Security radar target data processing method

Abstract

The invention discloses a security radar target data processing method, which comprises the following steps of: S1, preprocessing a radar echo signal, and converting the signal into frequency domain data through denoising, amplitude normalization and spectral analysis; s2, extracting salient region features by adopting a sparse attention mechanism, ignoring background data, and generating sparse feature representation; s3, feature representation is optimized through a self-supervised learning task, and learning of deep features of radar signals is achieved; s4, performing dynamic weight fusion on the sparse feature flow and the global feature flow by using a double-flow architecture to generate joint feature representation; s5, iteratively strengthening the salient region features through a recursive optimization method, and generating optimized target features; and S6, performing target classification and detection based on the optimized target features, and outputting the category and position information of the target. The method improves the radar target data processing efficiency and accuracy, and is suitable for the technical field of security monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method for processing security radar target data. Background Art

[0002] With the continuous development of security technology, radar is increasingly used in the field of security monitoring. Radar can monitor the movement trajectory, speed and spatial position of target objects in real time, providing important technical support for security systems. However, in the existing technology, security radar systems still face many technical challenges in target data processing, especially in terms of processing efficiency and detection accuracy.

[0003] In the existing technology, traditional radar target data processing methods mainly rely on rule-based signal analysis and fixed feature extraction mechanisms. These methods are incapable of coping with complex security scenarios. Specifically, the existing technology has obvious deficiencies in the following aspects: 1. Low efficiency of feature extraction: Traditional methods usually perform global processing on radar echo signals, which makes it difficult to effectively extract significant features of the target area. This not only increases the computational complexity, but also leads to the omission of key feature information.

[0004] 2. Insufficient ability to utilize unlabeled data: Existing methods usually rely on labeled data for training. However, in security scenarios, the cost of obtaining labeled data is high and limited, which limits the ability to represent deep features of radar signals.

[0005] 3. Weak information fusion capability: Traditional methods have difficulty in achieving efficient fusion of global and local information in target detection. This single feature stream processing mode often leads to insufficient accuracy in target classification and detection.

[0006] 4. Limited feature enhancement capabilities: Existing technologies lack a recursive optimization mechanism, making it difficult to focus on key features through multiple iterations, resulting in weak expression capabilities of salient area features, which ultimately affects the accuracy of target detection and classification.

[0007] 5. Insufficient real-time performance: In embedded security radar equipment, the high computational complexity of existing algorithms leads to insufficient real-time processing capabilities, which makes it difficult to meet the requirements of actual application scenarios.

[0008] Therefore, how to provide a security radar target data processing method is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a security radar target data processing method. The present invention adopts sparse attention mechanism, self-supervised learning, dual-stream architecture and recursive feature enhancement technologies, and describes in detail a processing method for extracting significant regional features from radar signals, dynamically fusing global and local information, and recursively optimizing target features. The method has the advantages of high computational efficiency, strong target detection accuracy and efficient use of unlabeled data.

[0010] A security radar target data processing method according to an embodiment of the present invention comprises the following steps: S1. Acquire a radar echo signal to be processed, and perform data preprocessing operations on the radar echo signal, including denoising, amplitude normalization, and spectrum analysis, to obtain preprocessed radar frequency domain data; S2, using the sparse attention mechanism to extract significant features from the preprocessed radar frequency domain data, and generating sparse feature representation by calculating the attention weights; S3, construct a self-supervised learning task to generate optimized deep feature representations by predicting predefined signal perturbations; S4. Use a dual-stream architecture to perform dynamic feature fusion on the optimized deep feature representation, where the sparse feature stream extracts salient regional features and the global feature stream extracts overall context information. The sparse feature stream and the global feature stream are fused through a dynamic weight allocation strategy to generate a joint feature representation. S5, recursively enhancing the joint feature representation, optimizing the expression ability of the salient region features through multiple iterations, so that the feature representation of the target region is gradually enhanced, and further optimized target features are generated; S6. Based on the further optimized target features, target classification and detection processing are performed, the target category is determined through the support vector machine classification algorithm, the spatial position information and category information of the target are output, and the target detection and recognition task is completed.

