Model training method, radar high-resolution range image feature extraction method and device

Abstract

This invention discloses a model training method, a radar high-resolution range image feature extraction method and apparatus, relating to the field of data processing technology. The model training method includes: acquiring N sets of training sample data; training the model to be trained based on the N sets of training sample data to obtain a loss value; adjusting the parameters of the model to be trained based on the loss value to obtain a currently trained model; and, if the loss value satisfies the loss convergence condition, determining the feature encoder network in the currently trained model as the radar high-resolution range image feature extraction model. The loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. This invention can improve the accuracy of feature extraction from radar high-resolution range image data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a model training method, a radar high-resolution range image feature extraction method and apparatus. Background Technology

[0002] Radar Automatic Target Recognition (RATR) is one of the core functions of modern radar systems. High Resolution Range Profile (HRRP), as a one-dimensional radar signal, contains information about the distribution of target scattering points along the radar's line-of-sight, making it a crucial data source for target identification.

[0003] High-resolution radar range profile data is obtained by using broadband radar signals to acquire the vector sum of the projections of the target's scattering point echoes onto the radar's line-of-sight direction. The range profile is formed by using the time and position information of the reflected signals, reflecting the target's size, structure, and the distribution information of scattering points along the range direction. It is a key technology for automatic radar target identification.

[0004] Based on the above characteristics, high-resolution radar range image data presents a core technical challenge that distinguishes it from general image data: extreme viewpoint sensitivity. The target's HRRP is a one-dimensional projection. When the radar observation viewpoint changes slightly, the shape, scattering point position, and amplitude of its one-dimensional range image may change drastically, resulting in extremely unstable features.

[0005] Existing end-to-end deep learning models typically perform feature extraction on ordinary image data, without taking into account the viewpoint sensitivity of high-resolution radar range image data. This makes it difficult to extract stable and invariant identity features from HRRP signals that change drastically with viewpoint, resulting in poor feature extraction performance. Summary of the Invention

[0006] The technical problem to be solved by the present invention is that the methods commonly used in the prior art are difficult to extract stable and invariant identity features from HRRP signals that change drastically with the viewing angle, resulting in poor feature extraction effect of HRRP data. In order to solve the above problem, the present invention provides a model training method, a radar high-resolution range image feature extraction method and device.

[0007] The content of this invention includes:

[0008] In a first aspect, embodiments of the present invention provide a method for training a model, comprising:

[0009] Obtain N sets of training sample data. The nth set of training sample data includes high-resolution range image data of the nth observed target under M different azimuth views. N and M are both positive integers, and n is a positive integer less than or equal to N.

[0010] The training model is trained based on N sets of training sample data to obtain a loss value. The training model includes a feature encoder network and a discriminant feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. The discriminant feature projection head is used to map the initial feature vector to obtain a discriminant feature vector.

[0011] Based on the loss value, the parameters of the model to be trained are adjusted to obtain the currently trained model;

[0012] If the loss value satisfies the loss convergence condition, the feature encoder network in the currently trained model to be trained is determined as the radar high-resolution range image feature extraction model.

[0013] The loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. The positive sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in the same group of training sample data, and the negative sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in different groups of training sample data.

[0014] Optionally, the instance-level loss for:

[0015] ;

[0016] ;

[0017] in, For temperature hyperparameters, For indicator functions, For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth perspectives.

[0018] Optionally, the loss value further includes kernel discriminant loss, the kernel discriminant loss for:

[0019] ;

[0020] ;

[0021] ;

[0022] in, Let Gaussian radial basis kernel function be used. For the first The mean of the discriminant feature vectors corresponding to the high-resolution range image data of each observed target. The mean of the discriminant feature vectors corresponding to all the high-resolution distance image data is given. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth perspectives.

[0023] Optionally, the model to be trained further includes a semantic label projection head, which is used to perform semantic mapping on the initial feature vector to obtain a semantic label vector;

[0024] The loss value also includes semantic loss, which is determined based on the sum of squares of the differences between discriminant feature similarity and semantic feature similarity. Discriminant feature similarity is the similarity between discriminant feature vectors corresponding to any two high-resolution distance image data, and semantic similarity is the similarity between semantic label vectors corresponding to any two high-resolution distance image data.

[0025] Optionally, the semantic loss for:

[0026] ;

[0027] ;

[0028] ;

[0029] in, For the first The observed target is at the first Semantic label vectors corresponding to high-resolution distance image data from each azimuth viewpoint. For the first The observed target is at the first Semantic label vectors corresponding to high-resolution distance image data from each azimuth viewpoint. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth perspectives.

