Intelligent target detection method based on two-dimensional difference characteristics of time-frequency domain signals
By constructing an intelligent target detection method based on the two-dimensional difference features of radar signals in the time and frequency domains and optimizing the parameters of deep neural networks, the problem of low accuracy in radar target detection is solved, and high-precision target detection in complex scenarios is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING INST OF REMOTE SENSING EQUIP
- Filing Date
- 2022-12-30
- Publication Date
- 2026-07-07
AI Technical Summary
Existing deep learning target detection algorithms suffer from low detection accuracy in radar target detection, especially in complex scenarios where high-precision target detection is difficult to achieve.
By constructing an intelligent target detection method based on the two-dimensional difference features of radar signals in the time and frequency domains, a deep neural network is used to optimize parameter updates. Combined with the two-dimensional difference features of time and frequency domain signals, a deep neural network model is designed to extract the target's distance and velocity parameters.
It effectively improves the accuracy and performance of radar target detection, especially enabling high-precision target detection in complex scenarios.
Smart Images
Figure CN116010849B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence target detection, and designs an intelligent target detection method based on the two-dimensional differential features of time-frequency domain signals. Background Technology
[0002] Deep learning algorithms can learn to select and extract sample features through multiple iterations and function mappings, and then use these extracted features for detection, effectively improving the accuracy and performance of object detection algorithms. In the field of object detection, the initial triumph of deep learning algorithms was achieved with the R-CNN model proposed by R. Girshick et al. in 2014, a groundbreaking algorithm in the field. This algorithm first introduced a two-stage object detection framework, which became the mainstream detection approach for several years afterward. This model extracts candidate regions of the target through heuristic search, then feeds all candidate regions as input into the convolutional network AlexNet to extract feature vectors. SVM is used for classification, and linear regression is used for offsets, ultimately yielding accurate category detection bounding boxes as the result. This algorithmic framework abandons the computationally complex, time-consuming, and weakly feature-expressive methods of traditional object detection, such as manually generated features and sliding windows, raising speed and accuracy to a new level and becoming the foundational model for many subsequent deep learning object detectors. Besides two-stage detectors that generate candidate boxes, there are also one-stage object detection algorithms in deep learning object detection. Popular methods include the YOLO series algorithms and the SSD algorithm. These algorithms do not generate candidate boxes; they directly predict the object's category and regress the bounding box position from the anchor boxes, improving detection speed by more than double. However, due to the lack of a candidate box filtering mechanism, all anchor boxes are considered as candidate boxes, which introduces a large number of negative samples, resulting in low detection accuracy. Anchor-free detection gradually emerged and matured starting in 2015. In that year, Baidu proposed the DenseBox model, pioneering the use of an end-to-end fully convolutional neural network to directly predict the bounding box of the object. In 2018, Hei et al. creatively proposed CornerNet, suggesting a method for localization by detecting a pair of corner points of the object's bounding box.
[0003] With the rapid development and excellent performance of intelligent target detection technology, radar target detection technology based on deep learning has become a new research hotspot in the field of radar applications. Research on radar target detection technology is of significant military importance, particularly for the precision guidance of missile-borne radar. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent target detection method based on the two-dimensional difference features of time-frequency domain signals. It aims to optimize the parameter update strategy of deep neural networks by utilizing radar target feature information and to solve the problem of high-precision target detection of missile-borne radar signals in complex scenarios using intelligent methods.
[0005] A smart target detection method based on two-dimensional difference features of time-frequency domain signals, the method comprising the following steps:
[0006] The first step is to acquire two-dimensional time-frequency domain data of the signal and perform intelligent target detection tasks;
[0007] The second step is to construct a training sample dataset for the time-frequency two-dimensional graph based on the time-frequency domain two-dimensional data of the signal.
[0008] The third step is to construct a label dataset based on the two-dimensional differential features of the signal, according to the training sample dataset.
[0009] The fourth step is to build a deep neural network model for the intelligent target detection method based on the intelligent target detection task.
[0010] The fifth step is to set the parameters of the deep neural network model and iteratively update the parameters of the deep neural network model according to the training sample dataset and the label dataset to obtain the intelligent target detection model.
[0011] The sixth step is to input the target detection task into the intelligent target detection model to obtain the target detection result.
[0012] In one embodiment, acquiring the two-dimensional time-frequency domain data of the signal includes:
[0013] To achieve intelligent target detection based on signal time-frequency domain images, a massive amount of complex scene echo signal time-frequency two-dimensional image sample simulation data is constructed for deep neural network model training, and the time-frequency two-dimensional image sample simulation data is used as signal time-frequency domain two-dimensional data.
