Radar time-frequency domain data feature prototype vector intelligent extraction method

Abstract

The invention provides a radar time-frequency domain data feature prototype vector intelligent extraction method, and relates to the field of radar intelligent detection. According to the method, through constructing radar time-frequency domain training data, a training sample is provided for an intelligent extraction method; dividing radar time-frequency domain training data into a support set and a query set; analyzing and processing the support set sample labels to generate segmentation masks of the support set samples; building a convolutional feature extraction network model, and initializing network model parameters; performing mask average pooling processing on the radar time-frequency domain data feature map extracted by the convolutional feature extraction network; designing a convolutional feature extraction network optimization function; and calculating a loss function by using the query set data label, carrying out iterative updating on the parameters of the convolutional feature extraction network model until the training result convergence meets the requirement, and outputting a feature prototype vector of the target signal. The problems that features extracted by an existing feature extraction method are insufficient in saliency, and classification and recognition are not facilitated are solved.

Description

Technical Field

[0001] This document relates to the field of radar intelligent detection, and in particular to a method for intelligent extraction of radar time-frequency domain data feature prototype vectors. Background Art

[0002] The signal processing and detection and recognition process of radar intelligent detection can be roughly divided into preprocessing, feature extraction and classification judgment. The purpose of preprocessing is that the original radar echo signal not only contains target information but is often accompanied by external clutter and internal noise. These unfavorable factors for target recognition will affect the final determination of the target category or model. Therefore, in order to highlight the target characteristics and suppress external clutter and internal noise, it is necessary to perform appropriate preprocessing according to the characteristics of the original radar echo signal and the recognition requirements to make it suitable for the subsequent feature extraction process. Feature extraction is to use the target characteristics to extract the features in the radar signal that help determine the category and model of the test set target through effective methods. The feature extraction methods are different for different target ranges and different forms of signals. Classification judgment is the step of classifying unknown targets based on the similarity between the features of the training data and the features of the test data. Accurate and efficient feature extraction methods can effectively support the subsequent classification judgment process and play a decisive role in the realization of radar high-precision and high-accuracy detection tasks.

[0003] Traditional feature extraction methods mainly extract key features of targets through manual analysis and design of feature rules. With the development of intelligent technology, various deep learning-based feature extraction methods have been proposed, and convolutional neural networks (CNN) have become the main feature extraction tool. Chen et al. realized feature extraction of targets on remote sensing images through improved SegNet and completed pixel-level semantic segmentation tasks. Zuo et al. proposed a hierarchical fusion full convolutional network for aerial images, which enhanced the feature extraction capability of convolutional neural networks for targets. The application of intelligent technology can improve the feature extraction capability of radar signals and help improve the judgment capability of classifiers. However, in the field of radar detection, there is still a problem that the extracted features are not significant enough, which is not conducive to classification and recognition.

[0004] Therefore, there is an urgent need for a feature extraction method to solve the problem that the features extracted by the existing feature extraction methods are not significant enough and are not conducive to classification and recognition. Summary of the invention

[0005] The present specification provides a method for intelligently extracting prototype vectors of radar time-frequency domain data features, which is used to solve the problem that the features extracted by existing feature extraction methods are not significant enough and are not conducive to classification and recognition.

[0006] In a first aspect, this specification provides a method for intelligently extracting a prototype vector of radar time-frequency domain data features, the method comprising:

[0007] Acquire radar time-frequency domain data and extract data feature vectors for detection and recognition tasks;

[0008] According to the radar time-frequency domain data, a radar time-frequency domain training data set is constructed, and the training data set data is divided into support set data and query set data;

[0009] Generate a binary segmentation mask for isolating a target signal from background noise and interference in the time-frequency domain according to the label information of the support set data;

[0010] According to the requirements of data feature vector extraction of the detection and recognition task, a convolutional feature extraction network model is built; and according to the requirements of radar time-frequency domain data and data feature vector extraction, an optimization function is designed for the convolutional feature extraction network;

[0011] Inputting the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature graph;

[0012] Performing mask average pooling processing according to the radar time-frequency domain data feature map and the support set data segmentation mask;

[0013] According to the masked average pooling processing result and the query set data, the optimization function is used to calculate the loss function value, and the convolutional feature extraction network model parameters are iteratively updated until the training results converge to meet the requirements, thereby obtaining the feature prototype vector results of the radar time-frequency domain data.