[0030] Secondly, embodiments of the present invention provide a method for extracting features from high-resolution radar range images, including:

[0031] Acquire high-resolution range image data to be processed;

[0032] The high-resolution range image data to be processed is input into the radar high-resolution range image feature extraction model for feature extraction to obtain target features;

[0033] The radar high-resolution range image feature extraction model is obtained by training the model using the training method described in the first aspect.

[0034] Thirdly, embodiments of the present invention provide a model training apparatus, comprising:

[0035] The first acquisition module is used to acquire N sets of training sample data. The nth set of training sample data includes high-resolution range image data of the nth observed target under M different azimuth views. N and M are both positive integers, and n is a positive integer less than or equal to N.

[0036] The training module is used to train the model to be trained based on N sets of training sample data to obtain a loss value. The model to be trained includes a feature encoder network and a discriminant feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. The discriminant feature projection head is used to map the initial feature vector to obtain a discriminant feature vector.

[0037] The parameter adjustment module is used to adjust the parameters of the model to be trained based on the loss value, so as to obtain the currently trained model to be trained.

[0038] The determination module is used to determine the feature encoder network in the currently trained model to be trained as a radar high-resolution range image feature extraction model when the loss value satisfies the loss convergence condition.

[0039] The loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. The positive sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in the same group of training sample data, and the negative sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in different groups of training sample data.

[0040] Fourthly, embodiments of the present invention provide a radar high-resolution range image feature extraction device, comprising:

[0041] The second acquisition module is used to acquire the high-resolution range image data to be processed.

[0042] A radar high-resolution range image feature extraction model is used to extract features from the high-resolution range image data to be processed, thereby obtaining target features;

[0043] The radar high-resolution range image feature extraction model is obtained by training the model using the training method described in the first aspect.

[0044] Fifthly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement steps in the training method of the model as described in the first aspect, or to implement steps in the radar high-resolution range image feature extraction method as described in the second aspect.

[0045] In a sixth aspect, embodiments of the present invention provide a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the model training method as described in the first aspect, or the steps in the radar high-resolution range image feature extraction method as described in the second aspect.

[0046] The beneficial effects of this invention are as follows: In this embodiment, firstly, N sets of training sample data are acquired; the model to be trained is trained based on the N sets of training sample data to obtain a loss value; then, the parameters of the model to be trained are adjusted based on the loss value to obtain the currently trained model; finally, if the loss value satisfies the loss convergence condition, the feature encoder network in the currently trained model is determined as the radar high-resolution range image feature extraction model. In this embodiment, the loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. By treating HRRP samples from the same observed target but under different azimuth views as positive examples, the model is forced to focus on the shared, essential target structure information behind the HRRP data from different views, while ignoring surface differences that change with the viewpoint, thereby learning features with viewpoint invariance and improving the feature extraction effect of HRRP data. Attached Figure Description

[0047] Appendix Figure 1 One of the flowcharts illustrating the training method for the model provided in this embodiment of the invention;

[0048] Appendix Figure 2 This is a schematic diagram of the structure of the model to be trained provided in an embodiment of the present invention;

[0049] Appendix Figure 3 A second schematic flowchart illustrating the training method for the model provided in this embodiment of the invention;

[0050] Appendix Figure 4 This is a flowchart illustrating the radar high-resolution range image feature extraction method provided in an embodiment of the present invention.

[0051] Appendix Figure 5a This is the result of directly performing dimensionality reduction visualization on the original HRRP data;

[0052] Appendix Figure 5b This is the result of dimensionality reduction visualization of the feature vectors obtained using the radar high-resolution range image feature extraction model;

[0053] Appendix Figure 6 A schematic diagram of the structure of a training device for a model provided in an embodiment of the present invention;

[0054] Appendix Figure 7 This is a schematic diagram of the radar high-resolution range image feature extraction device provided in an embodiment of the present invention;

[0055] Appendix Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0056] In the embodiments of this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first," "second," etc., in the specification of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are usually of the same class, without limiting the number of objects. For example, the first object can be one or multiple.

[0057] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0059] This application provides a model training method, a radar high-resolution range profile feature extraction method and apparatus, aiming to improve the accuracy of feature extraction from high-resolution range profile (HRRP) data.

[0060] Please see Figure 1 The training method for the model will be explained below. Figure 1 This is one of the flowcharts illustrating the training method for the model provided in this embodiment of the invention. The method specifically includes the following steps:

[0061] Step 101: Obtain N sets of training sample data. The nth set of training sample data includes high-resolution range image data of the nth observed target under M different azimuth views. N and M are both positive integers, and n is a positive integer less than or equal to N.

[0062] Step 102: Train the model to be trained based on the N sets of training sample data to obtain a loss value. The model to be trained includes a feature encoder network and a discriminant feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. The discriminant feature projection head is used to map the initial feature vector to obtain a discriminant feature vector.

