Radar target detection method based on contrast self-supervised learning

Abstract

The invention provides a radar target detection method based on contrast self-supervised learning, and the method comprises the steps: S1, data preparation: carrying out the preprocessing of an echo signal collected by a radar receiver, obtaining an R-D image corresponding to the signal, and carrying out the data augmentation operation, and obtaining an augmented R-D image; s2, building a self-supervised pre-training model: building a self-supervised comparative learning pre-training network based on a large number of unlabeled R-D graphs and augmented R-D graphs, and training to obtain a basic model; and S3, a supervised fine tuning model: inputting a small amount of target detection annotation data into the pre-training basic model, carrying out supervised fine tuning training to obtain a radar target detection model, adopting a pre-training-fine tuning two-stage training normal form, and fully utilizing a large amount of label-free data to carry out pre-training based on self-supervised learning to obtain a radar target detection model. Supervised fine tuning is carried out on a small number of labeled samples, dependence on a large amount of high-quality labeled data is effectively relieved, and radar target detection performance in a complex environment is further improved.

Description

Technical Field

[0001] The invention belongs to the field of radar target detection, and in particular relates to a radar target detection method based on contrastive self-supervised learning. Background Art

[0002] Radar, as an electronic device that uses electromagnetic waves to detect targets, has been widely used in many fields such as national defense, transportation, and security, and has high military and civilian value. Automatic target detection is a basic function of modern radars and the basis for subsequent target tracking and identification. Therefore, its detection performance directly affects the radar detection capability. The essence of radar target detection is to determine whether the target actually exists from the received radar echo data. In addition to the target signal, the echo data received by the radar also contains other signals such as clutter, noise, and even interference. Therefore, there are still great challenges in current radar target detection, and the detection performance still needs to be further improved.

[0003] Radar target detection methods can generally be divided into statistical-based methods and feature-based methods. Statistical-based methods are traditional radar detection methods, mainly including likelihood ratio test (LRT), track before detection (TBD) and constant false alarm rate (CFAR). However, with the increasing complexity of radar echo signals, the difficulty of statistical modeling of traditional methods has also increased dramatically, and the detection performance has also decreased, and the application has been limited. Feature-based radar detection methods extract useful features from echo data from the perspective of pattern classification, and use the differences between target and other signal features to achieve target detection. At present, feature detection methods represented by deep learning have been widely used, and common algorithms include R-CNN, Fast R-CNN, Faster R-CNN, and YOLO (You Only Look Once) series. Relevant literature and practice show that supervised deep learning methods have higher detection performance than traditional methods, but they rely on high-quality labeled data, which is difficult to obtain and costly.

[0004] In response to the problem of labeled data dependence in supervised learning, this technical invention proposes a radar target detection method based on contrastive self-supervised learning. Based on the pre-training-fine-tuning paradigm, contrastive self-supervised (Contrastive Self-Supervised) pre-training is first performed on unlabeled radar echo data to obtain a pre-trained basic model, and then supervised fine-tuning (Supervised Fine-Tuning, SFT) is performed based on the pre-trained basic model using a small amount of labeled data to obtain a radar target detection model. Self-supervised learning is an unsupervised learning method that automatically learns data features that are beneficial to target detection from unlabeled data through auxiliary tasks (for example, contrastive learning). Literature research shows that this technical invention is the first to use contrastive self-supervised learning methods for radar target detection tasks.

[0005] Different radars have different signal processing methods and imaging principles. After receiving the echo signal, they can obtain different forms of radar echo data, such as one-dimensional high-resolution range profiles (HRRP) and two-dimensional range-Doppler spectra (RD), micro-Doppler spectra (MD) and SAR images.

