it is trained for 10 rounds to obtain the trained classifier module.

[0047] In an embodiment of the present invention, performing ICEFS weighting on the log - odds values output by the fully - connected layer includes:

[0048] z ICEFS = z·IES(y)

[0049] where z represents the original log - odds value output by the fully - connected layer of the residual neural network, and z ICEFS represents the log - odds value after inverse - class effective feature - space weighting.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. The present invention organically combines the decoupled training strategy and the proposed inverse - class effective feature - space re - weighting strategy. After testing, on the HRRP data sets with mild class imbalance and severe class imbalance, the method proposed by the present invention can bring at least a 5% recognition - accuracy gain to the basic model. The inverse - class effective feature - space re - weighting strategy proposed by the present invention realizes the adjustment of the decision boundary of the recognition model, alleviates the dilemma that the decision boundary oppresses the minority classes due to the uneven distribution of data classes in the feature space, that is, realizes class re - balancing in the feature space.

[0052] 2. The HRRP target - recognition method based on inverse - class effective feature - space weighting proposed by the present invention only requires minor modifications to the basic recognition model, adding operations for calculating weights and re - weighting the log - odds values of the model. And compared with the training of the basic model, it only has 5% more training rounds, with a fast training speed, meeting the real - time requirements of the radar recognition system.

[0053] 3. The present invention focuses on feature - level class re - balancing rather than data - level class re - balancing, without modifying the dataset itself, thus avoiding the problem of interfering with model training due to poor - quality generated data. After testing, the recognition rate of all rare classes of the present invention on the specified test set has an increase of more than 6.5%.

[0054] The following will further elaborate on the present invention in detail with reference to the accompanying drawings and embodiments. Description of the Drawings

[0055] Figure 1 is a flowchart of a radar HRRP target recognition method based on inverse class - effective feature space weighting provided by an embodiment of the present invention;

[0056] Figure 2 is a schematic diagram of the processing process of a radar HRRP target recognition method based on inverse class - effective feature space weighting provided by an embodiment of the present invention;

[0057] Figure 3 is a schematic diagram of the structure of a residual neural network provided by an embodiment of the present invention;

[0058] Figure 4 is a schematic diagram of the calculation process of a class - effective feature space provided by an embodiment of the present invention. Detailed Embodiments

[0059] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the intended invention purpose, the following will, in combination with the accompanying drawings and specific embodiments, elaborate in detail on a radar HRRP target recognition method based on inverse class - effective feature space weighting proposed according to the present invention.

[0060] The foregoing and other technical contents, features, and effects of the present invention can be clearly presented in the following detailed description in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in - depth and specific understanding of the technical means and effects adopted by the present invention to achieve the intended purpose can be obtained. However, the accompanying drawings are only for reference and illustration, and are not used to limit the technical solution of the present invention.

[0061] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant is intended to cover non-exclusive inclusion, so that an article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the article or device comprising the said element.

[0062] Embodiment 1

[0063] This embodiment proposes a radar HRRP target recognition method based on inverse class effective feature space weighting, where the class effective feature space is a parameter used to characterize the coverage range of a class in the feature space. The purpose of this radar HRRP target recognition method is to propose a lightweight class imbalance training strategy. By making only minor modifications to the basic model, it can alleviate the phenomenon that the decision boundary obtained by training the model on a class-imbalanced dataset is overly biased towards the minority class, so as to increase the fault tolerance space for the classifier to judge the minority class, and then achieve more fair training. Finally, the goal of enabling the model to have good recognition performance on the class-imbalanced HRRP dataset is achieved. Compared with the prior art, in the scenario of a class-imbalanced dataset, the radar HRRP target recognition method of the present invention has better generalization, stability, and ease of implementation.

[0064] Most current class-imbalanced recognition methods focus on achieving class rebalancing and alleviating the oppression of the majority class on the minority class during the training process. For example, the log-odds value adjustment method and the loss function adjustment method, which usually perform additive transformations. In contrast, the present invention proposes a multiplicative transformation of the log-odds value of the inverse class frequency (ICF). Compared with the previous work that performs additive adjustments, ICF can also eliminate the bias of frequent classes and alleviate class imbalance, and because it uses multiplicative transformation, ICF increases the probability of rare classes at a faster rate than additive transformation, obtaining a stronger adjustment than additive transformation.