[0014] In the second aspect, the present specification provides a radar time-frequency domain data feature prototype vector intelligent extraction device, including: a data acquisition and demand extraction module, a data set construction module, a segmentation mask generation module, a network model construction and optimization module, a time-frequency domain data feature extraction module, a mask pooling processing module, a model training and application module; wherein:

[0015] The data acquisition and demand extraction module is used to acquire radar time-frequency domain data and extract requirements for data feature vectors of detection and recognition tasks;

[0016] The data set construction module is used to construct a radar time-frequency domain training data set based on the radar time-frequency domain data, and divide the training data set data into support set data and query set data;

[0017] The segmentation mask generating module is used to generate a segmentation mask of the support set data according to the label information of the support set data;

[0018] The network model building and optimization module is used to build a convolutional feature extraction network model according to the detection and recognition task data feature vector extraction requirements; and to design an optimization function for the convolutional feature extraction network according to the radar time-frequency domain data and data feature vector extraction requirements;

[0019] The time-frequency domain data feature extraction module is used to input the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature map;

[0020] The mask pooling processing module is used to perform mask average pooling processing according to the radar time-frequency domain data feature map and the support set data segmentation mask;

[0021] The model training and application module is used to calculate the loss function value using the optimization function according to the mask average pooling processing result and the query set data, and iteratively update the convolutional feature extraction network model parameters until the training results converge to meet the requirements, thereby obtaining the feature prototype vector results of the radar time-frequency domain data.

[0022] The beneficial effects of the present invention are as follows:

[0023] This specification provides a method for intelligent extraction of radar time-frequency domain data feature prototype vectors. The method provides training samples for the intelligent extraction method by constructing radar time-frequency domain training data; dividing the radar time-frequency domain training data into a support set and a query set; analyzing and processing the support set sample labels to generate a segmentation mask for the support set samples; building a convolutional feature extraction network model and initializing the network model parameters; performing masked average pooling processing on the radar time-frequency domain data feature map extracted by the convolutional feature extraction network; designing a convolutional feature extraction network optimization function; using the query set data labels to calculate the loss function, and iteratively updating the convolutional feature extraction network model parameters until the training results converge to meet the requirements and output the feature prototype vector of the target signal. This method uses the support set data labels to supervise the intelligent network, and combines the masked average pooling strategy to shield the clutter and interference in the background, which can effectively improve the extraction quality of the target echo signal feature prototype vector by the intelligent network, and can support the realization of intelligent target detection and recognition tasks with high performance requirements in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings described herein are used to provide a further understanding of this specification and constitute a part of this specification. The illustrative embodiments and descriptions of this specification are used to explain this specification and do not constitute an improper limitation on this specification. In the drawings:

[0025] Figure 1 It is a schematic diagram of an intelligent extraction method of radar time-frequency domain data feature prototype vector provided in an embodiment of this specification;

[0026] Figure 2 It is a schematic diagram of a method flow for intelligent extraction of radar time-frequency domain data feature prototype vectors provided in an embodiment of this specification;

[0027] Figure 3 It is a schematic diagram of an intelligent extraction device for radar time-frequency domain data feature prototype vector provided in an embodiment of this specification. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and their corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this document.

[0029] The technical solutions provided by the embodiments of this specification are described in detail below in conjunction with the accompanying drawings. Specific embodiment one:

[0031] This embodiment provides a method for intelligently extracting characteristic prototype vectors of radar time-frequency domain data. Figure 1 , the method comprising:

[0032] Step 102: Acquire radar time-frequency domain data and extract data feature vectors for detection and recognition tasks;

[0033] It should be noted that the radar data in complex electromagnetic scenes, as well as the corresponding target relative distance, speed and category information, are obtained through actual measurement or simulation, and the data is subjected to time-frequency analysis and processing; the data feature vector extraction requirements for detection and identification tasks are clarified, including the types of complex electromagnetic scenes and parameter ranges applicable to detection and identification tasks, detection and identification category settings, input data format requirements, output result format requirements, output result accuracy requirements, computational efficiency requirements, and other special requirements.