[0006] Target detection is an important and basic task in radar systems. The detection performance will directly affect the subsequent target tracking and target recognition tasks. As the radar working environment becomes more and more complex and the target types emerge in an endless stream, traditional target detection methods based on statistics and artificial features are difficult to effectively solve the problems faced by radar target detection. Deep learning is a machine learning method that does not rely on artificial features and can automatically learn feature representations to distinguish targets, clutter and interference from radar data. Compared with traditional detection methods, deep learning has better detection performance and generalization, and has been widely used in radar target detection. However, the current deep learning method is mainly supervised learning, which relies on a large amount of high-quality labeled training data. However, in real application environments, labeled target data is usually difficult to obtain, and professionals need to spend a lot of time and effort to label it, which is inefficient and costly, thus significantly affecting the actual application effect of deep learning methods. Summary of the invention

[0007] In view of the shortcomings of the above-mentioned radar target detection methods, the purpose of the present invention is to propose a radar target detection method based on contrastive self-supervised learning. The present invention is based on the pre-training-fine-tuning paradigm. It first performs unsupervised learning on unlabeled radar echo data, innovatively designs a contrastive self-supervised pre-training method to obtain a pre-trained basic model, and then uses a small amount of labeled data based on the pre-trained basic model to perform supervised fine-tuning to obtain a radar target detection model. The method proposed in the present invention belongs to the category of deep learning. It does not rely on a large amount of high-quality labeled data, and can make full use of a large amount of unlabeled radar echo data to learn multi-dimensional target representation features, which can ensure further improvement of radar target detection performance, robustness and generalization.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is: a radar target detection method based on contrastive self-supervised learning, comprising the following steps:

[0009] Step S1, data preparation: preprocessing the echo signal collected by the radar receiver to obtain the RD graph corresponding to the signal, and performing data augmentation operation to obtain the augmented RD graph;

[0010] Step S2, self-supervised pre-training model: based on a large number of unlabeled RD graphs and augmented RD graphs, a self-supervised contrastive learning pre-training network is constructed to train a basic model;

[0011] Step S3, supervised fine-tuning model: a small amount of target detection annotated data is input into the pre-trained basic model, and supervised fine-tuning training is performed to obtain a radar target detection model.

[0012] Furthermore, step S1 performs preprocessing and accumulation of unlabeled data for unsupervised training, specifically including the following steps:

[0013] S1-1, read radar echo data and perform routine preprocessing;

[0014] S1-2, for the pre-processed radar echo data, pulse compression and Doppler processing are performed in the time dimension to obtain an RD graph;

[0015] S1-3. For the RD graph generated by the radar echo data, the corresponding comparative RD graph is obtained by using the data augmentation method.

[0016] Furthermore, step S2 constructs a self-supervised contrastive learning pre-training network based on the accumulated large number of unlabeled RD graphs and the comparative RD graphs obtained through data augmentation to perform basic model pre-training. The specific training process is as follows:

[0017] S2-1. Build the encoder , respectively encode the input RD graph and the data augmented RD graph into feature vectors and , is the training parameter;

[0018] S2-2. Constructing learnable nonlinear transformations , the feature vector and Mapped to the same contrastive learning space, we get the contrastive learning vector and , is the training parameter;

[0019] S2-3. Calculate contrast loss: Assume sample , is an unlabeled dataset. After steps S2-1 and S2-2, we get the vector = ( ); the corresponding quantity for each time is For the small batch training data samples, the number of samples obtained in step S1 is 2 training samples, where each sample And its augmented sample are positive samples to each other, and with the other 2( -1) samples are negative samples of each other, so the learning optimization goal is to maximize and The similarity between them is minimized, and the similarity between them and other negative samples is minimized, that is:

[0020]

[0021] The contrastive learning loss function is calculated as follows:

[0022]

[0023] Among them, the similarity function Representation vector Between Regularized similarity, is the temperature parameter, and the cost function of each small batch training parameter update is:

[0024] .

[0025] Furthermore, step S2-1 constructs an encoder based on the Transformer network to perform feature encoding on each RD graph. The encoder consists of two parts: an input embedding layer and a Transformer Block layer. The specific implementation of each layer is as follows:

[0026] The first part is the input embedding layer, which realizes low-level and local feature extraction through convolutional neural network and generates the feature vector of the input RD graph. The structure of this layer is as follows:

[0027]

[0028]

[0029] Where x is the input RD graph, is the output vector sequence, H and W are the height and width of the input RD image, S is the convolution step, and C is the number of channels after the convolution operation; BN(·) is the batch normalization operation, and the Patch(·) operation cuts the convolution output into several image blocks of size P and converts them into an output vector sequence. The sequence length is , d is the dimension of the feature vector;