[0063] Step 103: Adjust the parameters of the model to be trained based on the loss value to obtain the currently trained model.

[0064] Step 104: If the loss value satisfies the loss convergence condition, the feature encoder network in the currently trained model to be trained is determined as the radar high-resolution range image feature extraction model.

[0065] In this embodiment, HRRP data of N observed targets are acquired from M different azimuth viewpoints to construct a training sample dataset. Specifically, in some embodiments, HRRP data of N different observed targets are acquired. For example, when N=6 and M=10, six different categories of aircraft are acquired as observed targets. For each category of aircraft, HRRP data of the aircraft are acquired from 10 different azimuth viewpoints. The HRRP data of each category of aircraft from the 10 azimuth viewpoints is used as a set of sample data.

[0066] It should be understood that the specific method of dividing the azimuth viewpoint is not limited here. For example, in some embodiments, the viewpoint is divided according to different angles, and 0° to 180° is evenly divided into M azimuth viewpoints. Depending on the division method, the value of M can be 3, 10, 18, etc., and can be set and adjusted according to the actual situation.

[0067] As a specific implementation, in order to construct positive sample pairs in a self-supervised learning framework, it is necessary to simulate a multi-view observation scenario, and uniformly divide the continuous azimuth angle range from 0° to 180° into M non-overlapping discrete intervals (for example, when M=12, each interval is 15°). Each interval can be regarded as an independent virtual sensor observation view.

[0068] It should be understood that the method of acquiring training sample data is not limited here. For example, in some embodiments, training sample data is generated using electromagnetic simulation software, covering the complete azimuth range from 0° to 180° to ensure data comprehensiveness. In other embodiments, training sample data is acquired through real-time observation of multiple different observation targets.

[0069] In some embodiments, to eliminate interference from irrelevant variables of the signal on feature extraction, the originally acquired HRRP data needs to be preprocessed. This preprocessing includes at least one of energy normalization and centroid alignment. Specifically, energy normalization normalizes the total energy of each HRRP data point to 1 to eliminate amplitude variations caused by factors such as target range and radar transmit power. Centroid alignment calculates the centroid of each HRRP data point and aligns it to the center of the range window through cyclic shifting to eliminate sensitivity caused by target translation.

[0070] In this embodiment, the model to be trained includes a feature encoder network and a discriminative feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. Specifically, the feature encoder network is the core of feature extraction, used to extract features from the one-dimensional HRRP data. (Where the superscript m indicates that it comes from the m-th azimuth viewpoint) is mapped to a low-dimensional but information-dense initial feature vector. .

[0071] ;

[0072] in, This is used to characterize the feature encoder network, and the specific structure of the feature encoder network is not limited here.

[0073] For example, in some embodiments, the feature encoder network is a one-dimensional deep residual network (ResNet-1D). In this embodiment, by using ResNet-1D as the feature encoder network, its residual connection (Shortcut) structure effectively solves the gradient vanishing problem during the training process of deep networks, while also capturing the hierarchical features from local details to global structure in the HRRP signal.

[0074] In this embodiment, the N sets of training sample data include For each HRRP data point, input all HRRP data into the feature encoder network, and extract features from each HRRP data point to obtain the initial feature vector corresponding to each HRRP data point.

[0075] The discriminative feature projection head is used to map the initial feature vector to obtain the discriminative feature vector. Specifically, the discriminative feature projection head is used to map the initial feature vector... Further mapping to a latent space specifically used to compute the contrastive loss yields the discriminative feature vector. .

[0076] ;

[0077] in, The specific structure of the discriminative feature projection head is not limited here.

[0078] For example, in some embodiments, the discriminative feature projection head consists of one or more fully connected layers and uses nonlinear activation functions such as rectified linear units (ReLU), which are not specifically limited here.

[0079] In this embodiment, the initial feature vector corresponding to each HRRP data is mapped through a discriminant feature projection head to obtain the discriminant feature vector corresponding to each HRRP data, and the loss value is further calculated based on the discriminant feature vector.

[0080] In this embodiment, the feature representation is decoupled from the target task of contrastive learning by using a discriminative feature projection head, which helps to learn features with better generalization ability.

[0081] In this embodiment, the loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. The positive sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in the same group of training sample data, and the negative sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in different groups of training sample data.

[0082] For any anchor point data, HRRP data from other azimuth perspectives that are in the same training sample group as the anchor point data (i.e., from the same observation target) are all positive samples, while HRRP data from different training sample groups that are in different observation targets are all negative samples.

[0083] For example, for a single anchor point data (No. The observed target is at the first HRRP data from various perspectives). With the same observation target (the first) (each observation target) but from different azimuth perspectives of (No. The observed target is at the first HRRP data from each directional perspective constitute a positive sample pair. Other negative samples (excluding the first of N observed targets) HRRP data of other observation targets besides the observed target, such as (No. The observed target is at the first HRRP data from each directional perspective constitute a negative sample pair.