[0034] Step 104: construct a radar time-frequency domain training data set based on the radar time-frequency domain data, and divide the training data set data into support set data and query set data;

[0035] Specifically, a specific implementation of step 104 may be:

[0036] S41. Calibrate the radar time-frequency domain data according to the relative distance, speed and category information of the known target; then divide the calibrated radar time-frequency domain data into a training set and a test set according to a set ratio, which together constitute a radar time-frequency domain training data set for the training and performance verification of the intelligent algorithm.

[0037] S42. Randomly extract data in proportion from the training set and test set of the radar time-frequency domain training data set as support set data and query set data, which are divided into training set support data, training set query data, test set support data and test set query data. Among them, the training set support data is used to extract feature prototype vectors during the model training process; the training set query data is used to verify the validity of the feature prototype vectors extracted during the model training process, and assist the model training process; the test set support data is used to extract feature prototype vectors during the model performance test verification process; the test set query data is used to verify the validity of the feature prototype vectors extracted during the model performance test verification process, and assist the model performance test verification process.

[0038] Step 106: Generate a binary segmentation mask for isolating the target signal from background noise and interference in the time-frequency domain according to the label information of the support set data;

[0039] Specifically, a specific implementation of step 106 may be:

[0040] For the data labels under the detection and recognition task, firstly, a full-zero segmentation mask template Y0 of size W×H is established, where W and H represent the width and height of the input data respectively;

[0041] Generate a w×h all-1 matrix mask label according to the label box width w and height h;

[0042] According to the morphological characteristics of the target signal in the time-frequency domain, the all-1 matrix mask labels are processed using mathematical transformation formulas to trim the target mask coverage so that it is exactly the same as the morphology of the target signal in the time-frequency domain;

[0043] Taking the coordinates (x0, y0) of the center point of the target bounding box as the center, on the template segmentation mask Y0, the logical operation rules are used to process the template segmentation mask Y0 and the processed target signal morphology mask label to obtain the segmentation mask Y that can completely segment the target signal and the background interference signal area in the time-frequency domain image.

[0044] Based on this, a segmentation mask that completely corresponds to the morphological characteristics of the target signal in the time-frequency domain can be obtained, thereby realizing the segmentation of the target signal and other environmental clutter interference in the time-frequency domain.

[0045] Step 108: building a convolutional feature extraction network model according to the detection and recognition task data feature vector extraction requirements; and designing an optimization function for the convolutional feature extraction network according to the radar time-frequency domain data and data feature vector extraction requirements;

[0046] Specifically, a specific implementation of step 108 may be:

[0047] S81, according to the detection and recognition task data feature vector extraction requirements, analyze the characteristics of the data to be extracted, and build a convolutional feature extraction network model;

[0048] The constructed convolutional feature extraction network model includes the convolutional network model structure, convolutional network random initialization parameters, and training hyperparameters.

[0049] Wherein, the convolutional feature extraction network optimization function includes: loss function, iterative optimization method, loss function hyperparameters;

[0050] First, define the loss function calculation formula. In order to capture more accurate target information, the loss function is defined using the Focal Loss optimization formula that can meet the fine-grained pixel level. c , L c The specific formula is as follows:

[0051]

[0052] Where N is the number of positive samples; p x ' y The confidence that the convolutional feature extraction network model predicts the data at the position (x, y) of the time-frequency domain image as the target signal; p xy is the true label at the position (x, y) of the time-frequency domain image and satisfies p xy ∈{0,1}; μ is used to solve the problem of balancing positive and negative samples; γ is used to balance the problem of balancing difficult and easy samples.