[0030] The second part is the Transformer Block layer, whose input is the output of the first part, that is, the feature vector sequence of length N. This layer is optimized based on the standard Transformer encoder, replacing the linear projection before the multi-head attention in the original Encoder with a convolutional projection; for each input, three convolution transformations are performed to generate the corresponding Query, Key, and Value values, that is , the convolutional projection structure is described as follows:

[0031]

[0032] Where s is the step length of the convolution kernel. After a multi-head attention MHA(·) operation and a fully connected feedforward neural network MLP(·) operation, the final RD graph encoding representation vector is obtained. , The calculation process is as follows:

[0033]

[0034]

[0035] Among them, LN(·) is the layer normalization operation, Drop(·) is the Dropout operation, ,

[0036] Specifically, step S2-2 projects the original RD graph and its augmented RD graph representation into the same contrastive learning space. The projection operation consists of two fully connected layers, and its structure is as follows:

[0037]

[0038] in is the ReLU activation function, and Both are learnable weight parameter matrices. The feature representation vectors of the input unlabeled RD graph and its augmented RD graph are obtained through nonlinear transformation. and = .

[0039] Furthermore, step S2-3 continues to perform small batch training and synchronously updates the training network and , until the pre-training on the entire unlabeled dataset is completed and the pre-trained basic model is obtained .

[0040] Furthermore, step S3 is based on the pre-trained basic model , for radar target detection tasks, a small number of labeled RD graphs are used to perform supervised fine-tuning of the model. The specific fine-tuning process is as follows:

[0041] S3-1. Input the labeled radar RD image annotation dataset into the encoder In the above example, the RD graph is represented and learned, and the encoder outputs the corresponding representation vector ;

[0042] S3-2. Construct a query representation learner, whose input is , first through a projection layer Converted into a feature vector corresponding to the number of targets to be matched K (greater than the actual number of targets) , and then query the content separately and query location Representation learning, that is: , ;

[0043] S3-3. Build a Transformer-based decoder. , , At the same time, the decoder is input. After layer-by-layer iterative learning, the number of targets in the sample and their corresponding position boxes are obtained. The output results are matched with the true value labels, the cost function is calculated, and the cost function is continuously optimized until the training is completed.

[0044] Furthermore, the supervised fine-tuning model adopted in step S3 is an "encoder-decoder" structure model as a whole. A query representation learner is added after the encoder output and before the decoder input to form an "encoder-learner-decoder" model structure.

[0045] Furthermore, the fine-tuning process adopts a small batch training method. Step S3-1 selects The model is fine-tuned by inputting samples until all the data in the labeled sample set are trained. The encoder used is consistent with the encoder in self-supervised contrastive learning. Through multi-level and multi-scale convolution operations and Transformer fusion representation learning, rich features related to small radar targets in the RD image are extracted.

[0046] Furthermore, step S3-2 performs representation learning on the query input to the decoder based on the encoder output result. For the radar target detection task, the decoder needs to output the number of targets and the corresponding target box positions in the RD map. , is the coordinate of the center point of the target frame, h and w are the height and width of the target frame respectively, so the input Query includes the query content and query location Two parts, including By targeting the encoder output Perform linear transformation to obtain; for For example, the position offset is predicted by the feed-forward neural network FFN. ,in represents the i-th possible target in a RD graph, is the i-th target box position, then the input query position corresponding to the target is:

[0047] .

[0048] Further, step S3-3 receives , , Input decoder network, where As the decoder network learns layer by layer, it is updated and recorded as , the output vector of the current Transformer module is used to generate the input query content and query position of the next Transformer module, and the position offset of the candidate target is recorded as , so that the input query position of the next Transformer module is recorded as ,Right now:

[0049]

[0050] And so on, each Transformer module is iteratively learned and updated until the final prediction result is output. .