[0011] Optionally, the S2 specifically includes: S21, based on the preprocessed radar frequency domain data, dividing the radar frequency domain data into a plurality of feature partitions, and extracting the frequency domain features of each partition as input features , construct the attention weight matrix , calculate the characteristic response value : ; in, Indicates the radar frequency domain data The characteristic response value of each partition, is the first The input features of the partitions, is the attention weight matrix, which is used to weight the input features. is the bias term, is a nonlinear activation function; S22, based on the calculated characteristic response value , through the sparse function For the attention weight matrix Perform sparse processing and only retain feature response values ​​greater than or equal to the preset sparse threshold The partition characteristics of S23. The sparse attention weight matrix It is called the salient region feature, and the salient region features are weighted and summed to generate a sparse feature representation : ; in, For the The input features of the partitions, Indicates the total number of significant partitions retained, It is a sparse feature representation used to characterize the significant target information in radar frequency domain data. is the sparse function; S24, representing the sparse features Perform standardized operations.

[0012] Optionally, the S22 specifically includes: S221, attention weight matrix The characteristic response values ​​of Perform statistical analysis and calculate the mean and standard deviation : ; ; in, Represents the total number of partitions of radar frequency domain data; S222: Dynamically determine the sparseness threshold based on the statistical analysis results : ; in, is the adjustment factor used to control the sensitivity of the threshold; S223, attention weight matrix The characteristic response value of Apply the sparsification function , calculate the sparse attention weight matrix : ; in, Represents the sparse attention weight matrix.

[0013] Optionally, the S3 specifically includes: S31, based on sparse feature representation , construct self-supervised learning tasks, design signal disturbance prediction tasks, and represent the input sparse features Apply predefined perturbations to generate perturbation signatures : ; in, is a random disturbance vector used to simulate the random noise in the radar signal; S32. Based on sparse feature representation and disturbance characteristics , construct a loss function based on weighted Euclidean distance , used to quantify the disturbance characteristics and sparse feature representation The difference between: ; in, represents the total dimension of the sparse feature representation, and Respectively represent The original and perturbed values ​​of the features, represents the Euclidean distance, is the feature weight; By calculating the characteristic response value of each feature With the mean The degree of deviation determines the feature weight : ; in, Indicates The characteristic response value of each feature, and are the mean and standard deviation of the characteristic response values, represents the absolute value operation, is an exponential function; S33. Loss function based on weighted Euclidean distance , by dynamically adjusting feature weights and optimize the parameters of the self-supervised learning network to minimize the perturbation characteristics and sparse feature representation The difference between the two generates an optimized deep feature representation .

[0014] Optionally, the S4 specifically includes: S41, based on optimized deep feature representation , build a two-stream architecture and extract sparse feature streams respectively and global feature flow , where the sparse feature flow From sparse feature representation Extracted from, the feature representation representing the local salient target, the global feature flow From the optimized deep feature representation Extract overall context information from S42, based on the characteristic response value calculation dynamic weight allocation strategy, for sparse feature flow and global feature flow Calculate dynamic weights separately and ; S43. Sparse feature flow and global feature flow Perform dynamic fusion processing to generate joint feature representation : ; in, Represents joint feature representation, which is used to simultaneously represent salient target features and global context information; S44. Representation of joint features Regularization processing is performed to adjust the feature distribution range.

[0015] Optionally, the S42 specifically includes: S421. Calculate sparse feature representation The characteristic response value of and global feature flow The characteristic response value of : ; ; in, Indicates the sparse feature representation Features, Indicates the first Features, and are the weights of the features, and are the number of features for sparse feature representation and global feature flow, respectively; S422, based on characteristic response value and Computing dynamic weights for sparse feature flows and the dynamic weights of the global feature flow : ; ; in, represents the exponential function; S423. Dynamic weights for sparse feature flows and the dynamic weights of the global feature flow Normalize it so that .

[0016] Optionally, the S5 specifically includes: S51. Joint feature representation , initialize the recursive feature representation , set the recursive iteration number threshold and recursively optimize the weight matrix and the bias vector ; S52, in each iteration, using the recursive optimization function to represent the recursive feature Perform significant area optimization to generate optimized recursive feature representation ; S53. After each iteration, the optimized recursive feature is expressed as Apply non-linear activation function , generating the activated recursive feature representation : ; in, is a predefined activation function, Indicates the number of iterations; S54, when the number of iterations Reached the recursive iteration threshold When , the final recursively optimized feature representation is output , as the further optimized target feature, used for target classification and detection tasks.