[0053] The iterative optimization method uses the gradient descent method of adaptive moment estimation to iteratively update the network model parameters. The update rules of the adaptive moment estimation algorithm are as follows:

[0054]

[0055] Among them, θ t represents the parameters to be optimized at time t, They represent the first-order moment and the second-order moment respectively, η represents the learning rate, and ε is a constant used to avoid the situation where the denominator is zero.

[0056] Set hyperparameters based on manual experience or analysis of experimental results.

[0057] Step 110: input the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature graph;

[0058] Specifically, a specific implementation of step 110 may be:

[0059] S101, inputting the support set data into the convolutional feature extraction network to obtain two-dimensional feature map prediction information;

[0060] S102. The two-dimensional feature map of radar time-frequency domain data retains the original spatial position attribute of the target signal in the time-frequency domain.

[0061] Step 112: performing mask average pooling processing according to the radar time-frequency domain data feature map and the support set data segmentation mask;

[0062] Specifically, a specific implementation of step 112 may be:

[0063] S121, based on the support set data, according to the generated support set data segmentation mask, eliminating the influence of background noise, extracting features related to the target object, and using the average pooling strategy to output a feature prototype representation vector related only to the target signal in the time-frequency domain;

[0064] The mask average pooling strategy transforms the feature map F'∈R with a channel number of c through bilinear interpolation. c×w'×h' The width and height are resized to the segmentation mask Y∈{0,1} W×H The width and height are the same; the resized feature map is represented by F∈R c ×W×H , by averaging the pixels of the target area on each channel i of the feature map, the i-th element v of the prototype representation vector v is calculated i ;

[0065] Among them, the calculation formula of the prototype representation vector is as follows:

[0066]

[0067] Among them, v i The prototype represents the i-th element of the vector v, F i,x,y is the element value at the i-th channel feature data (x, y) of the network output feature map F, Y x,y is the segmentation mask element value of the input image I at (x, y), W and H represent the width and height of the image.

[0068] Based on this, we can obtain the time-frequency domain feature prototype vector representation of the target signal that completely shields background clutter and interference, effectively improving the feature extraction capability of intelligent technology for radar time-frequency domain data, improving the recognition accuracy of subsequent intelligent classification tasks, and enhancing the interpretability of the classification network.

[0069] Step 114: According to the masked average pooling processing result and the query set data, the optimization function is used to calculate the loss function value, and the convolutional feature extraction network model parameters are iteratively updated until the training result converges to meet the requirements, thereby obtaining the feature prototype vector result of the radar time-frequency domain data.

[0070] Specifically, a specific implementation of step 114 may be:

[0071] S141, based on the query set data and the target prototype representation vector obtained by mask average pooling, the loss function is calculated using the query set data label;

[0072] S142, using a gradient descent-based method to reversely update the network model parameters, and using a quantized evaluation method to evaluate the training state until the results converge, to obtain a characteristic prototype vector result of the radar time-frequency domain data that shields background interference;

[0073] S143. Save the network model parameters with converged results as a reloadable file.

[0074] In summary, this embodiment uses the support set data label supervision intelligent network, combined with the mask average pooling strategy to shield the clutter and interference in the background, which can effectively improve the extraction quality of the target echo signal feature prototype vector by the intelligent network, and can support the realization of intelligent target detection and recognition tasks with high performance requirements in complex scenarios. Specific embodiment 2:

[0076] This embodiment provides a device for intelligently extracting prototype vectors of radar time-frequency domain data features. Figure 2 , including: data acquisition and demand extraction module 201, data set construction module 202, segmentation mask generation module 203, network model construction and optimization module 204, time-frequency domain data feature extraction module 205, mask pooling processing module 206, model training and application module 207; wherein:

[0077] The data acquisition and demand extraction module 201 is used to acquire radar time-frequency domain data and extract requirements for data feature vectors of detection and recognition tasks;

[0078] The data set construction module 202 is used to construct a radar time-frequency domain training data set based on the radar time-frequency domain data, and divide the training data set data into support set data and query set data;

[0079] The segmentation mask generating module 203 is used to generate a segmentation mask of the support set data according to the label information of the support set data;