[0084] Instance-level loss is based on the core idea of ​​contrastive learning, aiming to learn viewpoint-invariant features. First, it defines any two discriminative feature vectors... and Cosine similarity between for:

[0085] ;

[0086] in, For vectors transpose, Used to characterize discriminative feature vectors The Euclidean norm, Used to characterize discriminative feature vectors The Euclidean norm.

[0087] In some embodiments, the loss function is calculated using the Information Noise-Contrastive Estimation (InfoNCE) loss function. and Instance loss of the positive sample pairs :

[0088] ;

[0089] in, It's a temperature over-parameter. It is an indicator function. For the first The observed target is at the first Discriminant feature vectors corresponding to sample data from each directional viewpoint For the first The observed target is at the first Discriminant feature vectors corresponding to sample data from each directional viewpoint For the first The observed target is at the first Discriminant feature vectors corresponding to sample data from each directional viewpoint For the first The observed target is at the first The discriminant feature vectors corresponding to sample data from each directional perspective.

[0090] In some embodiments, the total instance-level loss It is obtained by averaging over all possible positive sample pairs:

[0091] .

[0092] In this embodiment of the invention, the discriminant feature vectors corresponding to HRRP data from the same observation target but from different azimuth viewpoints are defined as positive sample pairs, and the discriminant feature vectors corresponding to HRRP data from different observation targets are defined as negative sample pairs. The loss is calculated by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs. Through the above method, the instance-level loss utilizes the idea of ​​contrastive learning to force HRRP features from the same observation target but from different viewpoints to move closer to each other, while features from different observation targets move further apart, thereby learning features with viewpoint invariance.

[0093] Optionally, in some embodiments, the loss value also includes a kernel discriminant loss, calculated as a weighted sum of the instance-level loss and the kernel discriminant loss. The kernel discriminant loss is based on kernel discriminant analysis (KDA) theory. In some embodiments, it is obtained using a Gaussian radial basis function. Mapping the discriminant feature vectors to a high-dimensional space, within-class divergence and inter-class divergence Defined as:

[0094] ;

[0095] ;

[0096]

[0097]

[0098] in, It is the first The mean of the discriminant feature vectors corresponding to HRRP data for each observed target under M different azimuth views. It is the global mean of the discriminant feature vectors corresponding to all HRRP data.

[0099] Optionally, in some embodiments, the kernel determines the loss. Defined as intra-class divergence and inter-class divergence The ratio:

[0100] .

[0101] It should be understood that kernel discriminant loss originates from statistical discriminant analysis theory. Its optimization goal is to maximize a global statistical discriminant criterion, namely the ratio of inter-class divergence to intra-class divergence of all samples in the entire feature space. It starts from the perspective of the overall data distribution and seeks a feature mapping that optimizes the cluster structure.

[0102] The kernel discriminant loss in this embodiment is a batch-based global optimization method that utilizes the feature information of all samples within a batch to jointly calculate the scatter matrix, thereby reshaping the feature space as a whole. Therefore, the kernel discriminant loss of this invention improves the discriminative power of features in a more macroscopic and global way, especially when dealing with complex nonlinear data distributions, where kernel tricks can achieve better cluster separation results.

[0103] Optionally, in some embodiments, the model to be trained further includes a semantic label projection head, which is used to perform semantic mapping on the initial feature vector to obtain a semantic label vector;

[0104] The loss value also includes semantic loss, which is determined based on the sum of squares of the differences between discriminant feature similarity and semantic feature similarity. Discriminant feature similarity is the similarity between discriminant feature vectors corresponding to any two high-resolution distance image data, and semantic similarity is the similarity between semantic label vectors corresponding to any two high-resolution distance image data.

[0105] It should be understood that, in this embodiment, the model to be trained consists of a feature encoder network and two parallel projection heads. Specifically, the two parallel projection heads are a discriminative feature projection head and a semantic label projection head. The semantic label projection head is set up in parallel with the discriminative feature projection head and also receives the initial feature vector. As input, it is mapped to a semantic space, and the output is a semantic label vector. :

[0106] ;

[0107] in, This is used to characterize the semantic tag projection head; the specific structure of the semantic tag projection head is not limited here.

[0108] For example, in some embodiments, the semantic tag projection head consists of one or more fully connected layers, the output layer of which employs a Softmax activation function, such that the output semantic tag vector... This forms a probability distribution. The category corresponding to the dimension with the largest value in the semantic label vector is considered the pseudo-label of the current sample data.