[0017] Optionally, the S52 specifically includes: S521. Use recursive optimization weight matrix and the bias vector Initialize recursive optimization function parameters; S522, in In the iterations, the recursive feature representation Apply the recursive optimization function to generate an optimized recursive feature representation : ; in, Indicates the result of the previous round of recursive optimization; S523, calculating the optimized recursive feature representation by feature weight allocation The significant regional characteristic response value : ; in, Indicates The weight of the feature, Represents the total dimension of the recursive feature; S524, the significant area feature response value Return to the recursive optimization function and continue recursive iteration until the preset number of iterations is reached .

[0018] The beneficial effects of the present invention are: (1) The present invention combines sparse attention mechanism, self-supervised learning, dual-stream feature fusion architecture and recursive feature enhancement technology to achieve deep understanding and dynamic optimization of radar target features, enabling the system to efficiently extract significant regional features, dynamically combine global and local information, and gradually enhance the expression ability of target area features through recursive optimization, thereby effectively coping with complex security monitoring scenarios and significantly improving the accuracy of target detection and classification.

[0019] (2) The present invention significantly reduces the computational complexity by utilizing the sparse attention mechanism, focusing only on the features of the key target area and avoiding redundant processing of global calculations. At the same time, the self-supervised learning technology is used to efficiently utilize unlabeled data, thereby solving the problem of insufficient labeled data and improving the system's generalization ability and robustness for target features. The dual-stream feature fusion architecture further combines the information of the sparse feature stream and the global feature stream, enabling the system to effectively understand the relationship between targets and backgrounds in complex scenes.

[0020] (3) The present invention realizes multiple iterative optimization of significant regional features through recursive feature enhancement processing, so that the expression ability of target features is gradually enhanced, and the accuracy of target classification and detection is further improved. At the same time, based on the design of low computational complexity, this method can meet the strict requirements of embedded security radar equipment on real-time performance and computing resources, thereby improving the efficiency, accuracy and adaptability of radar target detection systems, and is widely applicable to real-time target recognition tasks in the field of security monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1The present invention provides a flow chart of a method for processing security radar target data. DETAILED DESCRIPTION

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0023] refer to Figure 1 , a security radar target data processing method, comprising the following steps: S1. Acquire a radar echo signal to be processed, and perform data preprocessing operations on the radar echo signal, including denoising, amplitude normalization, and spectrum analysis, to obtain preprocessed radar frequency domain data; S2, using the sparse attention mechanism to extract significant features from the preprocessed radar frequency domain data, and generating sparse feature representation by calculating the attention weights; Optionally, the S2 specifically includes: S21, based on the preprocessed radar frequency domain data, dividing the radar frequency domain data into a plurality of feature partitions, and extracting the frequency domain features of each partition as input features , construct the attention weight matrix , calculate the characteristic response value : ; in, Indicates the radar frequency domain data The characteristic response value of each partition, is the first The input features of the partitions, is the attention weight matrix, which is used to weight the input features. is the bias term, is a nonlinear activation function; S22, based on the calculated characteristic response value , through the sparse function For the attention weight matrix Perform sparse processing and only retain feature response values ​​greater than or equal to the preset sparse threshold The partition characteristics of S23. The sparse attention weight matrix It is called the salient region feature, and the salient region features are weighted and summed to generate a sparse feature representation : ; in, For the The input features of the partitions, Indicates the total number of significant partitions retained, It is a sparse feature representation used to characterize the significant target information in radar frequency domain data. is the sparse function; S24, representing the sparse features Perform standardized operations.

[0024] This implementation method performs sparse processing on the attention weight matrix and combines it with a dynamic threshold adjustment method to retain only the target features in the salient area, significantly reducing the computational complexity while improving the ability to capture key target areas. It provides a more efficient and accurate feature expression for subsequent processing and optimizes the performance of security radar target detection.

