Radar multi-feature graph representation and graph network fusion target detection system and method

Abstract

The invention relates to a radar multi-feature graph representation and graph network fusion target detection system and method, and belongs to the technical field of radar target detection methods. In order to solve the problem that in the prior art, a single-feature detection method is limited in target and background distinguishing capacity under a complex sea clutter background, the target detection system comprises a multi-feature extraction network and a graph fusion detection network, the multi-feature extraction network is used for extracting a plurality of hidden features from different model features of the same signal sample for graph representation to obtain multi-feature graph data MFG corresponding to each sample; and the graph fusion detection network performs feature extraction and aggregation on the multi-feature graph data to realize feature fusion detection. According to the target detection system and method, in different environment target detection experiments, the model generalization ability can be remarkably improved through multiple data sets, and the detection probability is improved by about 8%.

Description

Technical Field

[0001] The present invention relates to a target detection system and method, in particular to a radar multi-feature graph representation and graph network fusion target detection system and method, belonging to the technical field of radar target detection methods. Background Art

[0002] Affected by the time-varying and non-stationary nature of sea clutter, as well as the diversity of target types and motion states, the target detection method based on feature models and deep learning data-driven has limited ability to distinguish between targets and clutter signals in actual observation environments, and the detection performance is difficult to further improve. Many scholars have used the complementary effects of different features to conduct research from the perspective of multi-feature joint detection, providing a new way to improve detection performance. Therefore, feature detection has gradually evolved from single feature detection to multi-feature joint detection. The current multi-feature joint detection method is mainly based on convex hull learning algorithms, which are difficult to use in high-dimensional feature spaces. Among them, the time-frequency feature reflects the time-varying information of the frequency distribution of the sea surface background and the target. In the context of the gradual improvement of radar resolution, it provides effective support for the development of sea surface target detection methods. Micro-Doppler theoretical research shows that sea surface targets have micro-motion characteristics, which reflects the change of the target radial velocity over time, as well as the change of the target's attitude (such as roll, pitch, yaw, etc.) affected by waves, providing more information that can be used for detection. Therefore, the time-frequency feature is widely used in sea surface target detection by enhancing the difference in the trend of radial velocity change between the target and sea clutter.

[0003] However, the time-frequency feature detection method has similar problems to other feature detection methods. Due to the complex sea detection environment of radar and the diverse target characteristics, the target micro-motion characteristics are time-varying and do not show regular frequency modulation periodic characteristics, which affects the performance of the detection method. In addition, affected by factors such as sea spikes, sometimes the sea clutter signal has two-dimensional time-frequency characteristics similar to the target. For example, some scholars use CNN to process the time-frequency characteristics of several types of micro-motion targets to achieve target recognition and obtain good recognition performance. However, when analyzing the sea radar echo data, it is found that the time-frequency characteristics are unstable in distinguishing sea clutter and target signals. This performance difference is mainly related to the background characteristics. In the process of sea surface target detection, the target is affected by factors such as waves, the echo fluctuation is complex, and discontinuities may occur. In the time-frequency characteristics, the target signal energy concentration distribution area is discontinuous in time. Moreover, the target micro-motion state is related to the fluctuation of the sea surface, and its time-frequency spectrum correlation information is complex, lacking effective quantification or modeling methods. In addition, when the target speed is low or the angle between the target heading and the observation direction is large, resulting in low radial speed, it is easier to overlap with clutter in the time-frequency diagram and difficult to be effectively detected. Therefore, in the actual application environment of sea observation, a single time-frequency feature is sometimes difficult to distinguish between targets and clutter.