[0109] In this embodiment, the initial feature vector corresponding to each HRRP data is mapped through a semantic label projection head to obtain the semantic label vector corresponding to each HRRP data, and the loss value is further calculated based on the semantic label vector.

[0110] In this embodiment, the loss value includes not only instance-level loss and kernel-discriminative loss, but also semantic loss. The loss value is calculated based on the weighted sum of instance-level loss, kernel-discriminative loss, and semantic loss. The semantic loss uses a "self-distillation" mechanism to force the structure in the feature space to align with the semantic structure implied by the pseudo-label.

[0111] It should be understood that in the discriminative feature space, the first... The observed target in the azimuth view Discriminant feature vectors under , and the The observed target in the azimuth view Discriminant feature vectors under The discriminant feature similarity is In the semantic feature space, the first The observed target in the azimuth view semantic label vectors , and the The observed target in the azimuth view The following semantic label vectors The semantic feature similarity is .

[0112] As an optional implementation method, the feature similarity is determined as follows: The semantic feature similarity is All were calculated using cosine similarity:

[0113] ;

[0114] ;

[0115] Total semantic loss Defined as the mean of the sum of squares of the similarity differences between all sample pairs in the discriminative feature space and the semantic feature space:

[0116] .

[0117] In this embodiment, the similarity between the discriminative feature vectors of any two samples and the similarity between their corresponding semantic label vectors are calculated. By minimizing the sum of squares of the differences between these two similarities, the discriminative feature space and the semantic space are aligned. The semantic loss utilizes higher-level semantic information (pseudo-labels) generated by the network itself to supervise and correct the learning of discriminative features, ensuring that the learned features are not only separable at the instance level but also consistent at the category semantic level.

[0118] As a specific example, please refer to Figure 2 The model to be trained includes a feature encoder network, a discriminative feature projector head, and a semantic label projector head. The feature encoder network is used to map the preprocessed HRRP data into low-dimensional but information-dense initial feature vectors. In this embodiment, the feature encoder network is a one-dimensional deep residual network.

[0119] As a specific embodiment, the feature encoder network includes an input (In) layer, a convolutional (Conv) layer, a pooling layer, four residual block (ResBlock) layers, a global average pooling (GAP) layer, and a fully connected (FC) layer connected in sequence. The pooling layer performs max pooling. Each ResBlock includes two convolutional layers and a shortcut structure. Depending on the specific implementation, batch normalization (BN), ReLU, and / or dropout operations are added after different network layers. Specific additional operations and the detailed structure of each network layer can be found in the examples given in Table 1, and will not be elaborated further.

[0120] Table 1. Examples of feature encoder network structures

[0121]

[0122] The discriminative feature projection head and the semantic label projection head are configured in parallel and both are connected to the feature encoder network. The discriminative feature projection head is used to further map the initial feature vector output by the feature encoder network to a latent space (i.e., the discriminative feature space) specifically used to calculate the contrastive loss, thus obtaining the discriminative feature vector. In this embodiment, the discriminative feature projection head consists of at least two fully connected layers and uses nonlinear activation functions such as ReLU.

[0123] The semantic label projection head is used to further map the initial feature vector output by the feature encoder network to the semantic label space, outputting a semantic label vector. In this embodiment, the semantic label projection head consists of at least two fully connected layers, and the last layer of the semantic label projection head is connected to a Softmax activation function to generate the probability distribution of the target category as a pseudo-label.

[0124] like Figure 3 As shown, the continuous azimuth range from 0° to 180° is uniformly divided into 12 non-overlapping discrete intervals, each interval being 15°. Each interval can be regarded as an independent virtual sensor observation perspective. That is, in this embodiment, the value of M is 12.

[0125] HRRP data for N observed targets are acquired from 12 different azimuth viewpoints. The HRRP data from each of the 12 different azimuth viewpoints for each observed target are used as a set of training data samples. Energy normalization and centroid alignment are performed on all HRRP data to obtain preprocessed HRRP data. This preprocessed HRRP data is then input into a feature encoder network for feature extraction. The initial feature vector output by the feature encoder network is then input into a discriminant feature projection head and a semantic label projection head for further processing to obtain discriminant feature vectors and semantic label vectors.

[0126] Instance-level loss is calculated based on discriminative feature vectors and semantic label vectors. , and determine the loss and semantic loss The specific calculation method can be found in the preceding description, and will not be repeated here. Based on instance-level loss. , and determine the loss and semantic loss Determine the loss value obtained in this training. :

[0127] ;

[0128] in, , and The non-negative weighting coefficients used to balance the contributions of various losses are predetermined and their specific values ​​are not limited here. These loss values ​​are used to collaboratively optimize network parameters from three different dimensions.