[0065] On this basis, due to the complex correlation patterns among samples, the number of samples cannot accurately represent the size of the class feature space. The weights used in the ICF transformation are calculated based on the number of samples, and this deviation results in the continued oppression of the minority class by the majority class after using the ICF reweighting method. To break this dilemma, based on ICF, the present invention proposes using the parameter of the class effective feature space to replace the number of samples to more accurately represent the actual spanned feature space size of the class. To make the calculation of the class effective feature space feasible, two approximations are made to the class effective feature space: the cosine similarity between samples is used to approximately calculate the class effective feature space, solving the problem of insufficient available samples; the class effective feature space is calculated batch by batch on the batch data obtained by the effective number sampling method, avoiding the high storage resource requirements. Finally, combining the ICF method and the class effective feature space, the present invention proposes an inverse class effective feature space (ICEFS) logit reweighting method. This lightweight method is easy to implement, has a low computational complexity, and can significantly improve the recognition performance of the minority class.

[0066] The implementation of the ICEFS method is based on a decoupled two-stage training strategy: the first stage is the conventional training stage, which optimizes the basic cross-entropy loss to enable the feature extraction module of the model to learn better feature representations; the second stage is the classifier fine-tuning stage, which freezes the parameters of the feature extraction module of the model and only allows the parameters of the classifier module to participate in the gradient update. In this stage, the ICEFS reweighting operation is adopted to adjust the decision boundary learned by the classifier module. (Taking the cnn and resnet models as examples, the last few fully connected layers are regarded as the classifier module, and all the other modules in front are regarded as the feature extraction module).

[0067] Specifically, please refer to Figure 1 and Figure 2 , the radar HRRP target recognition method based on inverse class effective feature space weighting proposed in this embodiment includes the following steps:

[0068] S1: Obtain an HRRP echo data set with class imbalance and perform preprocessing to obtain a preprocessed HRRP echo data set.

[0069] The class-imbalanced HRRP echo dataset in this embodiment includes data of ten types of aircraft targets, which are composed of small and medium-sized civil airliners (majority class), large civil airliners (intermediate class), and non-cooperative flying targets of the other party (minority class). Each data sample in this HRRP echo dataset contains the radar one-dimensional high-resolution range profile. In the preprocessing stage, each data sample is truncated, and only the part containing target information is retained. Finally, the HRRP data dimension corresponding to each preprocessed data sample is 1×256. It should be noted that each data sample is labeled with the category of the target to which it belongs.

[0070] S2: Construct a residual neural network, which sequentially includes a feature extraction module and a classifier module.

[0071] Please refer to Figure 3 , Figure 3 FIG. is a schematic structural diagram of a residual neural network provided by an embodiment of the present invention. The residual neural network in this embodiment sequentially includes a feature extraction module and a classifier module. Among them, the feature extraction module includes a convolutional layer, a max pooling layer, a first residual unit, a second residual unit, a third residual unit, and a fourth residual unit connected in sequence. The classifier module includes a global average pooling layer, a fully connected layer, and a Softmax classification layer connected in sequence. The first residual unit includes three residual blocks connected in series. The second residual unit contains four residual blocks connected in series. The third residual unit contains six residual blocks connected in series. The fourth residual unit contains three residual blocks connected in series. Moreover, each residual block includes two convolutional layers with 3x3 convolutional kernels, and the output end of the previous residual block is connected to the output end of the next residual block.

[0072] Specifically, the size of the input image of the residual neural network in this embodiment is usually 224×224×3 (height×width×number of channels). The convolutional kernel of the convolutional layer in the feature extraction module is 7×7, the stride is 2, and the output channels are 64. The feature map obtained after convolution is downsampled by the max pooling layer (convolutional kernel 3×3, stride 2) to obtain a feature map with a size of 112×112×64.

[0073] The first residual unit includes three residual blocks connected in series. Each residual block consists of two convolutional layers with 3×3 convolutional kernels, which are respectively used for reducing the dimension, performing convolutional operations, and restoring the dimension. Moreover, the output end of the previous residual block is connected to the output end of the next residual block, and the output channels of the first residual unit are 64. The second residual unit contains four residual blocks, each with a 3×3 convolutional kernel. Also, the output end of the previous residual block is connected to the output end of the next residual block, and the output channels of the second residual unit are 128. The third residual unit contains six residual blocks, each with a 3×3 convolutional kernel. And the output end of the previous residual block is connected to the output end of the next residual block, and the output channels of the third residual unit are 256. The fourth residual unit contains three residual blocks, each with a 3×3 convolutional kernel. Additionally, the output end of the previous residual block is connected to the output end of the next residual block, and the output channels of the fourth residual unit are 512.