[0004] Under actual detection conditions, affected by complex environments and target characteristics, there are still many problems in the detection of amplitude-time-frequency features. The amplitude sequence feature significantly reduces the ability to distinguish between targets and clutter signals under conditions such as high sea conditions or weak targets, and the STFT time-frequency feature is also difficult to distinguish between targets and clutter signals under many conditions, such as Figure 1 As shown in Figure 2. Clutter may overwhelm the target because the clutter signal may have a wide Doppler spectrum and the radial velocity of the target may be low during movement. In addition, clutter sometimes exhibits time-frequency characteristics similar to those of the target, which can easily lead to missed alarms and false alarms. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the single feature detection method in the prior art has limited ability to distinguish targets and backgrounds under complex sea clutter background. From the perspective of neural network data-driven detection method, based on time-frequency features, a radar multi-feature graph representation and graph network fusion target detection system and method are proposed. The feature fusion detection method of multi-feature extraction network (Multi-Feature Extraction Network, MFEn) and graph fusion detection network (Graph Fusion and Detection Network, GFDn) is adopted to perform graph representation on multiple features of signal samples, and multi-feature fusion detection is realized through graph classification.

[0006] In order to solve the above problems, the present application is implemented through the following technical solutions: A radar multi-feature graph representation and graph network fusion target detection system, which is special in that it includes a multi-feature extraction network MFEn and a graph fusion detection network GFDn. The multi-feature extraction network MFEn is used to extract multiple latent features from different model features of the same signal sample for graph representation, and obtain multi-feature graph data MFG corresponding to each sample; The graph fusion detection network GFDn performs feature extraction and aggregation on multi-feature graph data to achieve feature fusion detection.

[0007] Furthermore, the multi-feature extraction network MFEn is composed of the encoder part of the convolutional autoencoder CAE, including 2 convolutional layers, 2 pooling layers and 1 fully connected layer, wherein the first convolutional layer has 64 3×3 convolutional kernels, the second convolutional layer has 128 3×3 convolutional kernels, and the fully connected layer outputs the latent features of 256-dimensional vectors; Furthermore, the multi-feature graph data MFG is expressed as , is a node set, is the feature matrix, is the edge set; Node set of multi-feature graph data MFG Each node in corresponds to a hidden feature of the sample; the feature matrix of MFG , is the set of all hidden feature information, in: is the node dimension of MFG, is the dimension of node features; The nodes are connected to form a fully connected graph, that is, the edge set After being expressed in the form of an adjacency matrix, it becomes a matrix of all 1s; Furthermore, the graph fusion detection network GFDn is composed of a feature fusion network and a feature detection network; Furthermore, the feature fusion network consists of a graph attention convolution layer, a graph pooling layer, and a feature reading module; The graph attention convolution layer and the graph pooling layer are each provided with three layers, which are arranged alternately. A feature reading module is provided after the first two graph pooling layers, and the embedded features obtained by splicing the output features of each graph pooling layer are input into the feature detection network; In the graph attention convolution layer, each node is used as a central node, and its features are fused with other node features with the corresponding attention coefficient; the graph pooling layer determines the nodes to be retained in the layer according to the pooling attention coefficient of each node. After multiple pooling layers, the multiple feature fusion results of a node are finally retained as the final fused embedded features; the graph attention convolution layer realizes the fusion of input feature class variables; Furthermore, the specific working process of the graph pooling layer is as follows: Sample features in the graph pooling process The node dimension, that is, the number of hidden features, is finally obtained from down to 1; In the In the graph pooling layer, first pass the network parameters of this layer And the node features of the input layer Calculate the pooling attention coefficient for each node : , Then according to the set graph pooling rate Get the index of the reserved node of this layer : , Where: is the image attention mask; Where: is the feature of each node according to the node index, Represents the operation of retaining node features of some nodes according to the mask, Represents the adjacency matrix of the graph data after retaining some nodes according to the index, and Output the node features and adjacency matrix of the graph data for this layer; After each pooling layer, the output features of the layer are read through the feature reading operation: ; Where: is the output feature, is the graph pooling rate, Input the number of graph data nodes for this layer, is the i-th node feature; Furthermore, the feature detection network is a multi-layer perceptron module MLP composed of two fully connected layers, the output dimensions of the two fully connected layers are 64 and 2 respectively, and the fused embedded features are detected.