[0051] Furthermore, each RD graph has K prediction results. is the actual target label in the figure, is the model prediction result, where , for K prediction results, the model learning goal is to minimize the matching cost function, that is:

[0052]

[0053] in is the true value label and predicted labels The pairwise matching cost function between each label , Indicates whether it is a target. is the target box position vector, since , when performing matching calculations, fill in the labels in the true value label set y , for the predicted label In terms of The probability of , To predict the location of the target box, the Hungarian algorithm is used to calculate the two-way matching loss between the true value label and the predicted label. The definitions are as follows:

[0054]

[0055] in GIoU is used for calculation.

[0056]

[0057] is the L1 regularization form, To generalize the IoU cost function, it is calculated as follows:

[0058]

[0059] Represents the target box area, Indicates that it also contains The maximum target box of the entire fine-tuning model cost function is defined as follows:

[0060]

[0061]

[0062] in , , are all hyperparameters.

[0063] Compared with the prior art, the present invention proposes a radar target detection method based on contrastive self-supervised learning, which adopts a two-stage training paradigm of "pre-training-fine-tuning", makes full use of a large amount of unlabeled data for pre-training based on self-supervised learning, and performs supervised fine-tuning on a small amount of labeled samples, effectively alleviating the dependence on a large amount of high-quality labeled data, and further improves the radar target detection performance in complex environments, and has the following beneficial effects:

[0064] 1. The present invention makes full use of a large amount of unlabeled RD graph data for self-supervised learning. Different from the traditional self-supervised learning method, it does not need to set pseudo labels and directly performs comparative learning from the RD graph feature level.

[0065] 2. The present invention does not need to construct negative samples when performing self-supervised learning. In each batch of training sample sets, the similarity between each sample and its comparison sample is maximized, and the similarity with other training samples is minimized, effectively reducing the introduction of noise data.

[0066] 3. When the present invention performs supervised fine-tuning, it only needs to be based on a small amount of labeled data sets, without the need for a large amount of sample labeling work, which significantly improves labeling efficiency and reduces labeling costs.

[0067] 4. The present invention constructs an encoder and a decoder based on the Transformer network, and does not need to use other target detection networks (for example, Yolo) for target detection.

[0068] 5. The present invention fully combines the local feature extraction of convolutional networks and the long-range dependency modeling capabilities of Transformer networks, which is more beneficial to radar small target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 1 is a flow chart of a radar target detection method based on contrastive self-supervised learning in this embodiment.

[0070] Figure 2 Schematic diagram of the pre-training network for comparative self-supervised learning in this embodiment.

[0071] Figure 3 Schematic diagram of the Transformer encoder structure in this embodiment.

[0072] Figure 4 Schematic diagram of the supervised fine-tuning training network structure in this embodiment.

[0073] Figure 5 Schematic diagram of the decoder network structure used in this embodiment. DETAILED DESCRIPTION

[0074] The technical solution of the present invention is further explained below in conjunction with the accompanying drawings and specific embodiments.

[0075] The present invention proposes a radar target detection method based on contrastive self-supervised learning, and its implementation steps are as follows: Figure 1 As shown, the details are as follows:

[0076] Step S1, data preparation: pre-process the echo signal collected by the radar receiver to obtain the corresponding range-Doppler (RD) map of the signal, and perform data augmentation operation to obtain the augmented RD map.

[0077] Step S2: self-supervised pre-trained model: Based on a large number of unlabeled RD graphs and augmented RD graphs, a self-supervised contrastive learning pre-trained network is constructed to obtain a basic model through training.

[0078] Step S3: supervised fine-tuning model: A small amount of target detection RD map annotation data is input into the pre-trained basic model, and supervised fine-tuning training is performed to obtain a radar target detection model.

[0079] The step S1 performs preprocessing and accumulation of unlabeled data for unsupervised training, specifically comprising the following steps:

[0080] S1-1. Read radar echo data for preprocessing, which mainly includes deleting abnormal data, completing missing information, normalizing data and other routine processing contents.

[0081] S1-2. For the pre-processed radar echo data, pulse compression and Doppler processing are performed in the time dimension to obtain an RD graph.

[0082] S1-3. For the RD map generated by radar echo data, the corresponding comparative RD map is obtained through data augmentation method. RD map is different from optical image, and conventional data augmentation methods such as rotation, inversion and cropping cannot be used directly. It needs to be augmented in combination with data features.

