A radar high-resolution range profile open set identification method and device based on cluster boundary detection

Abstract

The application discloses a radar high-resolution range profile open set identification method and device based on clustering boundary detection. The steps include: establishing a training and test set of radar high-resolution range profile; data preprocessing is performed on the training and test set; a convolutional neural network is constructed; the convolutional neural network is trained by using the training set; features of the training set are extracted; the Euclidean length of the vector sum of the same category feature center and sample features to the remaining features is calculated; the part sample with the maximum vector sum Euclidean length is the clustering boundary sample; features of the test set are extracted; the minimum Euclidean distance of the test features to the feature center and the corresponding category are calculated; the Euclidean length of the vector sum of the test features to the training features of the minimum distance category is calculated, and the Euclidean length corresponding to the clustering boundary sample is compared to perform open set identification. The application can not only be used for identifying and classifying the known category targets in the library, but also can reject the unknown category targets outside the library, and the target identification accuracy and the automation and intelligent level of the radar are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar, and particularly relates to a radar high-resolution range profile open set identification method and device based on clustering boundary detection. BACKGROUND

[0002] The distance unit length R of the radar transmitted signal can be expressed as Wherein, c is the speed of light, and B is the radar transmitted signal bandwidth. The large bandwidth of the radar transmitted signal provides a higher distance resolution. When the wavelength of the wideband radar is much smaller than the target size, the target can be approximated by a scattering point model. At this time, the part of the target located in each distance unit will be replaced by a scattering center in the distance unit. The echoes of the scattering centers in different distance units jointly constitute the high-resolution range profile (HRRP) of the target. The HRRP of the target contains information such as the geometric structure of the target, and has important value for the identification and classification of the target.

[0003] In recent years, the target identification method for radar high-resolution range profile data is mainly for closed set identification. This identification requires that the data categories in the test sample set and the data categories in the training sample set remain consistent. However, in actual applications, the radar not only captures the high-resolution range profile of the in-library target, but also captures the high-resolution range profile of many out-of-library unknown category targets. In this case, the existing closed set identification algorithm cannot reject the out-of-library unknown category data, but will misjudge it as a certain category in the in-library, which greatly reduces the target identification accuracy of the radar.

[0004] Therefore, some researchers begin to study the open set identification of radar high-resolution range profile. For example, Chai Jing et al. proposed a multi-kernel support vector domain description (Multi kernel SVDD) model based on support vector domain description (SVDD) to more flexibly describe the multi-mode distribution of HRRP data in a high-dimensional feature space, thereby improving the identification and rejection performance of radar HRRP. Zhang Xuefeng et al. proposed a multi-classifier fusion algorithm based on maximum correlation classifier (Maximum correlation classifier, MCC), support vector machine (Support vector machine, SVM) and relevance vector machine (Relevance vector machine, RVM) to realize the rejection and identification function of radar HRRP. However, the above two algorithms need to rely on a specific form of kernel function to extract features, which limits the ability of the model to extract enough separable features, thereby affecting the accuracy of target identification and the intelligent level of the radar.

[0005] To overcome the performance disadvantage of feature extraction of traditional machine learning methods, some works use deep neural networks for open set recognition of HRRP. The most basic method is to train a neural network using cross-entropy, and then to reject out-of-database targets by dividing the sample classification probability threshold after Softmax normalization. However, this loss function makes the sample feature distribution not tight enough, thereby restricting the open set performance of the model, so some works propose an open set recognition model based on prototype learning, which can improve the open set performance by compacting the intra-class features and expanding the inter-class differences, but this method is not reasonable to divide the decision boundary around the prototype points of each class, which also limits the open set performance of the model. SUMMARY

[0006] To solve the problem that the open set performance of the model is limited in the prior art, thereby affecting the accuracy of target recognition and the intelligent level of the radar, the application provides a radar high-resolution range profile open set recognition method and device based on clustering boundary detection.

[0007] To achieve the above purpose, the technical scheme provided by the application is as follows:

[0008] The first aspect of the application provides a radar high-resolution range profile open set recognition method based on clustering boundary detection, comprising the following steps:

[0009] Step 1: Establish a sample set:

[0010] including a first training sample set and a first test sample set; wherein the training sample set includes radar high-resolution range profiles of a plurality of known category targets, and the test sample set includes radar high-resolution range profiles of a plurality of known category targets and radar high-resolution range profiles of unknown category targets outside the database;

[0011] Step 2: Preprocess the radar high-resolution range profiles in the first training sample set and the first test sample set to obtain a second training sample set and a second test sample set;

[0012] Step 3: Construct a convolutional neural network;

[0013] Step 4: Train the convolutional neural network using the second training sample set to obtain a trained convolutional neural network;