[0008] The method for target detection using the above system is special in that it includes the following steps: Step 1: Construct a data set: take sample data at equal intervals on a long continuous pulse sequence with a certain pulse interval as the step length as the data set; Step 2, analyzing the sample data obtained in step 1 to obtain multiple model features of the signal; Step 3: Use the signal multiple model features in the training set to train MFEn-GFDn to obtain the trained MFEn-GFDn; Step 4: Use the signal multiple model features in the test set to test the trained MFEn-GFDn.

[0009] Further, in said step 1: the sample data corresponds to signals with overlapping parts; The data set contains both clutter units and target units, and the data set is divided into a training set and a test set, and the ratio of sample data in the training set to that in the test set is 2:1; The data in the data set includes radar signal data of four polarization modes: HH, HV, VH and VV; Further, the step 2 is specifically as follows: a total of 12 model features of four polarization mode signals are obtained by three time-frequency analysis methods: short Fourier Transform (STFT), smoothed pseudo Wigner-Ville Distribution (SPWVD), and Wigner-Ville Distribution (WVD); Furthermore, the step 3 is specifically as follows: using different learning rate, batch, and epoch combinations to ensure that MFEn-GFDn converges during the test process, and selecting a set of parameter combinations with the highest performance to adjust MFEn-GFDn.

[0010] This application uses the CAE structured MFEn to extract features from multiple time-frequency graphs of radar signals, constructs the radar signal MFG, and then uses GFDn to perform fusion detection on the MFG containing multiple feature information. In experiments on target detection in different environments, the generalization ability of the model can be significantly improved through multiple data sets. The MFEn-GFDn trained with a training set consisting of a mixture of samples from three data sets has an approximately 8% higher detection probability than DCCNN. In addition, MFEn-GFDn further improves detection performance by expanding feature dimensions, especially in environments where there is a lack of corresponding training samples, and has higher generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 The low-resolution STFT time-frequency characteristic diagram, where (a) is a schematic diagram of wide Doppler spectrum clutter, (b) is a schematic diagram of target-like Doppler characteristic clutter, (c) is a schematic diagram of low signal-to-clutter ratio target, and (d) is a schematic diagram of low-speed target; Figure 2 The network structure and data processing flow of the graph network multi-feature fusion detection method; Figure 3 This is the graph fusion detection network structure diagram; Figure 4 Schematic diagram of the graph pooling layer; Figure 5 are multi-feature radar signal data, where (a) is a schematic diagram of short-time Fourier transform of HH polarization mode, (b) is a schematic diagram of short-time Fourier transform of HV polarization mode, (c) is a schematic diagram of short-time Fourier transform of VH polarization mode, (d) is a schematic diagram of short-time Fourier transform of VV polarization mode, (e) is a schematic diagram of Wigner-Wiley distribution of HH polarization mode, (f) is a schematic diagram of Wigner-Wiley distribution of HV polarization mode, (g) is a schematic diagram of Wigner-Wiley distribution of VH polarization mode, (h) is a schematic diagram of Wigner-Wiley distribution of VV polarization mode, (i) is a schematic diagram of smoothed pseudo-Wigner-Wiley distribution of HH polarization mode, (j) is a schematic diagram of smoothed pseudo-Wigner-Wiley distribution of HV polarization mode, (k) is a schematic diagram of smoothed pseudo-Wigner-Wiley distribution of VH polarization mode, and (l) is a smoothed pseudo-Wigner-Wiley distribution of VV polarization mode; Figure 6 Schematic diagram of detection probability for different training data models; Figure 7 Schematic diagram of false alarm loss for different training data models. DETAILED DESCRIPTION

[0012] The following is a specific implementation of the present invention with reference to the accompanying drawings to further illustrate the composition of the present invention.

[0013] Embodiment 1. A radar multi-feature graph representation and graph network fusion target detection system, such as Figure 2 As shown, it includes a multi-feature extraction network MFEn and a graph fusion detection network GFDn. The multi-feature extraction network MFEn is used to extract multiple latent features from different model features of the same signal sample for graph representation, and obtain multi-feature graph data MFG corresponding to each sample; The graph fusion detection network GFDn performs feature extraction and aggregation on multi-feature graph data to achieve feature fusion detection.

