A radar target detection model training method based on incomplete label data

Abstract

This invention discloses a radar target detection model training method based on incomplete labeled data, proposing a learning framework based on dual-view collaboration and translation consistency. First, the radar echo signal is randomly cropped to form two sub-echo segments with partial overlap and translational relationships. Then, deep features are extracted using a Siamese Transformer network to predict the target probability at each range cell. After generating initial labels using a pseudo-label strategy, consistent and accurate mapping of pseudo-labels between sub-echoes is achieved based on equivariant translational characteristics. Finally, the model parameters are optimized by calculating the cross-entropy loss between the target probability and the pseudo-labels. This method solves the problem that existing deep learning detection models are highly dependent on and difficult to obtain accurate range cell-level labels in engineering. Relying only on incomplete labeled data, it improves reliability by supplementing fine-grained annotations with pseudo-labels and injecting translational consistency physical constraints, effectively overcoming training difficulties and showing promising application prospects.

Description

Technical Field

[0001] This invention belongs to the field of radar target detection technology, and specifically relates to a training method for a radar target detection model based on incomplete label data. Background Technology

[0002] Radar, with its advantages of long detection range and strong all-weather operation, plays an irreplaceable role in search and rescue, surveillance, and national defense. However, traditional radar target detection methods based on statistical models heavily rely on pre-set clutter distribution models. Once the actual clutter background does not match the pre-set model, the detection performance will drop sharply, leading to serious false alarms and missed target detections, making it difficult to meet the application requirements in complex environments.

[0003] In recent years, deep learning has been widely used to create radar target detection models due to its automatic feature extraction and pattern fitting capabilities. The paper "Marine target detection via spatial–temporal graph neural network, IEEE Transactions on Instrumentation and Measurement, 74 (2025): 1-14" proposes a spatiotemporal graph neural network detector for efficient detection. It can directly extract spatiotemporal features from radar without data preprocessing and uses these extracted features for target detection, achieving performance superior to traditional detectors. The paper "Maritime radar target detection in sea clutter based on CNN with dual-perspective attention, IEEE Geoscience and Remote Sensing Letters, 20 (2022): 1-5" proposes a radar target detection method based on convolutional neural networks and dual-perspective attention. It extracts relevant features from global and local perspectives through an attention mechanism, achieving excellent detection performance.

[0004] However, the superior performance of the aforementioned deep learning methods highly depends on supervised training with precise distance cell-level labels. Such fine-grained labeling is costly and extremely difficult to obtain in engineering, severely restricting the practical deployment of the models. In contrast, obtaining incomplete labels that only indicate the presence or absence of a target within a certain distance window is much more convenient. Therefore, researching a method to train a high-performance detection model on incomplete labeled data, reducing the reliance on fine-grained labeling, has significant theoretical implications and application prospects for improving the engineering practicality and promotional value of radar intelligent detection models. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a radar target detection model training method based on incomplete label data, which can train an intelligent detection model with excellent performance on radar echo data with only incomplete labels.

[0006] The objective of this invention is achieved through the following technical solution: a radar target detection model training method based on incomplete label data, comprising the following steps:

[0007] S1. Preprocess the received radar echo signal, and construct training and validation datasets based on the preprocessed radar echo signal and the corresponding incomplete labels.

[0008] The radar transmits a linear frequency modulated (LFM) signal for target detection. The echo signal is down-converted, matched-filtered, and sampled to obtain the baseband echo vector. The radar echo provides an incomplete label, specifically a range window-level label. : This indicates that a target exists within the distance window. This indicates that the target is not included; for items with window-level annotations... , length is The echo signal within the distance window expands to the left and right centered on that distance window. The label is obtained by cropping the distance unit. Corresponding training samples ;in The index of the distance window indicates the nth element in the baseband signal. Each distance window signal; map all training samples to their labels to obtain the training dataset;

[0009] When creating the validation dataset, we used fine-grained distance cell level prepared labels, that is, labels indicating whether the target exists within each distance cell;