[0080] The network model building and optimization module 204 is used to build a convolutional feature extraction network model according to the detection and recognition task data feature vector extraction requirements; and design an optimization function for the convolutional feature extraction network according to the radar time-frequency domain data and data feature vector extraction requirements;

[0081] The time-frequency domain data feature extraction module 205 is used to input the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature graph;

[0082] The mask pooling processing module 206 is used to perform mask average pooling processing according to the radar time-frequency domain data feature map and the support set data segmentation mask;

[0083] The model training and application module 207 is used to calculate the loss function value using the optimization function according to the mask average pooling processing result and the query set data, and iteratively update the convolutional feature extraction network model parameters until the training results converge to meet the requirements, thereby obtaining the feature prototype vector results of the radar time-frequency domain data.

[0084] Optionally, the data acquisition and demand extraction module 201 is specifically used for:

[0085] The types of complex electromagnetic scenes and parameter ranges applicable to detection and identification tasks, detection and identification category settings, input data format requirements, output result format requirements, output result accuracy requirements, and computational efficiency requirements.

[0086] Optionally, the support set data is used to extract feature prototype vectors during model training;

[0087] The query set data is used to verify the validity of the feature prototype vector extracted during the model training process and to assist the model training process.

[0088] Optionally, the segmentation mask generating module is specifically used to:

[0089] Create an all-0 segmentation mask template Y0 of size W×H, where W and H represent the width and height of the input data respectively;

[0090] Generate a w×h all-1 matrix mask label according to the label box width w and height h;

[0091] According to the morphological characteristics of the target signal in the time-frequency domain, the all-1 matrix mask labels are processed using mathematical transformation formulas to trim the target mask coverage so that it is exactly the same as the morphology of the target signal in the time-frequency domain;

[0092] Taking the coordinates (x0, y0) of the center point of the target bounding box as the center, on the template segmentation mask Y0, the logical operation rules are used to process the template segmentation mask Y0 and the processed target signal morphology mask label to obtain the segmentation mask Y that can completely segment the target signal and the background interference signal area in the time-frequency domain image.

[0093] Optionally, the network model building and optimization module 204 is specifically used to:

[0094] According to the requirements for extracting feature vectors of the detection and recognition task data, the characteristics of the data to be extracted are analyzed, and a convolutional feature extraction network model is built;

[0095] The constructed convolutional feature extraction network model includes the convolutional network model structure, convolutional network random initialization parameters, and training hyperparameters.

[0096] Optionally, the convolutional feature extraction network optimization function includes: a loss function, an iterative optimization method, and a loss function hyperparameter;

[0097] Among them, the loss function L c It is expressed as:

[0098]

[0099] Where N is the number of positive samples; p x ' y The confidence that the convolutional feature extraction network model predicts the data at the position (x, y) of the time-frequency domain image as the target signal; p xy is the true label at the position (x, y) of the time-frequency domain image and satisfies p xy ∈{0,1}; μ is used to solve the problem of balancing positive and negative samples; γ is used to balance the problem of balancing difficult and easy samples.

[0100] Optionally, the time-frequency domain data feature extraction module 205 is specifically used for:

[0101] Inputting the support set data into the convolutional feature extraction network to obtain two-dimensional feature map prediction information;

[0102] The two-dimensional feature map of radar time-frequency domain data retains the original spatial position attributes of the target signal in the time-frequency domain.

[0103] Optionally, the mask pooling processing module 206 is specifically used to:

[0104] Based on the support set data, the generated support set data segmentation mask is used to eliminate the influence of background noise, extract the features related to the target object, and use the average pooling strategy to output the feature prototype representation vector related only to the target signal in the time-frequency domain;

[0105] The mask average pooling strategy transforms the feature map F'∈R with a channel number of c through bilinear interpolation. c×w'×h' The width and height are resized to the segmentation mask Y∈{0,1} W×H The width and height are the same; the resized feature map is represented by F∈R c ×W×H , by averaging the pixels of the target area on each channel i of the feature map, the i-th element v of the prototype representation vector v is calculated i ;

[0106] Among them, the calculation formula of the prototype representation vector is as follows:

[0107]

[0108] Among them, v i The prototype represents the i-th element of the vector v, F i,x,y is the element value at the i-th channel feature data (x, y) of the network output feature map F, Y x,y is the segmentation mask element value of the input image I at (x, y), W and H represent the width and height of the image.