[0014] Furthermore, the multi-feature extraction network MFEn is composed of the encoder part of the convolutional autoencoder CAE, including 2 convolutional layers, 2 pooling layers and 1 fully connected layer, wherein the first convolutional layer has 64 3×3 convolutional kernels, the second convolutional layer has 128 3×3 convolutional kernels, and the fully connected layer outputs the latent features of 256-dimensional vectors; Furthermore, the multi-feature graph data MFG is expressed as , is a node set, is the feature matrix, is the edge set; Node set of multi-feature graph data MFG Each node in corresponds to a hidden feature of the sample; the feature matrix of MFG , is the set of all hidden feature information, in: is the node dimension of MFG, is the dimension of node features; The nodes are connected to form a fully connected graph, that is, the edge set After being expressed in the form of an adjacency matrix, it becomes a matrix of all 1s; Furthermore, if Figure 3 As shown, the graph fusion detection network GFDn is composed of a feature fusion network and a feature detection network; Furthermore, the feature fusion network consists of a graph attention convolution layer, a graph pooling layer, and a feature reading module; The graph attention convolution layer and the graph pooling layer are each provided with three layers, which are arranged alternately. A feature reading module is provided after the first two graph pooling layers, and the embedded features obtained by splicing the output features of each graph pooling layer are input into the feature detection network; In the graph attention convolution layer, each node is used as a central node, and its features are fused with other node features with the corresponding attention coefficient; the graph pooling layer determines the nodes to be retained in the layer according to the pooling attention coefficient of each node. After multiple pooling layers, the multiple feature fusion results of a node are finally retained as the final fused embedded features; the graph attention convolution layer realizes the fusion of input feature class variables; Furthermore, if Figure 4 As shown, the specific working process of the graph pooling layer is as follows: Sample features in the graph pooling process The node dimension, that is, the number of hidden features, is finally obtained from down to 1; In the In the graph pooling layer, first pass the network parameters of this layer And the node features of the input layer Calculate the pooling attention coefficient for each node : , Then according to the set graph pooling rate Get the index of the reserved node of this layer : , Where: is the image attention mask; , Where: is the feature of each node according to the node index, Represents the operation of retaining node features of some nodes according to the mask, Represents the adjacency matrix of the graph data after retaining some nodes according to the index, and Output the node features and adjacency matrix of the graph data for this layer; After each pooling layer, the output features of the layer are read through the feature reading operation: ; Where: is the output feature, is the graph pooling rate, Input the number of graph data nodes for this layer, is the i-th node feature; Furthermore, the feature detection network is a multi-layer perceptron module MLP composed of two fully connected layers, the output dimensions of the two fully connected layers are 64 and 2 respectively, and the fused embedded features are detected.

[0015] Embodiment 2. A method for performing target detection using the above-mentioned system, such as Figure 2 As shown, the following steps are included: Step 1: Construct a data set: take sample data at equal intervals on a long continuous pulse sequence with a certain pulse interval as the step length as the data set; The sample data corresponds to signals with overlapping parts; The data set contains both clutter units and target units, and the data set is divided into a training set and a test set, and the ratio of sample data in the training set to that in the test set is 2:1; The data in the data set includes radar signal data of four polarization modes: HH, HV, VH and VV; Step 2, analyzing the sample data obtained in step 1 to obtain multiple model features of the signal; A total of 12 model features of four polarization signals are obtained through three time-frequency analysis methods: Short Fourier Transform (STFT), Smoothed Pseudo Wigner-Ville Distribution (SPWVD), and Wigner-Ville Distribution (WVD), such as Figure 5 As shown; Step 3: Use the signal multiple model features in the training set to train MFEn-GFDn to obtain the trained MFEn-GFDn; Different learning rate, batch, and epoch combinations are used to ensure that MFEn-GFDn converges during the test process, and a set of parameter combinations with the highest performance are selected to adjust MFEn-GFDn; Step 4: Use the signal multiple model features in the test set to test the trained MFEn-GFDn.

[0016] Embodiment 3: This embodiment is a simulation experiment of Embodiment 1 and Embodiment 2.