[0083] Preferably, a generative artificial intelligence method is adopted, for example, a denoising autoencoder DAE (DenoisingAutoEncoder) is used to generate a positive RD graph that is close to the original RD graph.

[0084] Preferably, data augmentation is performed on the original radar echo data, a time warping method is used to generate radar I / Q echo data, and the newly generated radar echo data is subjected to pulse compression and Doppler processing to obtain an augmented comparative RD graph.

[0085] The step S2 constructs a self-supervised contrastive learning pre-training network based on the accumulated large number of unlabeled RD graphs and the comparative RD graphs obtained through data augmentation, performs basic model pre-training, and the overall architecture of the training network is as follows: Figure 2 The specific training steps are as follows:

[0086] S2-1. Build the encoder , respectively, the input RD graph (set as x) and the data augmented RD graph (set as ) is encoded as a feature vector and , is the training parameter.

[0087] S2-2. Constructing a learnable nonlinear transformation projector , the feature vector and Mapped to the same contrastive learning space, we get the contrastive learning vector and , is the training parameter.

[0088] S2-3, calculate contrast loss. Assume sample ( is an unlabeled dataset), after steps S2-1 and S2-2, the vector = ( ). The corresponding quantity for each time is =64 small batch training data samples, the number of which is 2 ( ) training samples, each of which And its augmented sample are positive samples to each other, and with the other 2( -1) samples are negative samples of each other, so the learning optimization goal is to maximize and The similarity between them is minimized, and the similarity between them and other negative samples is minimized, that is:

[0089]

[0090] The contrastive learning loss function is calculated as follows:

[0091]

[0092] Among them, the similarity function Representation vector Between Regularized similarity, is the temperature parameter. The cost function for each small batch training parameter update is:

[0093]

[0094] Specifically, step S2-1 constructs an encoder based on the Transformer network to perform feature encoding on each RD graph. Figure 3 As shown in the figure, the encoder consists of two parts: the input embedding layer and the Transformer Block layer. The specific implementation of each layer is as follows:

[0095] The first part is the input embedding layer, which uses a convolutional neural network to extract low-level and local features and generate the feature vector of the input RD graph. The structure of this layer is as follows:

[0096]

[0097]

[0098] Where x is the input RD graph, is the output vector sequence, H and W are the height and width of the input RD image, S is the convolution step, and C is the number of channels after the convolution operation. BN(·) is the batch normalization operation. The Patch(·) operation cuts the convolution output into several image blocks of size P and converts them into an output vector sequence. The sequence length is , d is the dimension of the feature vector. In this embodiment, H=W=480, C=1024, S=1, P=24, d=256.

[0099] The second part is the Transformer Block layer. This embodiment uses 3 layers for stacking, and its input is the output of the first part, that is, the feature vector sequence of length N. This layer is optimized based on the standard Transformer encoder, replacing the linear projection before the multi-head attention (MHA) in the original encoder with a convolutional projection; for each input, the corresponding Query, Key and Value values ​​are generated through 3 convolution transformations, that is, The convolutional projection structure is described as follows:

[0100]

[0101] Where s=1 is the step length of the convolution kernel. Figure 3 As shown in Figure 2, after a multi-head attention MHA(·) operation and a fully connected feedforward neural network MLP(·) operation, the final RD graph encoding representation vector is obtained. . The calculation process is as follows:

[0102]

[0103]

[0104] Among them, LN(·) is the layer normalization operation, Drop(·) is the Dropout operation, .

[0105] Specifically, step S2-2 projects the original RD graph and its augmented RD graph representation into the same contrastive learning space. The projection operation consists of two fully connected layers, whose structure is as follows:

[0106]

[0107] in is the ReLU activation function, and All are learnable weight parameter matrices. The feature representation vector of the input unlabeled RD graph and its augmented RD graph is obtained through nonlinear transformation. and = .

[0108] Specifically, step S2-3 continues to perform small batch training and synchronously updates the training network and , until the pre-training on the entire unlabeled dataset is completed and the pre-trained basic model is obtained .