[0014] Step 5: Extract high-dimensional features of the second training sample set using the trained convolutional neural network;

[0015] Step 6: Calculate the feature center of the high-dimensional features of N samples of the same category in the second training sample set, and the sum of the vector of the high-dimensional features of each sample to the high-dimensional features of the remaining samples, and calculate the Euclidean length of the sum of the N category vectors;

[0016] Step seven, sort the Euclidean lengths of the vector sums of the N category samples, and select the largest number of category samples as the detected cluster boundary samples;

[0017] Step eight, extracting high-dimensional features of the second test sample set using the trained convolutional neural network;

[0018] Step nine, calculating the Euclidean distances of the high-dimensional features of the second test sample to the N feature centers of the second training sample set in step six, and determining the category represented by the minimum value;

[0019] Step ten, performing open set recognition on the high-dimensional features of the second test sample set using the cluster boundary detection result, to obtain the recognition result of the second test sample set.

[0020] Further, in the above step three, the convolutional neural network comprises three convolutional layers and one fully connected layer; the convolutional steps of the three convolutional layers are the same, each convolutional layer comprises a plurality of convolutional kernels, and each convolutional kernel has the same size; and the loss function of the convolutional neural network has the expression:

[0021]

[0022] wherein, Θ(x) and Θ(x k ) are the output results of the convolutional neural network, O i (i=1,…,k,…,N) are the i-th prototypes randomly initialized according to Gaussian distribution; d(Θ(x),O k ) is the Euclidean distance of Θ(x) to O k , and λ is a hyperparameter.

[0023] Further, the training process of the above step four comprises:

[0024] The sample data of the second training sample set is randomly divided into q batches, and the data of each batch is an n×D-dimensional matrix data; wherein, floor( ) represents rounding down, and P represents the number of high-resolution distance images in the second training sample set;

[0025] The sample data of each batch is sequentially input into the convolutional neural network for processing to obtain the output result of the convolutional neural network;

[0026] The value of the loss function is calculated according to the output result of the convolutional neural network and the loss function of the convolutional neural network, and the parameter value of the convolutional neural network is updated by using the stochastic gradient method until the network converges, to obtain the trained convolutional neural network.

[0027] Further, the preprocessing of step one comprises: sequentially performing center of gravity alignment and normalization processing on the radar high resolution range profile to obtain the preprocessed second training sample set and the second test sample set.

[0028] Further, in steps five and eight, the process of extracting high-dimensional features using the trained convolutional neural network comprises:

[0029] convolving and down-sampling the input current batch of sample set data using the first layer of convolutional layers to obtain a first feature map;

[0030] convolving and down-sampling the first feature map using the second layer of convolutional layers to obtain a second feature map;

[0031] convolving and down-sampling the second feature map using the third layer of convolutional layers to obtain a third feature map;

[0032] performing nonlinear transformation processing on the third feature map using the fourth layer of fully connected layers to obtain the output result of the current sample set data;

[0033] repeating the above steps until the processing of multiple batches of sample data is completed to obtain the output result of the convolutional neural network.

[0034] Further, in step six, the expression of the feature center of the high-dimensional features of the second training sample set is:

[0035]

[0036] wherein, represents the i-th sample in the k-th class in the second training sample set, and the k-th class has a total of N k samples;

[0037] The vector sum expression of the high-dimensional features of the i-th sample in the k-th class to the high-dimensional features of the remaining samples is:

[0038]

[0039] The Euclidean length expression is:

[0040]

[0041] Further, in step seven, the number of selected clustering boundary samples is:

[0042] m k = floor(βN k )

[0043] wherein, m krepresents the number of samples detected as cluster boundary samples in the kth category, β is a hyperparameter, and represents the proportion of cluster boundary samples in the sample N k .

[0044] Further, in the above step nine, high-dimensional features of the second test sample set are extracted using the trained convolutional neural network, and the Euclidean distances of the high-dimensional features of the second test sample to the feature centers of the training sample set are calculated to determine the category represented by the minimum value:

[0045]

[0046] Further, in the above step ten, the open set recognition result includes:

[0047] When is established, then the test sample corresponding to the high-dimensional feature of the second test sample set is an unknown category outside the database;

[0048] When is not established, then the category of the test sample corresponding to the high-dimensional feature of the second test sample set is

[0049] wherein, is the Euclidean length of the vector sum of the high-dimensional feature of the mth largest sample in the kth category to the high-dimensional features of the remaining samples. k

[0050] The second aspect of the present application provides a radar high-resolution range image open set recognition device based on cluster boundary detection, comprising:

[0051] A data acquisition module is configured to establish a first training sample set and a first test sample set; wherein the training sample set includes radar high-resolution range images of targets of a plurality of known categories, and the test sample set includes radar high-resolution range images of targets of a plurality of known categories and radar high-resolution range images of targets of an unknown category outside the database;