[0010] S2. Construct a twin Transformer model for radar target detection; the twin Transformer consists of a dual-view echo random pruning module and two structurally identical branches, each branch containing a feature encoding module, a position encoding module, and... A cascaded Transformer module and a classifier, with parameters shared between the two branches;

[0011] For the input echo signal The dual-view echo random cropping module in Two echo segments are randomly cut, and each echo signal contains Each distance unit, and the two echo signals are denoted as follows: and ,in Because there is a clear relative translation relationship between the two clipped echoes, therefore... and Two echo perspectives considered as the same potential target; for and The real and imaginary parts of the echo signal are stacked as two channels respectively to obtain... This is used as the input sample pair for the two branches, and the class probabilities output by the two branches are denoted as follows: and ;

[0012] S3. Construct a loss function based on translational isovariance; based on incomplete labels at the distance window level. Constructing fine-grained distance cell-level pseudo-labels :when season ;when At that time, The distance cell with the highest probability is considered the target location, and the remaining distance cells are considered non-target locations. The following pseudo-labels are constructed:

[0013] ;

[0014] for The Middle The probability of a distance unit;

[0015] Subsequently, utilizing the clipping and translation relationship between the two echo viewpoints, a pseudo-label aligned with the pseudo-label of echo viewpoint 1 is constructed and used as the pseudo-label of echo viewpoint 2. ;

[0016] Based on the above pseudo-labels, the loss function is constructed as follows:

[0017] ;

[0018] in, For the binary cross-entropy loss per distance unit;

[0019] S4. Train the Siamese Transformer model using the training dataset and loss function; and obtain the final parameters of the Siamese Transformer model using the validation dataset: select the set of parameters with the highest accuracy on the validation dataset as the final parameters of the model;

[0020] S5. After pulse compression and range window segmentation of the radar echo, the signal is fed into the trained twin Transformer model to obtain the detection results.

[0021] The beneficial effects of this invention are as follows: The radar target detection model training method based on incomplete label data can effectively detect radar echoes with only incomplete labels; the invention extracts deep features from sub-echo signals using a Siamese Transformer network, enhancing the network's ability to extract deep semantic information from radar echoes; the dual-view echo random pruning module generates sub-echo signals with translational characteristics, increasing data diversity while injecting equivariant translational consistency constraints; the pseudo-label generation strategy and equivariant translational consistency constraints of this invention improve the accuracy of pseudo-labels, enabling the model to achieve complete-level supervised training using incomplete labels. This invention's method can train the model without requiring precise range cell-level labels, enabling high-precision radar target detection, while significantly reducing the cost and difficulty of acquiring labeled data in engineering practice. Attached Figure Description

[0022] Figure 1 This is a flowchart of a radar target detection model training method based on incomplete label data according to the present invention.

[0023] Figure 2 This is a diagram illustrating the structure of the radar intelligent target detection model and its training process in an embodiment of the present invention.

[0024] Figure 3 This is a diagram illustrating the structure of the radar intelligent target detection model in this embodiment of the invention and the workflow for obtaining detection results.

[0025] Figure 4 This is a diagram showing the detection results of radar echo signals in an embodiment of the present invention, with and without using the method of the present invention.

[0026] Figure 5 The figures show the detection performance curves of the method of the present invention and other methods under radar echo signals of different target intensities in this embodiment of the invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0028] like Figure 1 As shown, the present invention provides a radar target detection model training method based on incomplete label data, comprising the following steps:

[0029] S1. Preprocess the received radar echo signal, and construct training and validation datasets based on the preprocessed radar echo signal and the corresponding incomplete labels.

[0030] Radar transmits linear frequency modulated signals for target detection, represented as:

[0031] ;

[0032] in For signal carrier frequency, The frequency modulation frequency of a linear frequency modulation signal. This represents the pulse width.