[0017] In this embodiment, the performance of the proposed method is tested by measuring the IPIX dataset of resident mode radar signals. The resident mode radar signal has multiple coherent pulse signals within a long observation time, and multiple features of the detection samples can be extracted to verify the performance of the proposed method.

[0018] Table 1 IPIX radar parameters and data information , Each sample consists of 1024 continuous pulses. Samples are taken at equal intervals on a long continuous pulse sequence with a step length of 10 or 50 pulses, and the corresponding signals of the samples have overlapping parts. The IPIX data contains the radar echo signals in the dwell mode collected within 131 seconds. The training set samples are constructed with the signals in the first 90 seconds of observation time, and the last 41 seconds correspond to the test set samples.

[0019] The data includes radar signal data in four polarization modes: HH, HV, VH and VV. The model features used in this embodiment include three time-frequency diagrams of four types of polarization signals, STFT, WVD and SPWVD, a total of 12 features such as Figure 5 shown.

[0020] For the MFEn-GFDn detection method, the Figure 5 A variety of time-frequency model features are used for testing.

[0021] The experimental environment is TensorFlow1.13. After testing, the training parameters are finally determined, the batch size is 32, the learning rate is 0.01, epoch=10000, the parameter initialization uses Xavier, and the bias is initialized to 0. During the parameter initialization process, the parameter optimization strategy uses the gradient descent optimizer and the MSE loss function. GFDn uses the ADAM optimizer and the cross entropy loss function.

[0022] B. Analysis of data detection performance in different environments Four training data sets are constructed, which are composed of samples of a single IPIX01-03 data set and samples of the three data sets mixed in equal amounts, as shown in Table 2.

[0023] Table 2 Description of training sets for different sea conditions , Datasets 1, 2, and 3 are composed of observation data in a single environment, respectively, and dataset 4 is composed of an equal mixture of data from three environments. DCCNN and MFEn-GFDn are trained, and their object detection performance is tested on the IPIX03 test set.

[0024] Figure 6 Detection probability-false alarm rate curve for different detection methods. Figure 7 is the false alarm loss curve. The farther the curve is from the dotted line, the more significant the false alarm loss is. Figure 6 and Figure 7 As shown in Figure 3, the network trained with mixed data from training set 4 can achieve significantly improved detection performance and reduce false alarm loss compared to the single environment training set 3. Figure 6 As shown, the red curve represents the detection probability of MFEn-GFDn, and the false alarm rate is 10 -3The detection probability is about 0.68, while the blue curve represents the detection probability of the DCCNN method, with a false alarm rate of 10 -3 When the detection probability is 0.60, it can be seen that the detection probability of MFEn-GFDn is improved by 0.68-0.60=0.08 compared with the DCCNN method, that is, the detection probability of MFEn-GFDn is improved by about 8% compared with DCCNN. Figure 7 As shown in the figure, the proposed MFEn-GFDn can further suppress the false alarm loss, and the deviation between the actual false alarm rate and the set false alarm rate is smaller.

[0025] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A radar multi-feature graph representation and graph network fusion target detection system, characterized by: Including multi-feature extraction network MFEn and graph fusion detection network GFDn, The multi-feature extraction network MFEn is used to extract multiple latent features from different model features of the same signal sample for graph representation, and obtain multi-feature graph data MFG corresponding to each sample; The graph fusion detection network GFDn performs feature extraction and aggregation on multi-feature graph data to achieve feature fusion detection.

2. According to claim 1, a radar multi-feature graph representation and graph network fusion target detection system is characterized by: The multi-feature extraction network MFEn is composed of the encoder part of the convolutional autoencoder CAE, including 2 convolutional layers, 2 pooling layers and 1 fully connected layer, wherein the first convolutional layer has 64 3×3 convolutional kernels, the second convolutional layer has 128 3×3 convolutional kernels, and the fully connected layer outputs the latent features of a 256-dimensional vector.