[0109] The step S3 is based on the pre-trained basic model , for radar target detection tasks, a small number of labeled RD graphs are used to perform supervised fine-tuning of the model, such as Figure 4 The specific fine-tuning process is as follows:

[0110] S3-1. Input the labeled radar RD image annotation dataset into the encoder In the above example, the RD graph is represented and learned, and the encoder outputs the corresponding representation vector .

[0111] S3-2. Construct a query representation learner, whose input is First, we pass a projection layer Converted into a feature vector corresponding to the number of targets to be matched K (greater than the actual number of targets) , and then query the content separately and query location Representation learning, that is: , In this embodiment, K=50.

[0112] S3-3. Construct a Transformer-based decoder. This embodiment uses three layers of blocks for stacking. , , At the same time, the decoder is input, and after layer-by-layer iterative learning, the number of targets in the sample and their corresponding position boxes are obtained. The output result is matched with the true value label, the cost function is calculated, and the cost function is continuously optimized until the training is completed.

[0113] Specifically, step S3 is aimed at the target detection task of the radar RD map, and the target detection model is obtained through supervised fine-tuning training based on the pre-trained basic model generated in steps S1 and S2. The supervised fine-tuning model used in step S3 is an "encoder-decoder" structure model as a whole. The present invention adds a query representation learner after the encoder output and before the decoder input to form an "encoder-learner-decoder" model structure, which can further improve the radar small target detection performance in the RD map.

[0114] The fine-tuning process adopts a small batch training method. In this embodiment, the value In step S3-1, each time a labeled RD graph data set is selected The model is fine-tuned by inputting samples until all the data in the labeled sample set are trained. The encoder used is the same as the encoder in self-supervised contrastive learning. Through multi-level and multi-scale convolution operations and Transformer fusion representation learning, the rich features related to small radar targets in the RD image can be better extracted, which helps to improve the performance of radar small target detection.

[0115] Specifically, step S3-2 performs representation learning on the Query input to the decoder based on the encoder output result. For radar target detection tasks, the decoder needs to output the number of targets and the corresponding target box positions in the RD map. , is the coordinate of the center point of the target frame, h and w are the height and width of the target frame respectively, so the input Query includes the query content and query location Two parts, including By targeting the encoder output Perform a linear transformation to get. For example, the position offset is predicted by the feed-forward neural network FFN. ,in represents the i-th possible target in a RD graph, is the i-th target box position, then the input query position corresponding to the target is:

[0116]

[0117] Specifically, the step S3-3 receives , , Input decoder network, where As the decoder network learns layer by layer, it is updated and recorded as In this embodiment, the internal structure of the decoder is as follows: Figure 5 The output vector of the current Transformer module is used to generate the input query content and query position of the next Transformer module, and the position offset of the candidate target is recorded as , so that the input query position of the next Transformer module is recorded as ,Right now:

[0118]

[0119] And so on, each Transformer module is iteratively learned and updated until the final prediction result is output. .

[0120] Each RD graph has K prediction results (K=50 in this embodiment). is the actual target label in the figure, is the model prediction result, where For K prediction results, the model learning goal is to minimize the matching cost function, that is:

[0121]

[0122] in is the true value label and predicted labels The pairwise matching cost function between each label , Indicates whether it is a target. is the target box position vector. , when performing matching calculations, fill in the labels in the true value label set y For the predicted label In the present invention, it is defined as being predicted The probability of , To predict the target box position. The present invention uses the Hungarian algorithm to calculate the two-way matching loss between the true value label and the predicted label. The definitions are as follows:

[0123]

[0124] in The calculation is done using the GIoU (Generalized Intersection over Union, generalized IoU) method.

[0125]

[0126] is the L1 regularization form, To generalize the IoU cost function, it is calculated as follows:

[0127]

[0128] Represents the target box area, Indicates that it also contains The cost function of the entire fine-tuning model is defined as follows:

[0129]

[0130]

[0131] in , , are all hyperparameters.

[0132] Although the invention has been disclosed as above in terms of preferred embodiments, the embodiments and drawings are not intended to limit the invention. Anyone familiar with the art can make various changes or modifications without departing from the spirit and scope of the invention, but they are also within the scope of protection of the invention. Therefore, the scope of protection of the invention shall be based on the scope of protection of the claims of this application.