[0014] In one embodiment, the second step, constructing the training sample dataset for the time-frequency two-dimensional graph based on the signal time-frequency domain two-dimensional data, includes:
[0015] The two-dimensional time-frequency domain data of the signal are calibrated based on the known distance and velocity information of the target.
[0016] The calibrated two-dimensional time-frequency domain data of the signal is used as the training sample dataset for the two-dimensional time-frequency graph, which is then used for training intelligent algorithms.
[0017] In one embodiment, the sixth step, inputting the target detection task into the intelligent target detection model to obtain the target detection result, includes:
[0018] The target detection task is input into the intelligent target detection model to obtain two-dimensional prediction information. The two-dimensional prediction information is then processed to obtain the target detection result.
[0019] In one embodiment, the sixth step, in which the target detection task is input into the intelligent target detection model to obtain the target detection result, includes:
[0020] The intelligent target detection model outputs two-dimensional prediction results. The center point screening technology is used to extract the target's category, position, and size prediction results on the time-frequency dimension image, and finally the target distance and velocity detection results are obtained.
[0021] In one embodiment, after inputting the target detection task into the intelligent target detection model to obtain two-dimensional prediction information, and processing the two-dimensional prediction information to obtain the target detection result, the method further includes: evaluating the model performance based on the target label features according to the template detection result.
[0022] In one embodiment, evaluating the model performance based on target label features includes: evaluating the performance of the intelligent target detection network model based on the target time-frequency domain data features and using an optimized intersection-union ratio (IU / R) technique.
[0023] In one embodiment, the third step, which involves constructing a label dataset based on two-dimensional differential features of the signal according to the training sample dataset, includes:
[0024] Based on the target location and size labels marked in the training sample dataset of the time-frequency two-dimensional graph, and using a Gaussian model, a label dataset with two-dimensional difference feature information is constructed based on the two-dimensional difference feature analysis of the time-frequency domain signal.
[0025] In one embodiment, the fourth step, which involves building a deep neural network model for the intelligent target detection method based on the intelligent target detection task, includes:
[0026] Design a deep neural network model, and use the deep neural network model to analyze the two-dimensional time-frequency domain data features of the signal in multiple dimensions to extract target distance and velocity parameter information.
[0027] In one embodiment, the parameters of the deep neural network model include: network model initialization parameters and training parameter values:
[0028] The initialization parameters of the network model include: the network convolutional layers are random values, so that the network model has indiscriminate general characteristics;
[0029] The network model training parameters include: batch size, loss function, optimization strategy, and learning rate.
[0030] This invention innovatively proposes an intelligent target detection method based on the two-dimensional difference features of time-frequency domain signals. By utilizing the two-dimensional difference feature information of radar signals in the time-frequency domain, the method for updating the parameters of the intelligent target detection network model is optimized, effectively improving the performance of intelligent radar target detection. Attached Figure Description
[0031] Figure 1 A flowchart illustrating an intelligent target detection method based on two-dimensional differential features of time-frequency domain signals. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0033] In one embodiment, an intelligent target detection method based on two-dimensional difference features of time-frequency domain signals includes the following steps:
[0034] The first step is to acquire two-dimensional time-frequency domain data of the signal and perform intelligent target detection tasks;
[0035] The second step is to construct a training sample dataset for the time-frequency two-dimensional graph based on the time-frequency domain two-dimensional data of the signal.
[0036] The third step is to construct a label dataset based on the two-dimensional differential features of the signal, according to the training sample dataset.
[0037] The fourth step is to build a deep neural network model for the intelligent target detection method based on the intelligent target detection task.
[0038] The fifth step is to set the parameters of the deep neural network model and iteratively update the parameters of the deep neural network model according to the training sample dataset and the label dataset to obtain the intelligent target detection model.
[0039] The sixth step is to input the target detection task into the intelligent target detection model to obtain the target detection result.
[0040] In one embodiment, acquiring the two-dimensional time-frequency domain data of the signal includes:
[0041] To achieve intelligent target detection based on signal time-frequency domain images, a massive amount of complex scene echo signal time-frequency two-dimensional image sample simulation data is constructed for deep neural network model training, and the time-frequency two-dimensional image sample simulation data is used as signal time-frequency domain two-dimensional data.
[0042] In one embodiment, the second step, constructing the training sample dataset for the time-frequency two-dimensional graph based on the signal time-frequency domain two-dimensional data, includes:
[0043] The two-dimensional time-frequency domain data of the signal are calibrated based on the known distance and velocity information of the target.
[0044] The calibrated two-dimensional time-frequency domain data of the signal is used as the training sample dataset for the two-dimensional time-frequency graph, which is then used for training intelligent algorithms.
[0045] In one embodiment, the sixth step, inputting the target detection task into the intelligent target detection model to obtain the target detection result, includes:
[0046] The target detection task is input into the intelligent target detection model to obtain two-dimensional prediction information. The two-dimensional prediction information is then processed to obtain the target detection result.