3. According to claim 2, a radar multi-feature graph representation and graph network fusion target detection system is characterized by: The multi-feature graph data MFG is expressed as , is a node set, is the feature matrix, is the edge set; Node set of multi-feature graph data MFG Each node in corresponds to a hidden feature of the sample; the feature matrix of MFG , is the set of all hidden feature information, in: is the node dimension of MFG, is the dimension of node features; The nodes are connected to form a fully connected graph, that is, the edge set When expressed in adjacency matrix form, it is an all-1 matrix.

4. A radar multi-feature graph representation and graph network fusion target detection system according to claim 1, 2 or 3, characterized in that: The graph fusion detection network GFDn is composed of a feature fusion network and a feature detection network.

5. According to claim 4, a radar multi-feature graph representation and graph network fusion target detection system is characterized by: The feature fusion network consists of a graph attention convolution layer, a graph pooling layer, and a feature reading module; The graph attention convolution layer and the graph pooling layer are each provided with three layers, which are arranged alternately. A feature reading module is provided after the first two graph pooling layers, and the embedded features obtained by splicing the output features of each graph pooling layer are input into the feature detection network; In the graph attention convolution layer, each node is used as a central node, and its features are fused with other node features with the corresponding attention coefficient; The graph pooling layer determines the nodes to be retained in the layer according to the pooling attention coefficient of each node. After multiple pooling layers, the multiple feature fusion results of a node are finally retained as the final fused embedded features; The graph attention convolution layer realizes the fusion of input feature class variables.

6. The radar multi-feature graph representation and graph network fusion target detection system according to claim 5 is characterized by: The specific working process of the graph pooling layer is as follows: Sample features in the graph pooling process The node dimension, that is, the number of hidden features, is finally obtained from down to 1; In the In the graph pooling layer, first pass the network parameters of this layer And the node features of the input layer Calculate the pooling attention coefficient for each node : , Then according to the set graph pooling rate Get the index of the reserved node of this layer : ， Where: is the image attention mask; ， Where: is the feature of each node according to the node index, Represents the operation of retaining node features of some nodes according to the mask, Represents the adjacency matrix of the graph data after retaining some nodes according to the index, and Output the node features and adjacency matrix of the graph data for this layer; After each pooling layer, the output features of the layer are read through the feature reading operation: ， Where: is the output feature, is the graph pooling rate, Input the number of graph data nodes for this layer, is the feature of the i-th node.

7. The radar multi-feature graph representation and graph network fusion target detection system according to claim 4 is characterized by: The feature detection network is a multi-layer perceptron module MLP composed of two fully connected layers, the output dimensions of the two fully connected layers are 64 and 2 respectively, and the fused embedded features are detected.

8. A method for target detection using a radar multi-feature graph representation and graph network fusion target detection system as described in claim 4, characterized in that: The following steps are involved: Step 1: Construct a data set: take sample data at equal intervals on a long continuous pulse sequence with the pulse interval as the step length as the data set; Step 2, analyzing the sample data obtained in step 1 to obtain multiple model features of the signal; Step 3: Use the signal multiple model features in the training set to train MFEn-GFDn to obtain the trained MFEn-GFDn; Step 4: Use the signal multiple model features in the test set to test the trained MFEn-GFDn.

9. The method of radar multi-feature graph representation and graph network fusion target detection according to claim 8 is characterized by: In the step 1: the sample data corresponds to signals with overlapping parts; The data set contains both clutter units and target units, and the data set is divided into a training set and a test set, and the ratio of sample data in the training set to that in the test set is 2:1; The data in the data set includes radar signal data in four polarization modes: HH, HV, VH and VV.

10. The method of radar multi-feature graph representation and graph network fusion target detection according to claim 8, characterized in that: The step 2 is specifically as follows: obtaining a total of 12 model features of four polarization mode signals through three time-frequency analysis methods: short-time Fourier transform, smoothed pseudo Wigner-Wiley distribution, and Wigner-Wiley distribution; The step 3 is specifically as follows: using different learning rate, batch, and epoch combinations to ensure that MFEn-GFDn converges during the test process, and selecting a set of parameter combinations with the highest performance to adjust MFEn-GFDn.

Images