[0129] In practice, the pre-acquired multi-view HRRP dataset is input into the pre-built training model in batches for multiple iterative training iterations. In each training iteration, forward propagation is performed to obtain the loss value corresponding to the current training. Subsequently, the gradient of the loss function with respect to the parameters of each layer of the network is calculated using the backpropagation algorithm, and the network weights are updated using optimizers such as Adam. The training process continues, and when the loss value meets the loss convergence condition (e.g., model convergence or reaching a preset number of training rounds), the feature encoder network in the currently trained training model is determined as the radar high-resolution range image feature extraction model.

[0130] It should be understood that after training, the discriminative feature projection head and semantic label projection head can be discarded, retaining only the trained feature encoder network. At this point, the feature encoder network already possesses powerful feature extraction capabilities. For any new HRRP data to be tested, a single forward propagation through the feature encoder network is sufficient to obtain its high-quality final feature vector, rich in discriminative information.

[0131] In this embodiment, instance-level loss utilizes the idea of ​​contrastive learning to force HRRP features from the same observed target but different perspectives to be closer to each other, while features from different targets are further apart, thereby learning features with viewpoint invariance. By setting positive and negative sample pairs, the model to be trained can be forced to focus on the shared, essential target structure information behind HRRP data from different perspectives, while ignoring surface differences that change with viewpoint, thus learning features with viewpoint invariance. This goal of "learning viewpoint invariance" is essential and crucial in the HRRP field. Kernel discriminative loss, from the perspective of global distribution, further enhances the inter-class separability and intra-class compactness of features through nonlinear mapping. Semantic loss, by constructing pseudo-labels, constrains the similarity structure of features to be consistent with the semantic category structure of the target, achieving feature-semantic alignment.

[0132] This invention fundamentally changes the source and quantity of supervisory signals by introducing a self-supervised contrastive learning paradigm, thereby solving the problems of model underfitting and lack of interpretability in the model decision-making process caused by insufficient training samples. Specifically, the instance-level loss of this invention does not rely on manually labeled data such as "Aircraft A" or "Aircraft B," but its supervisory signals originate from the data itself—"HRRP samples from the same physical target should have similar characteristics even if the perspectives are different." For a given set of training samples... For the target of discrete perspective observation data, traditional supervised learning can only provide... training samples ( (Secondary supervision). The comparative learning method of this invention can be used in this... Constructing from a sample There are 100 positive sample pairs. This means that the number of supervisory signals has a linear relationship with the number of samples ( ). Transformed into a quadratic relationship of sample numbers ( By exponentially amplifying this intrinsic supervision signal, this invention cleverly transforms the problem of "sparse labeled samples" into a training task with "abundant training relation pairs (positive and negative sample pairs)". Even with a small amount of original data, the network can receive a large number of constraint gradients regarding feature invariance in each iteration, thus achieving sufficient training. This effectively avoids underfitting caused by insufficient supervision information and ensures that the deep model can still learn robust and highly discriminative features even under small sample conditions.

[0133] Please see Figure 4 This invention also provides a method for extracting features from high-resolution radar range images, the method specifically including the following steps:

[0134] Step 401: Obtain the high-resolution range image data to be processed.

[0135] Step 402: Input the high-resolution range image data to be processed into the radar high-resolution range image feature extraction model to extract features and obtain target features.

[0136] The radar high-resolution range image feature extraction model is obtained by training the model using the training method described above.

[0137] It should be understood that the radar high-resolution range image feature extraction model trained by the training method of the model provided in the foregoing embodiments is applied to the radar high-resolution range image feature extraction method provided in the foregoing embodiments. The structure of the radar high-resolution range image feature extraction model can be referred to the relevant description of the feature encoder network in the foregoing embodiments, and will not be repeated here to avoid repetition. Therefore, the specific process of the radar high-resolution range image feature extraction model processing HPPR data can be referred to the relevant process of the feature encoder network processing HPPR data in the foregoing embodiments, and will not be repeated here to avoid repetition.

[0138] In this embodiment of the invention, during training using the above-described model training method, end-to-end backpropagation training is performed under the supervision of multiple loss functions to continuously optimize the parameters in the model to be trained. After training, the high-performance radar high-resolution range image feature extraction model is retained for efficient and robust feature extraction of new HRRP data, providing high-quality input for subsequent classification, recognition, and other tasks.

[0139] The following uses specific experimental results as examples to illustrate the beneficial effects of the model training method and the radar high-resolution range image feature extraction method provided in this embodiment of the invention. To intuitively demonstrate the effectiveness of the method of this invention, t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction visualization technology is used to analyze the feature space in the experiment.

[0140] like Figure 5a As shown, this is the result of dimensionality reduction visualization of the original HRRP data. It can be seen that the sample points of different observation targets (represented by different colors) are severely mixed in the feature space, failing to form any meaningful clustering structure. This indicates that the original data is difficult to use directly for classification.