[0025] Optionally, the S22 specifically includes: S221, attention weight matrix The characteristic response values ​​of Perform statistical analysis and calculate the mean and standard deviation : ; ; in, Represents the total number of partitions of radar frequency domain data; S222: Dynamically determine the sparseness threshold based on the statistical analysis results : ; in, is the adjustment factor used to control the sensitivity of the threshold; S223, attention weight matrix The characteristic response value of Apply the sparsification function , calculate the sparse attention weight matrix : ; in, Represents the sparse attention weight matrix.

[0026] This implementation method dynamically optimizes the attention weight matrix through sparse processing, and utilizes the adaptive calculation of feature response values ​​and dynamic thresholds to effectively extract the features of significant target areas, thereby reducing the impact of non-critical features, significantly improving the accuracy and computational efficiency of target data processing, and providing a more efficient solution for security radar target detection.

[0027] S3, construct a self-supervised learning task to generate optimized deep feature representations by predicting predefined signal perturbations; Optionally, the S3 specifically includes: S31, based on sparse feature representation , construct self-supervised learning tasks, design signal disturbance prediction tasks, and represent the input sparse features Apply predefined perturbations to generate perturbation signatures : ; in, is a random disturbance vector used to simulate the random noise in the radar signal; S32. Based on sparse feature representation and disturbance characteristics , construct a loss function based on weighted Euclidean distance , used to quantify the disturbance characteristics and sparse feature representation The difference between: ; in, represents the total dimension of the sparse feature representation, and Respectively represent The original and perturbed values ​​of the features, represents the Euclidean distance, is the feature weight; By calculating the characteristic response value of each feature With the mean The degree of deviation determines the feature weight : ; in, Indicates The characteristic response value of each feature, and are the mean and standard deviation of the characteristic response values, represents the absolute value operation, is an exponential function; S33. Loss function based on weighted Euclidean distance , by dynamically adjusting feature weights and optimize the parameters of the self-supervised learning network to minimize the perturbation characteristics and sparse feature representation The difference between the two generates an optimized deep feature representation .

[0028] This implementation method constructs self-supervised learning tasks, uses sparse feature representation for deep feature optimization, and combines weighted Euclidean distance and adaptive weight allocation to effectively capture the deep-level significant features of radar signals, improve the robustness and generalization ability of target detection, and provide more accurate and efficient technical support for security radar target data processing.

[0029] S4. Use a dual-stream architecture to perform dynamic feature fusion on the optimized deep feature representation, where the sparse feature stream extracts salient regional features and the global feature stream extracts overall context information. The sparse feature stream and the global feature stream are fused through a dynamic weight allocation strategy to generate a joint feature representation. Optionally, the S4 specifically includes: S41, based on optimized deep feature representation , build a two-stream architecture and extract sparse feature streams respectively and global feature flow , where the sparse feature flow From sparse feature representation Extracted from, the feature representation representing the local salient target, the global feature flow From the optimized deep feature representation Extract overall context information from S42, based on the characteristic response value calculation dynamic weight allocation strategy, for sparse feature flow and global feature flow Calculate dynamic weights separately and ; S43. Sparse feature flow and global feature flow Perform dynamic fusion processing to generate joint feature representation : ; in, Represents joint feature representation, which is used to simultaneously represent salient target features and global context information; S44. Representation of joint features Regularization processing is performed to adjust the feature distribution range.

[0030] This implementation method dynamically fuses the sparse feature stream and the global feature stream and generates a joint feature representation using a dynamic weight allocation strategy. It not only retains the local features of the salient area targets, but also integrates the global context information, achieves accurate expression and efficient coordination of features, improves the accuracy and robustness of target detection, and provides high-quality input support for subsequent classification.

[0031] Optionally, the S42 specifically includes: S421. Calculate sparse feature representation The characteristic response value of and global feature flow The characteristic response value of : ; ; in, Indicates the sparse feature representation Features, Indicates the first Features, and are the weights of the features, and are the number of features for sparse feature representation and global feature flow, respectively; S422, based on characteristic response value and Computing dynamic weights for sparse feature flows and the dynamic weights of the global feature flow : ; ; in, represents the exponential function; S423. Dynamic weights for sparse feature flows and the dynamic weights of the global feature flow Normalize it so that .

[0032] This implementation achieves the coordinated optimization of salient regions and global context information by calculating the dynamic weight distribution of sparse features and global features, which not only improves the accuracy of feature fusion, but also ensures the balance and stability of feature expression, providing a better feature basis for subsequent target detection tasks.