[0033] When the point target is at a distance of At that time, the echo signal is:

[0034] ;

[0035] in , At the speed of light, The coefficient for target reflection and propagation attenuation. For clutter, It is additive white Gaussian noise;

[0036] The baseband echo vector is obtained after down-conversion, matched filtering, and sampling of the echo signal. ,in The number of distance units, Represents the complex field. The target echo is mainly concentrated in the first... One distance unit, Sampling rate, Indicates rounding up;

[0037] The labels provided by the radar echo are incomplete; specifically, they are range window level labels. : This indicates that a target exists within the distance window. Indicates no target; for a distance cell number of The baseband echo vector, with the selected range window length being... At that time, it can be cut out Non-overlapping distance windows. For windows with window-level annotations. , length is The echo signal within the distance window expands to the left and right centered on that distance window. The label is obtained by cropping the distance unit. Corresponding training samples ;in The index of the distance window indicates the nth element in the baseband signal. Each distance window signal is used to map all training samples to their corresponding labels, thus obtaining the training dataset. The premise of this invention is that each labeled distance window contains at most one target, and the distances to the left and right of the window are... There is no target within the distance cell.

[0038] When creating the validation dataset, since its purpose is to test the performance of the detection model, it is prepared with fine-grained distance cell level labels, that is, labels indicating whether the target exists within each distance cell.

[0039] In this example implementation, the radar maritime detection dataset from the first issue of the Journal of Radar Science in 2022 was selected to create the training and validation datasets according to the requirements.

[0040] S2. Construct a twin Transformer model for radar target detection; such as... Figure 2 As shown, the twin Transformer consists of a dual-view echo random cropping module and two structurally identical branches. Each branch sequentially contains a feature encoding module, a position encoding module, and... A cascaded Transformer module and a classifier, with parameters shared between the two branches;

[0041] For the input echo signal The dual-view echo random cropping module in Two echo segments are randomly cut, and each echo signal contains Each distance unit, and the two echo signals are denoted as follows: and ,in This makes each length of All cropped samples can be made to a length of The label already has incomplete distance window tags. The echo signal is included, thus ensuring and Inheriting the same tags Because there is a clear relative translation relationship between the two clipped echoes, it is possible to... and Two echo perspectives considered as the same potential target. For and The real and imaginary parts of the echo signal are stacked as two channels respectively to obtain... , Let represent the real number field; and use it as a pair of input samples, inputting them into the two branches respectively; the class probabilities output by the two branches are denoted as . and .

[0042] Input from one side of the branch For example, among which To determine the number of range cells in the input echo, the feature encoding module first employs a single convolutional layer to encode the input radar echo data into feature representations. ,in This represents the number of feature channels. Since the Transformer module is insensitive to the absolute position of the input features, while positional information in the range dimension is crucial in radar target detection, the position encoding module of this invention enables the model to utilize positional information between features for modeling. Specifically, it encodes the position sequence... Mapped to position encoding via a fully connected layer (FC). and with Adding them together yields the input features with location information:

[0043] ;

[0044] in This indicates a convolutional layer.

[0045] Subsequently, the Transformer module... As input, more discriminative deep features are extracted. Each Transformer module consists of a layer normalization (Norm), a multi-head attention module (MHA), and a feedforward network (FFN), and residual connections are used to stabilize training.

[0046] ;

[0047] Norm normalizes features to improve model training stability and reduce the impact of clutter-induced feature amplitude fluctuations on model convergence and detection decisions; MHA integrates feature information from different distance units across the entire distance dimension, modeling the target's response patterns in neighboring units and their relationship with surrounding clutter; FFN enhances the separability of the target and clutter in the feature space through nonlinear transformation. Layer normalization affects the input feature vector. The calculation formula is as follows:

[0048] ;

[0049] in, For feature dimensions; and These are the mean and variance of the feature, respectively; Minimum constant (e.g.) (), used to prevent the denominator from being zero; and These are learnable affine transformation parameters; This represents element-wise multiplication. The Multi-Head Attention Module (MHA) captures feature dependencies in different subspaces in parallel using multiple attention heads.

[0050] ;

[0051] in The dimension of the input vector; For the first The linear projection matrix of the head; For the number of heads; This is the output projection matrix after multi-head splicing; The function normalizes the attention weights to a probability distribution. FFN consists of two linear transformations and a nonlinear activation function, used to enhance the model's nonlinear expressiveness. Its formula is expressed as:

[0052] ;

[0053] in, The weights and biases are for the first-level linear transformation; The weights and biases for the second-level linear transformation; For non-linear activation functions:

[0054] ;

[0055] in, It is the cumulative distribution function of the standard normal distribution.