Claims

1. A radar target detection method based on contrastive self-supervised learning, characterized in that: The steps include: Step S1, data preparation: preprocessing the echo signal collected by the radar receiver to obtain the RD graph corresponding to the signal, and performing data augmentation operation to obtain the augmented RD graph; Step S2, self-supervised pre-training model: based on a large number of unlabeled RD graphs and augmented RD graphs, a self-supervised contrastive learning pre-training network is constructed to train a basic model; Step S3, supervised fine-tuning model: a small amount of target detection annotated data is input into the pre-trained basic model, and supervised fine-tuning training is performed to obtain a radar target detection model.

2. A radar target detection method based on contrastive self-supervised learning according to claim 1, characterized in that: Step S1 performs preprocessing and accumulation of unlabeled data for unsupervised training, specifically including the following steps: S1-1, read radar echo data and perform routine preprocessing; S1-2, for the pre-processed radar echo data, pulse compression and Doppler processing are performed in the time dimension to obtain an RD graph; S1-3. For the RD graph generated by the radar echo data, the corresponding comparative RD graph is obtained by using the data augmentation method.

3. A radar target detection method based on contrastive self-supervised learning according to claim 1, characterized in that: Step S2 builds a self-supervised contrastive learning pre-training network based on the accumulated large number of unlabeled RD graphs and the comparative RD graphs obtained through data augmentation, and performs basic model pre-training. The specific training process is as follows: S2-1. Build the encoder , respectively encode the input RD graph and the data augmented RD graph into feature vectors and , is the training parameter; S2-2. Constructing learnable nonlinear transformations , the feature vector and Mapped to the same contrastive learning space, we get the contrastive learning vector and , is the training parameter; S2-3. Calculate contrast loss: Assume sample , is an unlabeled dataset. After steps S2-1 and S2-2, we get the vector = ( ); the corresponding quantity for each time is For the small batch training data samples, the number of samples obtained in step S1 is 2 training samples, where each sample And its augmented sample are positive samples to each other, and with the other 2( -1) samples are negative samples of each other, so the learning optimization goal is to maximize and The similarity between them is minimized, and the similarity between them and other negative samples is minimized, that is: The contrastive learning loss function is calculated as follows: Among them, the similarity function Representation vector Between Regularized similarity, is the temperature parameter, and the cost function of each small batch training parameter update is: 。 4. The radar target detection method based on contrastive self-supervised learning according to claim 3, characterized in that: Step S2-1 constructs an encoder based on the Transformer network to encode the features of each RD graph. The encoder consists of two parts: one is the input embedding layer, and the other is the Transformer Block layer. The specific implementation of each layer is as follows: The first part is the input embedding layer, which realizes low-level and local feature extraction through convolutional neural network and generates the feature vector of the input RD graph. The structure of this layer is as follows: Where x is the input RD graph, is the output vector sequence, H and W are the height and width of the input RD image, S is the convolution step, and C is the number of channels after the convolution operation; BN(·) is the batch normalization operation, Conv(·) represents the convolution operation, and the Patch(·) operation cuts the convolution output into several image blocks of size P and converts them into an output vector sequence. The sequence length is , d is the dimension of the feature vector; The second part is the Transformer Block layer, whose input is the output of the first part, that is, the feature vector sequence of length N. This layer is optimized based on the standard Transformer encoder, replacing the linear projection before the multi-head attention in the original Encoder with a convolutional projection; for each input, three convolution transformations are performed to generate the corresponding Query, Key, and Value values, that is , the convolutional projection structure is described as follows: Where s is the step length of the convolution kernel, the Reshape2D(·) operation reshapes the input data into a two-dimensional shape, Conv2D(·) represents a two-dimensional convolution operation, and the Flatten(·) operation converts multi-dimensional data into one dimension. After a multi-head attention MHA(·) operation and a fully connected feedforward neural network MLP(·) operation, the final RD graph encoding representation vector is obtained. , The calculation process is as follows: Among them, LN(·) is the layer normalization operation, Drop(·) is the Dropout operation, , Specifically, step S2-2 projects the original RD graph and its augmented RD graph representation into the same contrastive learning space. The projection operation consists of two fully connected layers, and its structure is as follows: in is the ReLU activation function, and Both are learnable weight parameter matrices. The feature representation vectors of the input unlabeled RD graph and its augmented RD graph are obtained through nonlinear transformation. and = .