[0033] S5, recursively enhancing the joint feature representation, optimizing the expression ability of the salient region features through multiple iterations, so that the feature representation of the target region is gradually enhanced, and further optimized target features are generated; Optionally, the S5 specifically includes: S51. Joint feature representation , initialize the recursive feature representation , set the recursive iteration number threshold and recursively optimize the weight matrix and the bias vector ; S52, in each iteration, using the recursive optimization function to represent the recursive feature Perform significant area optimization to generate optimized recursive feature representation ; S53. After each iteration, the optimized recursive feature is expressed as Apply non-linear activation function , generating the activated recursive feature representation : ; in, is a predefined activation function, Indicates the number of iterations; S54, when the number of iterations Reached the recursive iteration threshold When , the final recursively optimized feature representation is output , as the further optimized target feature, used for target classification and detection tasks.

[0034] This implementation method uses recursive feature enhancement processing, combined with salient area optimization and nonlinear activation, to gradually enhance the expressive power of target features, making the features of salient target areas clearer, supporting higher-precision target classification and detection tasks, and improving the detection efficiency and accuracy of the security radar system.

[0035] Optionally, the S52 specifically includes: S521. Use recursive optimization weight matrix and the bias vector Initialize recursive optimization function parameters; S522, in In the iterations, the recursive feature representation Apply the recursive optimization function to generate an optimized recursive feature representation : ; in, Indicates the result of the previous round of recursive optimization; S523, calculating the optimized recursive feature representation by feature weight allocation The significant regional characteristic response value : ; in, Indicates The weight of the feature, Represents the total dimension of the recursive feature; S524, the significant area feature response value Return to the recursive optimization function and continue recursive iteration until the preset number of iterations is reached .

[0036] This implementation method realizes the gradual optimization of recursive feature representation through recursive optimization of the weight matrix and dynamic adjustment of the bias vector, combined with the adaptive weight allocation of the feature response value of the significant area, which can effectively enhance the significant feature expression of the target area and improve the accuracy and robustness of target classification and detection in security radar target data processing.

[0037] S6. Based on the further optimized target features, target classification and detection processing are performed, the target category is determined through the support vector machine classification algorithm, the spatial position information and category information of the target are output, and the target detection and recognition task is completed.

[0038] Optionally, the S6 specifically includes: S61, based on the further optimized target feature, normalizing the target feature to adjust the value range of the target feature to within a predefined standard range; S62, inputting the normalized target features into a support vector machine classification algorithm, constructing a classification decision model, and determining the category label of the target by calculating the similarity between the support vector classification algorithm and the target features; S63, outputting the category label of the target according to the calculation result of the classification decision model, and completing the target classification task; S64, predicting the spatial position of the classified target in combination with the target detection method, including determining the center coordinates and bounding box range of the target; S65: Integrate the target category label and spatial location information into the detection result, and output the target classification and location prediction information.

[0039] Embodiment 1: In order to verify the feasibility and superiority of the present invention, we applied the security radar target data processing method of the present invention to the perimeter security system of a certain airport. Since the airport perimeter covers a wide area, the environment is complex, and the target types are diverse and frequently changing, the traditional radar monitoring system often cannot meet the requirements of real-time performance and accuracy in this scenario. Therefore, the airport management department decided to introduce the method of the present invention to improve the radar target detection and classification capabilities.

[0040] In this scenario, the radar system needs to monitor illegal intrusion targets (such as personnel and vehicles) around the runway and low-altitude flying targets (such as drones) in real time. Traditional methods are prone to missed reports and false alarms due to low data processing efficiency and weak feature extraction capabilities, especially in rainy and foggy weather or when there is severe environmental interference. In order to solve these problems, we deployed a radar target data processing system based on the present invention.

[0041] First, the system receives the radar echo signal, performs denoising, amplitude normalization, and spectrum analysis on it to generate radar frequency domain data. Subsequently, through the sparse attention mechanism, only the significant regional features in the signal are extracted to reduce the interference of background noise and generate sparse feature representation. Next, the system uses self-supervised learning methods to perform feature pre-training on unlabeled data, automatically learn the deep feature representation of radar signals, and improve the system's target recognition ability in complex scenes. The dual-stream feature fusion architecture dynamically combines the advantages of sparse feature streams and global feature streams, and generates a more accurate joint feature representation through a dynamic weight allocation strategy. Finally, the recursive feature enhancement module performs multiple iterative optimizations on the joint feature representation, so that the target feature expression capability is gradually enhanced, and finally the target classification and detection are completed.