[0074] After passing through the above-mentioned residual blocks, a global average pooling layer is used to pool each channel, outputting a feature map with a size of 1×1×2048. After the global average pooling layer, a fully connected layer is connected, and then a Softmax layer is connected to output the final classification result.

[0075] It should be noted that the above convolutional layer, max-pooling layer, first residual unit, second residual unit, third residual unit, and fourth residual unit are used as the feature extraction module to obtain the feature map of the input image, and the global average pooling layer, fully connected layer, and Softmax classification layer are used as the classifier module to output the final classification result. Subsequently, two-stage training of ICEFS will be carried out based on this division.

[0076] S3: Obtain the class-effective feature space of the HRRP echo dataset.

[0077] Step S3 of this embodiment includes:

[0078] Obtain the sampling probability of the data samples with class y and normalize it to obtain the normalized sample sampling probability; obtain multiple batches of HRRP echo data from the preprocessed HRRP echo dataset based on the normalized sample sampling probability, and obtain the effective feature space of class y in each batch; obtain the effective feature space of the class in the complete dataset composed of all batches based on the effective feature space of class y in each batch, and use the above steps to obtain the effective feature space of each class in the complete HRRP echo dataset.

[0079] Specifically, the present invention proposes to use the class effective feature space to replace the number of samples to characterize the potential coverage of the class in the feature space. First, the exact calculation formula of the class effective feature space in the ideal case is derived.

[0080] To calculate the effective space corresponding to the class y in the data samples, it is necessary to use multiple different data sets to represent the distribution of the class y. Among them, each data set contains N y data samples with the label y. When the number of data sets used is large enough, the data sample at any position in the class y can be regarded as a random variable following the distribution of the class y, and the feature f i corresponding to the i-th data sample is also a random variable. The prototype variance under non-independent and identically distributed can be calculated as:

[0081]

[0082] where μ y is the prototype of the class y, is the variance corresponding to the class y, f i represents the feature corresponding to the i-th data sample, f j represents the feature corresponding to the j-th data sample, N y represents the number of data samples containing the class y in the data set, a y is a vector of size 1*N y , and R y represents the correlation matrix of {f1, f2,... f Ny}.

[0083] It should be noted that in deep learning, class prototypes usually refer to a method in clustering algorithms that can be used to generate data representations. This method uses clustering algorithms to cluster the input data into multiple clusters, and then calculates the centroid of each cluster, that is, the prototype. These prototypes are usually considered to be "typical" points representing different classes in the data set.

[0084] After introducing the concept of the class effective feature space , the calculation of the prototype variance under non-independent and identically distributed can be expressed as:

[0085]

[0086] where N y es represents the class effective feature space of the class y.

[0087] Combining Equation (1) and Equation (2), the effective feature space of the class y can be calculated as:

[0088]

[0089] According to Equation (3), to obtain the class-effective feature space of class y, only the correlation matrix R needs to be calculated. y . Ideally, if a large number of data samples can be collected for class y, the correlation matrix R can be accurately calculated. y . However, in the class-imbalanced HRRP dataset faced by the present invention, the samples of the minority class are severely scarce. Therefore, it is unrealistic to collect a large number of data samples to calculate the correlation matrix R. y . When the data samples of each class are limited, the correlation between the same-class variables has a large uncertainty, resulting in the dynamic change of the correlation matrix R. y . Based on this dilemma, the calculation of the class-effective feature space will be approximated twice next to make it realizable in the real scenario.

[0090] Specifically, the cosine similarity between the features f i and f j corresponding to the data samples i and j is used to replace the correlation matrix R. y , so the calculation formula of R y is shown in Equation (4):

[0091]

[0092] where represents the cosine similarity between the features f i and f j corresponding to the data samples i and j in the HRRP echo dataset, and E represents taking the expectation.

[0093] Furthermore, in order to reduce the requirement of the calculation of the class-effective feature space for storage resources, the present invention adopts the following strategy: First, an effective sample number sampling strategy is adopted. In this sampling strategy, the sampling probability ρ y of the samples from class y is calculated as follows:

[0094]

[0095] where ρ y represents the sampling probability of class y, that is, the probability of including the data samples of class y in each batch, and β y reflects the effectiveness of the samples of different classes and is a hyperparameter with a value range of [0, 1]. Generally, the empirical values are 0.9, 0.99 or 0.999. Since the sum of the sampling probabilities of all data needs to be 1, ρ y needs to be normalized.