[0047] In one embodiment, the sixth step, in which the target detection task is input into the intelligent target detection model to obtain the target detection result, includes:
[0048] The intelligent target detection model outputs two-dimensional prediction results. The center point screening technology is used to extract the target's category, position, and size prediction results on the time-frequency dimension image, and finally the target distance and velocity detection results are obtained.
[0049] In one embodiment, after inputting the target detection task into the intelligent target detection model to obtain two-dimensional prediction information, and processing the two-dimensional prediction information to obtain the target detection result, the method further includes: evaluating the model performance based on the target label features according to the template detection result.
[0050] In one embodiment, evaluating the model performance based on target label features includes: evaluating the performance of the intelligent target detection network model based on the target time-frequency domain data features and using an optimized intersection-union ratio (IU / R) technique.
[0051] In one embodiment, the third step, which involves constructing a label dataset based on two-dimensional differential features of the signal according to the training sample dataset, includes:
[0052] Based on the target location and size labels marked in the training sample dataset of the time-frequency two-dimensional graph, and using a Gaussian model, a label dataset with two-dimensional difference feature information is constructed based on the two-dimensional difference feature analysis of the time-frequency domain signal.
[0053] In one embodiment, the fourth step, which involves building a deep neural network model for the intelligent target detection method based on the intelligent target detection task, includes:
[0054] A deep neural network model is designed to analyze the two-dimensional time-frequency domain data features of the signal from multiple dimensions, extracting target distance and velocity parameters. The network design techniques include designing a network structure model, employing hourglass network technology, and using center-attention-based width-height regression technology.
[0055] In one embodiment, the parameters of the deep neural network model include: network model initialization parameters and training parameter values:
[0056] The initialization parameters of the network model include: the network convolutional layers are random values, so that the network model has indiscriminate general characteristics;
[0057] The network model training parameters include: batch size, loss function, optimization strategy, and learning rate.
[0058] This invention innovatively proposes an intelligent target detection method based on the two-dimensional difference features of time-frequency domain signals. By utilizing the two-dimensional difference feature information of radar signals in the time-frequency domain, the method for updating the parameters of the intelligent target detection network model is optimized, effectively improving the performance of intelligent radar target detection.
[0059] In one embodiment, an intelligent target detection method based on two-dimensional difference features of time-frequency domain signals is characterized by the following specific steps:
[0060] The first step is to construct a massive time-frequency two-dimensional graph training sample dataset.
[0061] To address the need for intelligent target detection in time-frequency domain images of signals, a massive amount of complex scene echo signal time-frequency two-dimensional image sample simulation data is constructed for training deep neural network models. The training data is calibrated based on the actual distance of the target in the simulation data. The calibrated data sample set is used as the training sample dataset for intelligent target detection methods based on the two-dimensional differential features of time-frequency domain signals, and is used for training intelligent algorithms.
[0062] The second step is to construct a labeled dataset based on the two-dimensional differential features of the signal.
[0063] A two-dimensional label dataset is constructed by using the target location and size labels marked in the training sample dataset with time-frequency two-dimensional graphs. Based on the two-dimensional differential feature analysis of time-frequency domain signals, a two-dimensional Gaussian model method is used to assign target center points. Surrounding points Weights are assigned to construct a labeled dataset with two-dimensional differential feature information.
[0064] The third step is to build a deep neural network model for intelligent object detection.
[0065] Based on the time-frequency domain signal characteristics of radar, a deep neural network model structure for intelligent target detection is constructed. Through network design techniques, the two-dimensional time-frequency domain data features of the signal are analyzed from multiple dimensions, thereby effectively extracting target distance and velocity parameters. The network design techniques include designing a network structure model, employing hourglass network technology, and using width-height regression technology based on center attention.
[0066] The fourth step is to initialize the parameters and train the network model parameters for the intelligent object detection method.
[0067] To address the characteristics of two-dimensional time-frequency domain images in complex scenarios, based on the intelligent target detection method network model from the second step, parameters suitable for the data characteristics are set for the intelligent target detection method network model. These parameters include two parts: network model initialization parameters and training parameter values.
[0068] Initializing network model parameters includes setting random values for network convolutional layers to ensure the network model possesses indiscriminate general characteristics.
[0069] Network model training parameters include batch size, loss function, optimization strategy, and learning rate, which are set based on the data and network model characteristics. The loss function is determined by the center loss function. Dimensional loss and center offset loss It consists of three parts. It employs gradient descent and adaptive moment estimation techniques to optimize the objective function. .
[0070] Using the deep neural network model structure designed in step three and the network model parameters set in step four, the deep neural network model parameters are iteratively updated using the label dataset information and the deep neural network output information, and the network model parameters are saved as a reloadable file.