[0052] A preprocessing module is configured to preprocess the radar high-resolution range images in the training sample set and the test sample set to obtain a second training sample set and a second test sample set;

[0053] A model construction module is configured to construct a convolutional neural network;

[0054] A training module is configured to train the convolutional neural network using the second training sample set to obtain a trained convolutional neural network;

[0055] A first extraction module is configured to extract high-dimensional features of the second training sample set using the trained convolutional neural network;

[0056] ​a clustering boundary detection module, configured to calculate a feature center of the high-dimensional features of the second training sample set, and a Euclidean length of a vector sum of the high-dimensional features of each sample in the same category to the high-dimensional features of the remaining samples, and to detect the clustering boundary samples by sorting the Euclidean lengths;

[0057] a second extraction module, configured to extract the high-dimensional features of the second test sample set by using the trained convolutional neural network;

[0058] a target recognition module, configured to perform open set recognition on the high-dimensional features of the second test sample set by using the clustering boundary samples, and to obtain a recognition result of the second test sample set.

[0059] Compared with the prior art, the method has the following beneficial effects:

[0060] 1. The radar high-resolution range profile open set recognition method provided by the application extracts high-dimensional features of data by using a one-dimensional convolutional neural network, first, the data is preprocessed by energy normalization and alignment, which can mine high-level features of high-resolution range profile data, remove the amplitude sensitivity, translation sensitivity and posture sensitivity of radar high-resolution range profile data, and has strong robustness compared with the traditional direct classification method; second, the primary features of each layer can be combined to obtain higher-level features for recognition, and compared with the traditional machine learning feature extraction method, the network structure and the end-to-end nonlinear mapping method can help improve the target recognition rate of the model.

[0061] 2. The one-dimensional convolutional neural network is trained by using a loss function based on prototype learning, compared with the conventional cross-entropy loss function, the known class feature compactness in the feature space can be further improved, and the class distribution difference can be further expanded, so that the target recognition rate is further improved.

[0062] 3. The application detects the boundary samples of the known class feature clustering globally by calculating the Euclidean distance of the features of the to-be-tested sample to the features of all samples in the known class clustering, and can point-to-point detect the boundary samples of the known class feature clustering, and determine the clustering boundary to a specific sample point; on this basis, the features of the test sample are compared with the features of the detected clustering boundary samples, so as to judge the specific category of the to-be-tested target, compared with the previous distance measurement-based method, this strategy can effectively improve the recognition rate of the samples at the edge of the known class clustering, and has strong recognition robustness.

[0063] 4、The application puts forward innovative technical points of mutual combination and implementation in the two links of feature extraction and decision boundary from the basic principle of pattern recognition. Specifically, in the design of the feature extraction link, a one-dimensional convolutional neural network is trained by using a loss function based on prototype learning, which helps to extract more original data of separable high-dimensional features; in addition, in the design of the decision boundary link, an original clustering boundary detection algorithm is used to accurately detect the feature clustering boundary, thereby improving the classification accuracy of the test samples.

[0064] 5、Wide range of applications: not only can be used for identifying and classifying known category targets in the library, but also can reject unknown category targets outside the library, thereby improving the target recognition accuracy and effectively improving the automation and intelligence level of the radar. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a flowchart of a radar high-resolution range image open set recognition method based on clustering boundary detection provided by an embodiment of the application;

[0066] Figure 2 is a schematic diagram of a device for implementing the method of the application;

[0067] Figure 3 is the simulation test result of the embodiment. DETAILED DESCRIPTION

[0068] The application will be further described in detail below in combination with specific embodiments and drawings, but the embodiments of the application are not limited thereto.

[0069] Embodiment 1, as shown in the drawings, the application provides a radar high-resolution range image open set recognition method based on clustering boundary detection, comprising the following steps: Figure 1

[0070] Step one, establish a sample set:

[0071] including a first training sample set and a first test sample set; wherein the training sample set includes radar high-resolution range images of a plurality of known category targets, and the test sample set includes radar high-resolution range images of a plurality of known category targets and radar high-resolution range images of unknown category targets outside the database;

[0072] Step two, pre-process the radar high-resolution range images in the first training sample set and the first test sample set to obtain a second training sample set and a second test sample set;

[0073] Step three, construct a convolutional neural network;

[0074] Step four, train the convolutional neural network using the second training sample set to obtain a trained convolutional neural network;​

[0075] Step five, extracting high-dimensional features of the second training sample set by using the trained convolutional neural network;

[0076] Step six, calculating the feature center of the high-dimensional features of N samples of the same category in the second training sample set and the sum of the vectors of the high-dimensional features of each sample in N categories to the high-dimensional features of the remaining samples, and calculating the Euclidean length of the sum of N category vectors;