[0056] The classifier processes the deep features output by the last Transformer module. Linear decision-making is performed unit-by-unit, and the decision score is mapped to the posterior probability using the sigmoid function. The process is as follows:

[0057]

[0058] in Indicates the first Each distance cell contains the posterior probability of the target. For the two echo viewpoints of the twin Transformer input... The output class probabilities are denoted as follows: and .

[0059] S3. Construct a loss function based on translational isovariance; obtain the posterior probability of the Siamese Transformer for the input echo viewpoint 1. Subsequently, the adaptive pseudo-label generation strategy of the present invention is based on incomplete labels at the distance window level. Constructing fine-grained distance cell-level pseudo-labels :when season ;when At that time, The distance cell with the highest probability is considered the target location, and the remaining distance cells are considered non-target locations. The following pseudo-labels are constructed:

[0060] ;

[0061] for The Middle The probability of a distance unit;

[0062] Subsequently, utilizing the clipping and translation relationship between the two echo viewpoints, a pseudo-label aligned with the pseudo-label of echo viewpoint 1 is constructed and used as the pseudo-label of echo viewpoint 2. The specific method is as follows: Let the clipping offset between echo viewpoint 1 and viewpoint 2 on the distance cell be... And define the translation operator. for:

[0063] ;

[0064] The pseudo-label for echo viewpoint 2 is: ; This indicates that the pseudo-labels obtained earlier will be used. Translation distance units, i.e. In the The label on the distance unit became in Labels on each distance cell. This construction explicitly injects an equivariant prior on the translation of the echo in the distance dimension: when the input echo is translated in the distance dimension, the model's detection result should produce the same translation, thus ensuring physical consistency between the two perspectives.

[0065] Based on the above pseudo-labels, the loss function is constructed as follows:

[0066] ;

[0067] in, For the binary cross-entropy loss per distance unit:

[0068] ;

[0069] By constraining the prediction results of the two perspectives to be consistent with their translation-aligned pseudo-labels, and by performing equal-weighted fusion of the two losses, the model can not only learn the distance-by-distance unit response under a single perspective, but also enhance the translation consistency across perspectives, thereby improving the ability to detect targets.

[0070] S4. Train the Siamese Transformer model constructed in step S2 using the training dataset constructed in step S1 and the loss function constructed in step S3; and obtain the final parameters of the Siamese Transformer model through the validation dataset: select the set of parameters with the highest accuracy on the validation dataset as the final parameters of the model.

[0071] S5. After pulse compression and range window segmentation of the radar echo, the signal is fed into the trained twin Transformer model to obtain the detection results.

[0072] like Figure 3 As shown, during the inference phase of acquiring the model's detection results, only a single branch of the Siamese network is retained as the target detector. Unlike the dual-view random cropping of the input echoes during the training phase, the echoes are cropped along the distance cells at lengths of [missing information]. The sliding window is cropped, and the resulting samples are sequentially input into the model to obtain the posterior probability vector of each distance unit within the corresponding window. Subsequently, a threshold was adopted. Make a decision for each distance cell and output the detection result vector. :

[0073]

[0074] in This is an indicator function that returns 1 only if the condition within the parentheses is met; otherwise, it returns 0.

[0075] In this embodiment, the file "20221114050022_stare_HH.mat" from the maritime detection dataset published in the first issue of the Journal of Radar in 2022 is selected to verify the detection performance of the present invention. Figure 4 (a) Visualization of the radar echo and the location of the target; Figure 4 (b) is the detection result of the present invention; Figure 4 (c) The results of detection using the traditional statistical model-based OS-CFAR method; Figure 4 (d) shows the detection results of Transformer trained with precise distance labels. The comparison results show that, under the condition of training with incomplete label data, the detection performance of the present invention is not only better than that of traditional statistical model-based detection methods, but also reaches a similar performance level to that of models trained with precise distance unit-level labels, effectively verifying the superiority of the present invention.