[0096]

[0097] Among them, ρ' y represents the sampling probability of class y after normalization, that is, the probability that the final class y sample is sampled, and C represents the number of classes.

[0098] Then, based on the above sampling strategy, multiple batches of data are obtained. Each batch includes multiple data samples, and the class effective feature space of each batch of data is calculated batch by batch Its calculation formula is shown in Equation (7):

[0099]

[0100] Among them, represents the effective feature space of class y in the b-th batch in the HRRP echo dataset, B≥b≥1, a yb is a vector of size 1*N yb R yb represents the correlation matrix of {f1,f2,...f Nyb}, N yb is the number of samples belonging to class y within the b-th batch, and the i-th element value of a yb is i = 1,2,...,N yb , and B represents the number of batches.

[0101] As Figure 4 shown, in a complete training round process, the operation is repeated until all the data in the training set is traversed. Finally, all the batch class effective feature spaces are superimposed to obtain the total class effective feature space under the complete dataset, and the formula is shown in Equation (8):

[0102]

[0103] Among them, B is the number of batches in a complete training round. That is, assuming that a complete dataset (i.e., the HRRP echo dataset obtained in step S1) has 10,000 training data and the batch size is set to 100, then the dataset is divided into 100 batches for training. After 100 batches, one round of training is completed. The ultimate goal of this step is to estimate the class effective space under the complete dataset (i.e., an entire round). This goal is achieved by calculating the class effective space batch by batch and finally adding the 100 batch class effective spaces.

[0104] S4: Based on the number of samples of each class in the HRRP echo dataset, obtain the inverse class frequency weight of each class, and then obtain the inverse class effective feature space weight of each class.

[0105] It should be noted that in actual situations, it is usually impossible to directly obtain the class distribution (source distribution) of all data in a certain scenario, and only a limited number of training data sets can be sampled from it. These training data sets are obtained by randomly sampling from the source distribution. Therefore, the training data can be considered as a finite sampling approximation of the source distribution. The marginal distribution p D (y) on the training data set D can be calculated by uniformly sampling a training data set D from the source distribution data set s to approximate the true marginal distribution p s (y) of the target domain, which provides a theoretical basis for the derivation of the weight calculation formula of the ICF method. Specifically, given that the training data set D is a set of training HRRP echo data sampled from the source distribution data set s, in this embodiment, it is the class-imbalanced HRRP echo data set obtained in step S1. Define N(y, D) as the total number of samples o y in the training data set D that belong to class y, then there is:

[0106] N(y, D) = |{o y ∈ D}| (9)

[0107] According to formula (9), the class probability p D (y) of class y can be defined as:

[0108]

[0109] where, C is the total number of classes in the HRRP echo data set, and N(y, D) is the total number of data samples in the HRRP echo data set that belong to class y.

[0110] Subsequently, the ICF weight is obtained by taking the logarithm of the inverse of p D (y), that is:

[0111]

[0112] Subsequently, the inverse class effective feature space weight for each class is obtained based on the inverse class frequency weight.

[0113] Combining the inferences of step S3 and step S4, replace the number of samples N(y, D) in the ICF weight calculation formula with the effective feature space to obtain the calculation formula for the inverse class effective feature space weight:

[0114]

[0115] where, represents the effective feature space of class y in the complete HRRP echo data set.

[0116] Then, the logit value of the residual neural network model can be subjected to the IES transformation, and this operation is the Inverse Class’s Effective Feature Space (ICEFS) reweighting method:

[0117] z ICEFS = zIES(y) (13)

[0118] where z represents the original logit value output by the fully connected layer of the residual neural network, and z ICEFS is the logit value after being weighted by the inverse class effective feature space.

[0119] According to different forms of logarithmic calculations, several variants of the ICEFS reweighting method are proposed in the present invention as shown in Table 1.

[0120] Table 1 Weight calculation formulas of different ICEFS variant methods

[0121]

[0122]

[0123] where Φ -1 represents the inverse cumulative distribution function of the standard normal distribution.

[0124] S6: Use the cross-entropy loss and the inverse class effective feature space weights to train the residual neural network in two stages to obtain the trained residual neural network model.