[0042] During the system implementation, we collected actual data from different scenarios around the airport perimeter, including monitoring data on sunny days, rainy days, nights, and high-interference environments. The experimental data was processed in real time at the airport security control center, and the results are as follows: Table 1 Comparison of target detection performance of airport perimeter security systems

[0043] As can be seen from Table 1, after the deployment of the system of the present invention, the number of target detections has increased, especially the detection accuracy of illegal targets has been significantly improved, from 87.4% to 96.8%. The false alarm rate has been reduced from 5.7% to 1.3%, and the false alarm rate has been reduced from 4.5% to 1.1%. The system's ability to identify illegal targets has been significantly enhanced. At the same time, the detection response time has been greatly shortened, from 18 seconds in traditional systems to 6 seconds, effectively meeting the needs of real-time security at the airport perimeter. In addition, in rainy and foggy weather and under strong interference conditions, the system of the present invention can still maintain stable detection performance, while the traditional system shows a significant decline.

[0044] In a specific case, the system of the present invention detected an abnormal target at the edge of the runway at night. The target features were quickly extracted under the sparse attention mechanism, and the feature representation optimized by the recursive feature enhancement module was accurately classified as an illegal intrusion. Finally, the system completed the detection and alarm within 6 seconds, and the relevant security personnel arrived at the scene in time to avoid potential risks.

[0045] It can be seen from the above experimental data and actual application cases that the method of the present invention not only solves the shortcomings of traditional radar target detection methods in terms of efficiency, accuracy and environmental adaptability, but also significantly improves the overall performance of the security system, providing reliable technical support for high-demand scenarios such as airports.

[0046] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A security radar target data processing method, characterized in that: The steps include: S1. Acquire a radar echo signal to be processed, and perform data preprocessing operations on the radar echo signal, including denoising, amplitude normalization, and spectrum analysis, to obtain preprocessed radar frequency domain data; S2, using the sparse attention mechanism to extract significant features from the preprocessed radar frequency domain data, and generating sparse feature representation by calculating the attention weights; S3, construct a self-supervised learning task to generate optimized deep feature representations by predicting predefined signal perturbations; S4. Use a dual-stream architecture to perform dynamic feature fusion on the optimized deep feature representation, where the sparse feature stream extracts salient regional features and the global feature stream extracts overall context information. The sparse feature stream and the global feature stream are fused through a dynamic weight allocation strategy to generate a joint feature representation. S5, recursively enhancing the joint feature representation, optimizing the expression ability of the salient region features through multiple iterations, so that the feature representation of the target region is gradually enhanced, and further optimized target features are generated; S6. Based on the further optimized target features, target classification and detection processing are performed, the target category is determined through the support vector machine classification algorithm, the spatial position information and category information of the target are output, and the target detection and recognition task is completed.

2. A security radar target data processing method according to claim 1, characterized in that: The S2 specifically includes: S21, based on the preprocessed radar frequency domain data, dividing the radar frequency domain data into a plurality of feature partitions, and extracting the frequency domain features of each partition as input features , construct the attention weight matrix , calculate the characteristic response value : ； in, Indicates the radar frequency domain data The characteristic response value of each partition, is the first The input features of the partitions, is the attention weight matrix, which is used to weight the input features. is the bias term, is a nonlinear activation function; S22, according to the calculated characteristic response value , through the sparse function For the attention weight matrix Perform sparse processing and only retain feature response values ​​greater than or equal to the preset sparse threshold The partition characteristics of S23. The sparse attention weight matrix It is called the salient region feature, and the salient region features are weighted and summed to generate a sparse feature representation : ； in, For the The input features of the partitions, Indicates the total number of retained significant partitions, It is a sparse feature representation used to characterize the significant target information in radar frequency domain data. is the sparse function; S24, representing the sparse features Perform standardized operations.