5. The radar target detection method based on contrastive self-supervised learning according to claim 3, characterized in that: Step S2-3: Continue small batch training and update the training network synchronously and , until the pre-training on the entire unlabeled dataset is completed and the pre-trained basic model is obtained .

6. The radar target detection method based on contrastive self-supervised learning according to claim 5, characterized in that: Step S3 is based on the pre-trained basic model , for radar target detection tasks, a small number of labeled RD graphs are used to perform supervised fine-tuning of the model. The specific fine-tuning process is as follows: S3-1. Input the labeled radar RD image annotation dataset into the encoder In the above example, the RD graph is represented and learned, and the encoder outputs the corresponding representation vector ; S3-2. Construct a query representation learner, whose input is , first through a projection layer Converted into a feature vector corresponding to the number of targets to be matched K (greater than the actual number of targets) , and then query the content separately and query location Representation learning, that is: , , where Linear(·) represents linear transformation and FFN(·) represents feedforward network; S3-3. Build a Transformer-based decoder. , , At the same time, the decoder is input. After layer-by-layer iterative learning, the number of targets in the sample and their corresponding position boxes are obtained. The output results are matched with the true value labels, the cost function is calculated, and the cost function is continuously optimized until the training is completed.

7. The radar target detection method based on contrastive self-supervised learning according to claim 1, characterized in that: The supervised fine-tuning model used in step S3 is an "encoder-decoder" structure model as a whole. A query representation learner is added after the encoder output and before the decoder input to form an "encoder-learner-decoder" model structure.

8. The radar target detection method based on contrastive self-supervised learning according to claim 6, characterized in that: The fine-tuning process adopts a small batch training method. Step S3-1 selects The model is fine-tuned by inputting samples until all the data in the labeled sample set are trained. The encoder used is consistent with the encoder in self-supervised contrastive learning. Through multi-level and multi-scale convolution operations and Transformer fusion representation learning, rich features related to small radar targets in the RD image are extracted.

9. The radar target detection method based on contrastive self-supervised learning according to claim 6, characterized in that: Step S3-2 performs representation learning on the Query input to the decoder based on the encoder output result. For radar target detection tasks, the decoder needs to output the number of targets and the corresponding target box positions in the RD map. , is the coordinate of the center point of the target frame, h and w are the height and width of the target frame respectively, so the input Query includes the query content and query location Two parts, including By targeting the encoder output Perform linear transformation to obtain; for For example, the position offset is predicted by the feed-forward neural network FFN. ,in represents the i-th possible target in a RD graph, is the i-th target box position, then the input query position corresponding to the target is: 。 10. The radar target detection method based on contrastive self-supervised learning according to claim 6, characterized in that: Step S3-3: Receiving , , Input decoder network, where As the decoder network learns layer by layer, it is updated and recorded as , the output vector of the current Transformer module is used to generate the input query content and query position of the next Transformer module, and the position offset of the candidate target is recorded as , so that the input query position of the next Transformer module is recorded as ,Right now: And so on, each Transformer module is iteratively learned and updated until the final prediction result is output. .

11. A radar target detection method based on contrastive self-supervised learning according to claim 10, characterized in that: Each RD graph has K prediction results, assuming is the actual target label in the figure, is the model prediction result, where , for K prediction results, the model learning goal is to minimize the matching cost function, that is: in is the true value label and predicted labels The pairwise matching cost function between each label , Indicates whether it is a target. is the target box position vector, since , when performing matching calculations, fill in the labels in the true value label set y , for the predicted label In terms of The probability of , To predict the location of the target box, the Hungarian algorithm is used to calculate the two-way matching loss between the true value label and the predicted label. The definitions are as follows: in GIoU is used for calculation. is the L1 regularization form, To generalize the IoU cost function, it is calculated as follows: Represents the target box area, Indicates that it also contains The maximum target box of the entire fine-tuning model cost function is defined as follows: in , , are all hyperparameters.

Images