[0125] In the first-stage training, the training strategy of using the cross-entropy loss to train the entire model is as follows: Instance-balanced sampling is adopted for the complete training dataset, that is, each sample has the same probability of being sampled. The first-stage training has a total of 200 rounds. In the later stage of training, a learning rate decay strategy is adopted. At the 160th round, the learning rate is multiplied by a decay factor such as 0.1, and at the 180th round, the learning rate is multiplied by the decay factor 0.01, and this setting is continued until the training ends. This approach can help the residual neural network model to more finely adjust the parameters in the later stage of training, improving the stability and accuracy of the residual neural network model. The training of this stage Figure 3 for the entire model shown is aimed at enabling the feature extraction module to learn good representation capabilities.

[0126] In the second - stage training, a classifier module is trained using the inverse - class - sample - space weighting strategy. According to the partitioning of the residual neural network structure in step S3, the feature extraction module of the residual neural network will be frozen in the second - stage training, so that it cannot participate in gradient update. Only the parameters of the classifier module participate in gradient update. The logarithmic logits output by the fully - connected layer of the residual neural network are weighted by ICEFS using equation (13) to achieve class re - balance, enabling the classifier module to learn a robust balanced classification ability. This stage only needs to be trained for 10 epochs to enable the recognition model to achieve good unbalanced recognition performance.

[0127] It should be noted that the log - odds value is the input of the final softmax layer of the residual neural network. For example, if it is a 10 - classification problem, the log - odds value is a 10×1 vector, and each element represents the odds of the corresponding class. At this time, the inverse - class effective feature - space weight calculated in step S4 is also a 10×1 weight vector, which means multiplying the log - odds value by the inverse - class effective feature - space weight before entering the softmax.

[0128] S6: Use the trained residual neural network model to perform target recognition on the HRRP echo data to be recognized.

[0129] Specifically, directly input the HRRP echo data to be recognized into the trained residual neural network model, calculate the probabilities that the target to be recognized is classified into various classes through the Softmax layer, and select the class corresponding to the highest probability as the recognition result.

[0130] The present invention organically combines the decoupled training strategy and the proposed inverse - class effective feature - space re - weighting strategy. After testing, on the HRRP datasets with mild class imbalance and severe class imbalance, the method proposed by the present invention can bring at least a 5% increase in recognition accuracy to the basic model. The inverse - class effective feature - space re - weighting strategy proposed by the present invention realizes the adjustment of the decision boundary of the recognition model, alleviates the dilemma that the decision boundary oppresses the minority classes caused by the uneven distribution of data classes in the feature space, that is, realizes class re - balance in the feature space.

[0131] The HRRP target recognition method based on inverse - class effective feature - space weighting proposed by the present invention only requires minor modifications to the basic recognition model, adding operations to calculate weights and re - weight the log - odds value of the model. Compared with the training of the basic model, it only has 5% more training epochs, has a fast training speed, and meets the real - time requirements of the radar recognition system.

[0132] The present invention focuses on feature-level class rebalancing rather than data-level class rebalancing, without modifying the dataset itself, thus avoiding the problem of interfering with model training due to poor quality of generated data. After testing, the recognition rate of all rare classes in the specified test set by the present invention has an increase of more than 6.5%.

[0133] Another embodiment of the present invention provides a storage medium in which a computer program is stored, and the computer program is used to execute the steps of the radar HRRP target recognition method based on inverse-class effective feature space weighting described in the above embodiment. Another aspect of the present invention provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps of the radar HRRP target recognition method based on inverse-class effective feature space weighting described in the above embodiment are realized. Specifically, the above integrated module implemented in the form of a software functional module can be stored in a computer-readable storage medium. The above software functional module is stored in a storage medium and includes several instructions for causing an electronic device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute some steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs and other various media that can store program codes.

[0134] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A radar HRRP target recognition method based on weighted inverse-class effective feature space, characterized in that Including: S1: Obtain a class-imbalanced HRRP echo dataset and perform preprocessing to obtain the preprocessed HRRP echo dataset; S2: Construct a residual neural network, which successively includes a feature extraction module and a classifier module; S3: Obtain the class-effective feature space of the HRRP echo dataset; S4: Based on the number of samples of each class in the HRRP echo dataset, obtain the inverse class frequency weight of each class, and further obtain the inverse class-effective feature space weight of each class; S5: Use cross-entropy loss and the inverse class-effective feature space weight to train the residual neural network in two stages to obtain the trained residual neural network model; S6: Use the trained residual neural network model to perform target recognition on the HRRP echo data to be recognized.