3. A security radar target data processing method according to claim 2, characterized in that: The S22 specifically includes: S221, attention weight matrix The characteristic response values ​​of Perform statistical analysis and calculate the mean and standard deviation : ； ； in, Represents the total number of partitions of radar frequency domain data; S222. Dynamically determine the sparseness threshold based on statistical analysis results : ； in, is the adjustment factor used to control the sensitivity of the threshold; S223, attention weight matrix The characteristic response value of Apply the sparsification function , calculate the sparse attention weight matrix : ； in, Represents the sparse attention weight matrix.

4. A security radar target data processing method according to claim 3, characterized in that: The S3 specifically includes: S31, based on sparse feature representation , construct self-supervised learning tasks, design signal disturbance prediction tasks, and represent the sparse features of the input Apply predefined perturbations to generate perturbation signatures : ； in, is a random disturbance vector used to simulate the random noise in the radar signal; S32. Based on sparse feature representation and disturbance characteristics , construct a loss function based on weighted Euclidean distance , used to quantify the disturbance characteristics and sparse feature representation The difference between: ； in, represents the total dimension of the sparse feature representation, and Respectively represent The original and perturbed values ​​of the features, represents the Euclidean distance, is the feature weight; By calculating the characteristic response value of each feature With the mean The degree of deviation determines the feature weight : ； in, Indicates The characteristic response value of each feature, and are the mean and standard deviation of the characteristic response values, Represents absolute value operation, is an exponential function; S33. Loss function based on weighted Euclidean distance , by dynamically adjusting feature weights and optimize the parameters of the self-supervised learning network to minimize the perturbation characteristics and sparse feature representation The difference between the two generates an optimized deep feature representation .

5. A security radar target data processing method according to claim 4, characterized in that: The S4 specifically includes: S41, based on optimized deep feature representation , build a two-stream architecture and extract sparse feature streams respectively and global feature flow , where the sparse feature flow From sparse feature representation Extracted from, the feature representation representing the local salient target, the global feature flow From the optimized deep feature representation Extract overall context information from S42, based on the characteristic response value calculation dynamic weight allocation strategy, for sparse feature flow and global feature flow Calculate dynamic weights separately and ; S43. Sparse feature flow and global feature flow Perform dynamic fusion processing to generate joint feature representation : ； in, Represents joint feature representation, which is used to simultaneously represent salient target features and global context information; S44. Representation of joint features Regularization processing is performed to adjust the feature distribution range.

6. A security radar target data processing method according to claim 5, characterized in that: The S42 specifically includes: S421. Calculate sparse feature representation The characteristic response value of and global feature flow The characteristic response value of : ； ； in, Indicates the sparse feature representation Features, Indicates the first Features, and are the weights of the features, and are the number of features for sparse feature representation and global feature flow, respectively; S422, based on characteristic response value and Computing dynamic weights for sparse feature flows and the dynamic weights of the global feature flow : ； ； in, represents the exponential function; S423. Dynamic weights for sparse feature flows and the dynamic weights of the global feature flow Normalize it so that .

7. A security radar target data processing method according to claim 6, characterized in that: The S5 specifically includes: S51. Joint feature representation , initialize the recursive feature representation , set the recursive iteration number threshold and recursively optimize the weight matrix and the bias vector ; S52, in each iteration, using the recursive optimization function to represent the recursive feature Perform significant area optimization to generate optimized recursive feature representation ; S53. After each iteration, the optimized recursive feature is expressed as Apply non-linear activation function , generating the activated recursive feature representation : ； in, is a predefined activation function, Indicates the number of iterations; S54, when the number of iterations Reached the recursive iteration threshold When , the final recursive optimized feature representation is output , as the further optimized target feature, used for target classification and detection tasks.

8. The method for processing security radar target data according to claim 7, characterized in that: The S52 specifically includes: S521. Use recursive optimization weight matrix and the bias vector Initialize recursive optimization function parameters; S522, in In the iterations, the recursive feature representation Apply the recursive optimization function to generate an optimized recursive feature representation : ； in, Indicates the result of the previous round of recursive optimization; S523, calculating the optimized recursive feature representation by feature weight allocation The significant regional characteristic response value : ； in, Indicates The weight of the feature, Represents the total dimension of the recursive feature; S524, the significant area feature response value Return to the recursive optimization function and continue recursive iteration until the preset number of iterations is reached .

Images