[0141] like Figure 5bAs shown, this is the result of dimensionality reduction visualization of the feature vectors obtained by using the feature encoder trained by the model training method provided in this invention as a feature extraction model for radar high-resolution range images. It can be observed that the feature space presents a very clear structure: sample points of the same observed target are closely clustered together, forming independent clusters; while clusters between different observed targets are significantly separated with clear boundaries.

[0142] This comparative result strongly demonstrates that the method proposed in this invention can effectively learn feature representations with strong discriminativeness and robustness, laying a solid foundation for subsequent high-precision target recognition tasks.

[0143] like Figure 6 As shown, this embodiment of the invention also provides a model training device 600, comprising:

[0144] The first acquisition module 601 is used to acquire N sets of training sample data. The nth set of training sample data includes high-resolution range image data of the nth observed target under M different azimuth views. N and M are both positive integers, and n is a positive integer less than or equal to N.

[0145] Training module 602 is used to train the model to be trained based on N sets of training sample data to obtain a loss value. The model to be trained includes a feature encoder network and a discriminant feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. The discriminant feature projection head is used to map the initial feature vector to obtain a discriminant feature vector.

[0146] The parameter adjustment module 603 is used to adjust the parameters of the model to be trained based on the loss value to obtain the currently trained model to be trained.

[0147] The determination module 604 is used to determine the feature encoder network in the currently trained model to be trained as a radar high-resolution range image feature extraction model when the loss value satisfies the loss convergence condition.

[0148] The loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. The positive sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in the same group of training sample data, and the negative sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in different groups of training sample data.

[0149] Optionally, the instance-level loss for:

[0150] ;

[0151] ;

[0152] in, For temperature hyperparameters, For indicator functions, For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth perspectives.

[0153] Optionally, the loss value further includes kernel discriminant loss, the kernel discriminant loss for:

[0154] ;

[0155] ;

[0156] ;

[0157] in, Let Gaussian radial basis kernel function be used. For the first The mean of the discriminant feature vectors corresponding to the high-resolution range image data of each observed target. The mean of the discriminant feature vectors corresponding to all the high-resolution distance image data is given. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth perspectives.

[0158] Optionally, the model to be trained further includes a semantic label projection head, which is used to perform semantic mapping on the initial feature vector to obtain a semantic label vector;

[0159] The loss value also includes semantic loss, which is determined based on the sum of squares of the differences between discriminant feature similarity and semantic feature similarity. Discriminant feature similarity is the similarity between discriminant feature vectors corresponding to any two high-resolution distance image data, and semantic similarity is the similarity between semantic label vectors corresponding to any two high-resolution distance image data.

[0160] Optionally, the semantic loss for:

[0161] ;

[0162] ;

[0163] ;

[0164] in, For the first The observed target is at the first Semantic label vectors corresponding to high-resolution distance image data from each azimuth viewpoint. For the first The observed target is at the first Semantic label vectors corresponding to high-resolution distance image data from each azimuth viewpoint. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth perspectives.

[0165] The training device 600 for the model provided in this application embodiment can execute the above-described model training method embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0166] like Figure 7 As shown, this embodiment of the invention also provides a radar high-resolution range image feature extraction device 700, comprising:

[0167] The second acquisition module 701 is used to acquire high-resolution range image data to be processed;

[0168] Radar high-resolution range image feature extraction model 702 is used to extract features from the high-resolution range image data to be processed to obtain target features;

[0169] The radar high-resolution range image feature extraction model 702 is obtained by training the model using the training method described above.

[0170] The radar high-resolution range image feature extraction device 700 provided in this application embodiment can execute the above-described radar high-resolution range image feature extraction method embodiment. Its implementation principle and technical effect are similar, and will not be described again here.

[0171] It should be noted that the division of units in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0172] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0173] like Figure 8 As shown, this application embodiment provides an electronic device 800, including: a memory 802, a processor 801, and a program stored in the memory 802 and executable on the processor 801; the processor 801 is used to read the program in the memory 802 to implement the steps in the training method of the model as described above, or to implement the steps in the radar high-resolution range image feature extraction method as described above.

[0174] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the various processes of the training methods for the model described above, or the various processes of the radar high-resolution range image feature extraction methods described above, and achieves the same technical effect. To avoid repetition, further details are omitted here. The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as compact disks (CDs), digital video discs (DVDs), Blu-ray discs (BDs), high-definition universal discs (HVDs), etc.), and semiconductor storage (such as read-only memory (ROMs), erasable programmable read-only memory (EPROMs), electrically erasable programmable read-only memory (EEPROMs), non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0175] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0177] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this application without departing from its spirit, and all of these modifications are within the scope of protection of this application.