2. The radar HRRP target recognition method based on inverse-class effective feature space weighting according to claim 1, wherein The feature extraction module includes a convolutional layer, a max pooling layer, a first residual unit, a second residual unit, a third residual unit, and a fourth residual unit connected in sequence. The classifier module includes a global average pooling layer, a fully connected layer, and a Softmax classification layer. Among them, The first residual unit includes three residual blocks connected in series. The second residual unit contains four residual blocks connected in series. The third residual unit contains six residual blocks connected in series. The fourth residual unit contains three residual blocks connected in series. And each residual block includes two convolutional layers with a 3×3 convolutional kernel, and the output end of the previous residual block is connected to the output end of the next residual block.

3. The radar HRRP target recognition method based on inverse class effective feature space weighting according to claim 1, wherein The S3 includes: Obtain the sampling probability of data samples with class y and perform normalization to obtain the normalized sample sampling probability; Based on the normalized sample sampling probability, obtain multiple batches of HRRP echo data from the preprocessed HRRP echo dataset, and obtain the effective feature space of class y in each batch; Based on the effective feature space of class y in each batch, obtain the effective feature space of class y in the complete HRRP echo dataset composed of all batches; Obtain the effective feature space of each class in the complete HRRP echo dataset.

4. The radar HRRP target recognition method based on weighted inverse-class effective feature space according to claim 3, wherein The sampling probability is expressed as: where ρ y represents the sampling probability of class y, that is, the probability of data samples containing class y in each batch, and β y is a hyperparameter reflecting the effectiveness of samples of different classes, with a value in [0, 1], and N y represents the number of data samples containing class y in the collected dataset; The normalized sample sampling probability is: where ρ' y represents the sampling probability of class y after normalization, and C represents the number of classes in the HRRP echo dataset.

5. The radar HRRP target recognition method based on inverse-class effective feature space weighting according to claim 3, characterized in that The effective feature space of class y in each batch is expressed as: Among them, represents the effective feature space of the b-th batch of class y in the HRRP echo dataset, where B ≥ b ≥ 1, a yb is a vector of size 1*N yb R yb represents the correlation matrix of {f1, f2,... f Nyb}, N yb is the number of samples belonging to class y within the b-th batch, and the i-th element value of a yb is B represents the number of batches.

6. The radar HRRP target recognition method based on inverse class effective feature space weighting according to claim 5, characterized in that Correlation matrix R yb The expression is as follows: Among them, represents the cosine similarity between the features f corresponding to the data samples i and j in the HRRP echo dataset i and f j , μ y is the prototype of class y, and E represents taking the expectation.

7. The radar HRRP target recognition method based on inverse class effective feature space weighting according to claim 1, wherein Based on the number of samples of each class in the HRRP echo dataset, obtain the inverse class frequency weight of each class, including: Obtain the class probability of class y in the HRRP echo dataset: Among them, C is the total number of categories in the HRRP echo dataset, and N(y, D) is the total number of data samples belonging to category y in the HRRP echo dataset; Obtain the inverse class frequency weight of class y:

8. The radar HRRP target recognition method based on inverse class effective feature space weighting according to claim 7, wherein The expression of the inverse class-effective feature space weight is: Among them, represents the effective feature space of the complete category y in the HRRP echo dataset; Or, Or, Or, where Φ -1 represents the inverse cumulative distribution function of the standard normal distribution.

9. The method for radar HRRP target recognition based on inverse-class effective feature space weighting according to claim 7, wherein The S6 includes: Perform the first-stage training on the residual neural network using cross-entropy loss. The training strategy is: perform instance-balanced sampling on the complete HRRP echo dataset for a total of 200 rounds. In the later stage of training, adopt a learning rate decay strategy. Multiply the learning rate by the first decay factor at the 160th round, and multiply the learning rate by a second decay factor smaller than the first decay factor at the 180th round until the training ends; The second-stage training of the classifier module is carried out by adopting an inverse class sample space weighting strategy. The training strategy is as follows: freeze the feature extraction module, and only let the parameters of the classifier module participate in gradient update. Perform ICEFS weighting on the logit values output by the fully connected layer, and train for 10 rounds to obtain the trained classifier module.

10. The radar HRRP target recognition method based on inverse class effective feature space weighting according to claim 9, characterized in that, Performing ICEFS weighting on the logit values output by the fully connected layer includes: z ICEFS = z · IES(y) Among them, z represents the original logit value output by the fully connected layer of the residual neural network, and z ICEFS represents the logit value after being weighted by the inverse class effective feature space.

Images