[0071] The fifth step is to process the two-dimensional prediction information output by the model, obtain the target detection results, and evaluate the model performance.
[0072] The center point screening technique is adopted, and the two-dimensional prediction results are output based on the network model. The prediction results of the target category, position and size on the time-frequency dimension image are extracted. The prediction results are processed according to the design parameters of the simulated radar system to finally obtain the target distance and velocity detection results. The cross-union ratio technique based on the target time-frequency domain data feature optimization is used to evaluate the performance of the intelligent target detection network model.
[0073] Thus, intelligent target detection based on the two-dimensional difference features of time-frequency domain signals has been completed.
[0074] This invention innovatively proposes an intelligent target detection method based on the two-dimensional difference features of time-frequency domain signals. By utilizing the two-dimensional difference feature information of radar signals in the time-frequency domain, the method for updating the parameters of the intelligent target detection network model is optimized, effectively improving the performance of intelligent radar target detection.
Claims
1. A smart target detection method based on two-dimensional difference features of time-frequency domain signals, characterized in that, The method includes the following steps: The first step is to acquire two-dimensional time-frequency domain data of the signal and intelligent target detection task. The acquisition of two-dimensional time-frequency domain data of the signal includes: constructing a massive amount of complex scene echo signal time-frequency two-dimensional map sample simulation data for deep neural network model training, and using the time-frequency two-dimensional map sample simulation data as the signal time-frequency domain two-dimensional data. The second step involves constructing a training sample dataset of time-frequency two-dimensional images based on the aforementioned signal time-frequency domain two-dimensional data. To meet the requirements of intelligent target detection tasks using signal time-frequency domain images, a massive amount of complex scene echo signal time-frequency two-dimensional image sample simulation data is constructed for training deep neural network models. The training data is then calibrated based on the actual distance of the target in the simulation data. This calibrated data sample set is used as the training sample dataset for an intelligent target detection method based on the two-dimensional differential features of time-frequency domain signals, for training the intelligent algorithm. The third step involves constructing a label dataset based on the two-dimensional difference features of the signal, using the target location and size labels marked in the time-frequency two-dimensional graph training sample dataset. Based on the two-dimensional difference feature analysis of the time-frequency domain signal, a two-dimensional Gaussian model method is used to assign the target center point. Surrounding points Assign weights to construct a labeled dataset with two-dimensional differential feature information; The fourth step involves constructing a deep neural network model for the intelligent target detection method based on the intelligent target detection task. This model is built upon the characteristics of radar time-frequency domain signals. Through network design techniques, the two-dimensional time-frequency domain data features of the signal are analyzed from multiple dimensions, thereby enabling the effective extraction of target distance and velocity parameter information. The network design techniques include designing a network structure model, employing hourglass network technology, and using center-attention-based width-height regression technology. The fifth step involves setting the parameters of the deep neural network model and iteratively updating these parameters based on the training sample dataset and the label dataset to obtain an intelligent object detection model. Considering the characteristics of two-dimensional time-frequency domain images in complex scenes, and based on the deep neural network model of the intelligent object detection method in the fourth step, parameters suitable for the data characteristics of the network model are set. These parameters include two parts: network model initialization parameters and training parameter values. Initialization parameters include setting random values for the network convolutional layers to ensure the network model possesses indiscriminate general characteristics. Training parameters include batch size, loss function, optimization strategy, and learning rate, set in conjunction with the data and network model characteristics. The loss function is determined by the central loss function. Dimensional loss and center offset loss It consists of three parts and employs gradient descent and adaptive moment estimation techniques to optimize the objective function. ; The sixth step is to input the target detection task into the intelligent target detection model to obtain the target detection result.
2. The method according to claim 1, characterized in that, The sixth step involves inputting the target detection task into the intelligent target detection model to obtain the target detection results, including: The target detection task is input into the intelligent target detection model to obtain two-dimensional prediction information. The two-dimensional prediction information is then processed to obtain the target detection result.
3. The method according to claim 2, characterized in that, In the sixth step, inputting the target detection task into the intelligent target detection model to obtain the target detection result includes: The intelligent target detection model outputs two-dimensional prediction results. The center point screening technology is used to extract the target's category, position, and size prediction results on the time-frequency dimension image, and finally the target distance and velocity detection results are obtained.
4. The method according to claim 3, characterized in that, The step of inputting the target detection task into the intelligent target detection model to obtain two-dimensional prediction information, processing the two-dimensional prediction information to obtain the target detection result, further includes: evaluating the model performance based on the target label features according to the target detection result.
5. The method according to claim 4, characterized in that, The evaluation of model performance based on target label features includes: evaluating the performance of the intelligent target detection network model based on the target time-frequency domain data features and using an optimized intersection-union ratio (IU / R) technique.