[0109] Optionally, the model training and application module 207 is specifically used for:

[0110] Based on the query set data and the target prototype representation vector obtained by mask average pooling, the loss function is calculated using the query set data label;

[0111] The network model parameters are updated inversely using a gradient descent method, and the training status is evaluated using a metric evaluation method until the results converge, obtaining the characteristic prototype vector results of the radar time-frequency domain data that shields background interference;

[0112] Save the converged network model parameters as a reloadable file.

[0113] In summary, this embodiment uses the support set data label supervision intelligent network, combined with the mask average pooling strategy to shield the clutter and interference in the background, which can effectively improve the extraction quality of the target echo signal feature prototype vector by the intelligent network, and can support the realization of intelligent target detection and recognition tasks with high performance requirements in complex scenarios.

[0114] The above description is only a preferred embodiment of this specification and is not intended to limit this specification. For those skilled in the art, this specification may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this specification shall be included in the protection scope of this specification.

Claims

1. A method for intelligently extracting characteristic prototype vectors of radar time-frequency domain data, characterized in that: The method comprises: Acquire radar time-frequency domain data and extract data feature vectors for detection and recognition tasks; According to the radar time-frequency domain data, a radar time-frequency domain training data set is constructed, and the training data set data is divided into support set data and query set data; Generate a binary segmentation mask for isolating a target signal from background noise and interference in the time-frequency domain according to the label information of the support set data; According to the requirements of data feature vector extraction of the detection and recognition task, a convolutional feature extraction network model is built; and according to the requirements of radar time-frequency domain data and data feature vector extraction, an optimization function is designed for the convolutional feature extraction network; Inputting the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature graph; Performing mask average pooling processing according to the radar time-frequency domain data feature map and the support set data segmentation mask; According to the masked average pooling processing result and the query set data, the optimization function is used to calculate the loss function value, and the convolutional feature extraction network model parameters are iteratively updated until the training results converge to meet the requirements, thereby obtaining the feature prototype vector results of the radar time-frequency domain data.

2. The method according to claim 1, characterized in that The data feature vector extraction requirements for detection and recognition tasks include: The types of complex electromagnetic scenes and parameter ranges applicable to detection and identification tasks, detection and identification category settings, input data format requirements, output result format requirements, output result accuracy requirements, and computational efficiency requirements.

3. The method according to claim 2, characterized in that The support set data is used to extract feature prototype vectors during model training; The query set data is used to verify the validity of the feature prototype vector extracted during the model training process and to assist the model training process.

4. The method according to claim 3, characterized in that: The step of generating a binary segmentation mask for isolating a target signal from background noise and interference in a time-frequency domain according to the label information of the support set data comprises: Create an all-0 segmentation mask template Y0 of size W×H, where W and H represent the width and height of the input data respectively; Generate a w×h all-1 matrix mask label according to the label box width w and height h; According to the morphological characteristics of the target signal in the time-frequency domain, the all-1 matrix mask labels are processed using mathematical transformation formulas to trim the target mask coverage so that it is exactly the same as the morphology of the target signal in the time-frequency domain; Taking the coordinates (x0, y0) of the center point of the target bounding box as the center, on the template segmentation mask Y0, the logical operation rules are used to process the template segmentation mask Y0 and the processed target signal morphology mask label to obtain the segmentation mask Y that can completely segment the target signal and the background interference signal area in the time-frequency domain image.

5. The method according to claim 4, characterized in that According to the detection and recognition task data feature vector extraction requirements, building a convolutional feature extraction network model includes: According to the requirements for extracting feature vectors of the detection and recognition task data, the characteristics of the data to be extracted are analyzed, and a convolutional feature extraction network model is built; The constructed convolutional feature extraction network model includes the convolutional network model structure, convolutional network random initialization parameters, and training hyperparameters.