Claims

1. A method for training a model, characterized in that, include: Obtain N sets of training sample data. The nth set of training sample data includes high-resolution range image data of the nth observed target under M different azimuth views. N and M are both positive integers, and n is a positive integer less than or equal to N. The training model is trained based on N sets of training sample data to obtain a loss value. The training model includes a feature encoder network and a discriminant feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. The discriminant feature projection head is used to map the initial feature vector to obtain a discriminant feature vector. Based on the loss value, the parameters of the model to be trained are adjusted to obtain the currently trained model; If the loss value satisfies the loss convergence condition, the feature encoder network in the currently trained model to be trained is determined as the radar high-resolution range image feature extraction model. The loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. The positive sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in the same group of training sample data, and the negative sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in different groups of training sample data. Among them, the instance-level loss for: ； ； in, For temperature hyperparameters, For indicator functions, For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. for and The instance loss of the positive sample pairs constituted. Used to calculate cosine similarity.

2. The method as described in claim 1, characterized in that, The loss value also includes kernel discriminant loss, the kernel discriminant loss for: ； ； ； in, Let Gaussian radial basis kernel function be used. For the first The mean of the discriminant feature vectors corresponding to the high-resolution range image data of each observed target. The mean of the discriminant feature vectors corresponding to all the high-resolution distance image data is given. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For intra-class divergence, This represents the inter-class divergence.

3. The method as described in claim 1 or 2, characterized in that, The model to be trained also includes a semantic label projection head, which is used to perform semantic mapping on the initial feature vector to obtain a semantic label vector. The loss value also includes semantic loss, which is determined based on the sum of squares of the differences between discriminant feature similarity and semantic feature similarity. Discriminant feature similarity is the similarity between discriminant feature vectors corresponding to any two high-resolution distance image data, and semantic similarity is the similarity between semantic label vectors corresponding to any two high-resolution distance image data.

4. The method as described in claim 3, characterized in that, The semantic loss for: ； ； ； in, For the first The observed target is at the first Semantic label vectors corresponding to high-resolution distance image data from each azimuth viewpoint. For the first The observed target is at the first Semantic label vectors corresponding to high-resolution distance image data from each azimuth viewpoint. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. for and The similarity of discriminative features, for and semantic feature similarity Used to calculate cosine similarity.

5. A method for extracting features from high-resolution radar range images, characterized in that, include: Acquire high-resolution range image data to be processed; The high-resolution range image data to be processed is input into the radar high-resolution range image feature extraction model for feature extraction to obtain target features; The radar high-resolution range image feature extraction model is obtained by training the model using the training method described in any one of claims 1 to 4.

6. A training device for a model, characterized in that, include: The first acquisition module is used to acquire N sets of training sample data. The nth set of training sample data includes high-resolution range image data of the nth observed target under M different azimuth views. N and M are both positive integers, and n is a positive integer less than or equal to N. The training module is used to train the model to be trained based on N sets of training sample data to obtain a loss value. The model to be trained includes a feature encoder network and a discriminant feature projection head. The feature encoder network is used to extract features from the high-resolution distance image data to obtain an initial feature vector. The discriminant feature projection head is used to map the initial feature vector to obtain a discriminant feature vector. The parameter adjustment module is used to adjust the parameters of the model to be trained based on the loss value, so as to obtain the currently trained model to be trained. The determination module is used to determine the feature encoder network in the currently trained model to be trained as a radar high-resolution range image feature extraction model when the loss value satisfies the loss convergence condition. The loss value includes instance-level loss, which is determined based on the similarity between positive sample pairs and the similarity between negative sample pairs. The positive sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in the same group of training sample data, and the negative sample pair is the discriminant feature vector corresponding to any two high-resolution distance image data in different groups of training sample data. Among them, the instance-level loss for: ； ； in, For temperature hyperparameters, For indicator functions, For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. For the first The observed target is at the first Discriminant feature vectors corresponding to high-resolution distance image data from various azimuth views. for and The instance loss of the positive sample pairs constituted. Used to calculate cosine similarity.

7. A radar high-resolution range image feature extraction device, characterized in that, include: The second acquisition module is used to acquire the high-resolution range image data to be processed. A radar high-resolution range image feature extraction model is used to extract features from the high-resolution range image data to be processed, thereby obtaining target features; The radar high-resolution range image feature extraction model is obtained by training the model using the training method described in any one of claims 1 to 4.

8. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; characterized in that the processor is configured to read the program from the memory to implement the steps in the training method of the model as described in any one of claims 1 to 4, or to implement the steps in the radar high-resolution range image feature extraction method as described in claim 5.

9. A readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps in the training method of the model as described in any one of claims 1 to 4, or the steps in the radar high-resolution range image feature extraction method as described in claim 5.

Images