[0076] To conduct a detailed evaluation of the detection performance of this method over a wider signal-to-noise ratio (SCR) range, we extended the target SCR to 0–11 dB based on the "20221114050022_stare_HH.mat" file, and expanded the traditional statistical model-based CFAR methods for comparison to CA-CFAR, OS-CFAR, SO-CFAR, and GO-CFAR. The corresponding detection probability versus false alarm probability curves are shown below. Figure 5 As shown, Figure 5 (a) is the detection probability curve. Figure 5 (b) is the probability curve of Xu Jing. Figure 5 The boundary performance in this context refers to the detection performance of a model trained using precise distance unit-level labels. Experimental results show that across the entire SCR range, the method of this invention maintains a low false alarm rate while consistently achieving a significantly higher detection probability than various CFAR algorithms. Compared to detectors trained with fine annotations, the method of this invention still lags behind in the low SCR range of 0–2 dB, but when the SCR reaches 3 dB and above, the detection performance of both methods becomes very close.

[0077] In summary, the method of the present invention can achieve detection performance similar to that of training with accurate distance unit level label data when using incomplete distance window level label data for training.

[0078] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for training a radar target detection model based on incomplete label data, characterized in that, Includes the following steps: S1. Preprocess the received radar echo signal, and construct training and validation datasets based on the preprocessed radar echo signal and the corresponding incomplete labels. The radar transmits a linear frequency modulated (LFM) signal for target detection. The echo signal is down-converted, matched-filtered, and sampled to obtain the baseband echo vector. The radar echo provides an incomplete label, specifically a range window-level label. : This indicates that a target exists within the distance window. This indicates that the target is not included; for items with window-level annotations... , length is The echo signal within the distance window expands to the left and right centered on that distance window. The label is obtained by cropping the distance unit. Corresponding training samples ;in The index of the distance window indicates the nth element in the baseband signal. Each distance window signal; map all training samples to their labels to obtain the training dataset; When creating the validation dataset, we used fine-grained distance cell level prepared labels, that is, labels indicating whether the target exists within each distance cell; S2. Construct a twin Transformer model for radar target detection; the twin Transformer consists of a dual-view echo random pruning module and two structurally identical branches, each branch containing a feature encoding module, a position encoding module, and... A cascaded Transformer module and a classifier, with parameters shared between the two branches; For the input echo signal The dual-view echo random cropping module in Two echo segments are randomly cut, and each echo signal contains Each distance unit, and the two echo signals are denoted as follows: and ,in Because there is a clear relative translation relationship between the two clipped echoes, therefore... and Two echo perspectives considered as the same potential target; for and The real and imaginary parts of the echo signal are stacked as two channels respectively to obtain... The input consists of two branches; the class probabilities output by the two branches are denoted as follows: and ; S3. Construct a loss function based on translational isovariance; based on incomplete labels at the distance window level. Constructing fine-grained distance cell-level pseudo-labels :when season ;when At that time, The distance cell with the highest probability is considered the target location, and the remaining distance cells are considered non-target locations. The following pseudo-labels are constructed: ； for The Middle The probability of a distance unit; Subsequently, utilizing the clipping and translation relationship between the two echo viewpoints, a pseudo-label aligned with the pseudo-label of echo viewpoint 1 is constructed and used as the pseudo-label of echo viewpoint 2. ; Based on the above pseudo-labels, the loss function is constructed as follows: ； in, For the binary cross-entropy loss per distance unit; S4. Train the Siamese Transformer model using the training dataset and loss function; and obtain the final parameters of the Siamese Transformer model using the validation dataset: select the set of parameters with the highest accuracy on the validation dataset as the final parameters of the model; S5. After pulse compression and range window segmentation of the radar echo, the signal is fed into the trained twin Transformer model to obtain the detection results.

2. The radar target detection model training method based on incomplete label data according to claim 1, characterized in that, The feature encoding module is implemented using a single convolutional layer; The classifier performs linear decision-making on the deep features output by the last Transformer module on a distance-by-distance basis, and uses the sigmoid function to map the decision scores to posterior probabilities.

Images