6. The method according to claim 5, characterized in that The convolutional feature extraction network optimization function includes: Loss functions, iterative optimization methods, loss function hyperparameters; Among them, the loss function L c It is expressed as: Where N is the number of positive samples; p x ' y The confidence that the convolutional feature extraction network model predicts the data at the position (x, y) of the time-frequency domain image as the target signal; p xy is the true label at the position (x, y) of the time-frequency domain image and satisfies p xy ∈{0,1}; μ is used to solve the problem of balancing positive and negative samples; γ is used to balance the problem of balancing difficult and easy samples.

7. The method according to claim 6, characterized in that The step of inputting the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature graph comprises: Inputting the support set data into the convolutional feature extraction network to obtain two-dimensional feature map prediction information; The two-dimensional feature map of radar time-frequency domain data retains the original spatial position attributes of the target signal in the time-frequency domain.

8. The method according to claim 7, characterized in that The mask average pooling process is performed according to the radar time-frequency domain data feature map and the support set data segmentation mask, including: Based on the support set data, the generated support set data segmentation mask is used to eliminate the influence of background noise, extract the features related to the target object, and use the average pooling strategy to output the feature prototype representation vector related only to the target signal in the time-frequency domain; The mask average pooling strategy transforms the feature map F'∈R with a channel number of c through bilinear interpolation. c×w'×h' The width and height are resized to the segmentation mask Y∈{0,1} W×H The width and height are the same; the resized feature map is represented by F∈R c×W×H , by averaging the pixels of the target area on each channel i of the feature map, the i-th element v of the prototype representation vector v is calculated i ; Among them, the calculation formula of the prototype representation vector is as follows: Among them, v i The prototype represents the i-th element of the vector v, F i,x,y is the element value at the i-th channel feature data (x, y) of the network output feature map F, Y x,y is the segmentation mask element value of the input image I at (x, y), W and H represent the width and height of the image.

9. The method according to claim 8, characterized in that The loss function value is calculated using the optimization function according to the mask average pooling processing result and the query set data, and the convolution feature extraction network model parameters are iteratively updated until the training result converges to meet the requirements, and the feature prototype vector result of the radar time-frequency domain data is obtained, including: Based on the query set data and the target prototype representation vector obtained by mask average pooling, the loss function is calculated using the query set data label; The network model parameters are updated inversely using a gradient descent method, and the training status is evaluated using a metric evaluation method until the results converge, obtaining the characteristic prototype vector results of the radar time-frequency domain data that shields background interference; Save the converged network model parameters as a reloadable file.

10. An intelligent device for extracting characteristic prototype vectors of radar time-frequency domain data, applied to the method according to any one of claims 1 to 9, characterized in that: include: Data acquisition and demand extraction module, data set construction module, segmentation mask generation module, network model construction and optimization module, time-frequency domain data feature extraction module, mask pooling processing module, model training and application module; Among them: The data acquisition and demand extraction module is used to acquire radar time-frequency domain data and extract requirements for data feature vectors of detection and recognition tasks; The data set construction module is used to construct a radar time-frequency domain training data set based on the radar time-frequency domain data, and divide the training data set data into support set data and query set data; The segmentation mask generating module is used to generate a segmentation mask of the support set data according to the label information of the support set data; The network model building and optimization module is used to build a convolutional feature extraction network model according to the detection and recognition task data feature vector extraction requirements; and to design an optimization function for the convolutional feature extraction network according to the radar time-frequency domain data and data feature vector extraction requirements; The time-frequency domain data feature extraction module is used to input the support set data into the convolutional feature extraction network to obtain a radar time-frequency domain data feature map; The mask pooling processing module is used to perform mask average pooling processing according to the radar time-frequency domain data feature map and the support set data segmentation mask; The model training and application module is used to calculate the loss function value using the optimization function according to the mask average pooling processing result and the query set data, and iteratively update the convolutional feature extraction network model parameters until the training results converge to meet the requirements, thereby obtaining the feature prototype vector results of the radar time-frequency domain data.

Images