[0077] Step seven, sorting the Euclidean lengths of the sum of N category sample vectors, and selecting the largest number of category samples as detected cluster boundary samples;

[0078] Step eight, extracting high-dimensional features of the second test sample set by using the trained convolutional neural network;

[0079] Step nine, calculating the Euclidean distance of the high-dimensional features of the second test sample to the N feature centers of the second training sample set in step six, and determining the category represented by the minimum value;

[0080] Step ten, calculating the Euclidean length of the sum of all vectors of the high-dimensional features of the second test sample to the high-dimensional features of the second training sample of the category determined in step nine, comparing with the Euclidean length of the sum of the vectors of the number of cluster boundary samples in step seven, and performing open set recognition to obtain the recognition result of the second test sample set.

[0081] Embodiment 2: A radar high-resolution range profile open set recognition method based on cluster boundary detection, specifically comprising the following steps:

[0082] Step one, establishing a sample set:

[0083] Step 1-1, obtaining P radar high-resolution range profile original data of N categories as a first training sample set, wherein N≥3, P≥900;

[0084] Step 1-2, obtaining Q radar high-resolution range profile original data of N categories and R radar high-resolution range profile original data of M unknown categories as a second test sample set, wherein Q≥900, M≥1, L≥300.

[0085] The P radar high-resolution range profile original data and the Q radar high-resolution range profile original data are data of known categories in the database, and the L radar high-resolution range profile original data is data of unknown categories outside the database.

[0086] Step two, pre-process the radar high-resolution range profile sample data in the first training sample set and the first test sample set, in this embodiment, the pre-processing is to sequentially perform the gravity center alignment and normalization processing on the original data in the first training sample set and the first test sample set, to obtain the pre-processed training sample set and test sample set, that is, to obtain the second training sample set and the second test sample set after pre-processing.

[0087] Specifically, the original data in the first training sample set or the first test sample set is denoted as x0, first, the original data x0 is subjected to the gravity center alignment, to obtain the data x'0 after the gravity center alignment; then, the data x'0 after the gravity center alignment is subjected to the two-norm normalization processing, to obtain the sample data x after the normalization processing, and the expression is:

[0088]

[0089] Wherein, x can represent the sample data of the second training sample set x train or the sample data of the second test sample set x test , the second training sample set and the second test sample set are respectively PxD, (Q+R)xD dimensional matrices, wherein D represents the total number of distance units contained in one radar high-resolution range profile original data.

[0090] Step three, construct a convolutional neural network:

[0091] In this embodiment, the set convolutional neural network includes three convolutional layers and one fully connected layer, which are respectively denoted as the first layer convolutional layer, the second layer convolutional layer, the third layer convolutional layer and the fourth layer fully connected layer. Wherein, each convolutional layer has the same convolutional step, and each convolutional layer includes a plurality of convolutional kernels, and each convolutional kernel has the same size.

[0092] Specifically, for the first layer convolutional layer:

[0093] It is set to include C convolutional kernels, and the C convolutional kernels of the first layer convolutional layer are denoted as K, and the size of K is set to 1xwxl, w represents the window of each convolutional kernel in the first layer convolutional layer, 1<w<D; C is a positive integer greater than 0; the convolutional step of the first layer convolutional layer is set to L; at the same time, the kernel window size of the down-sampling processing of the first layer convolutional layer is all set to mxm, 1<m<D, D represents the total number of distance units contained in one radar high-resolution range profile imaging data in the second training sample set, and m is a positive integer greater than 0; the step of the down-sampling processing of the first layer convolutional layer is set to I, and I is equal to m in value.

[0094] The activation function of the first layer convolutional layer is set to x represents the pre-processed sample data (the sample data of the second training sample set or the second test sample set), represents the convolution operation, and b represents the all-1 bias of the third convolutional layer.

[0095] For the second convolutional layer:

[0096] The convolutional layer is configured to have C′ convolutional kernels. The C′ convolutional kernels of the second convolutional layer are denoted as K′, and their size is the same as that of the convolutional kernel K of the first convolutional layer. The convolutional stride of the second convolutional layer is denoted as L′, w≤L′≤Dw, and L′ is equal to the convolutional stride L of the first convolutional layer. At the same time, the kernel window size of the downsampling processing of the second convolutional layer is set to m′×m′, 1<m′<D, where m′ is a positive integer greater than 0. The stride of the downsampling processing of the second layer is I′, and I′ is equal to the value of m′.

[0097] Set the activation function of the second convolutional layer to... This represents the first feature map output by the first convolutional layer. b' represents the convolution operation, and b' represents the all-1 bias of the second convolutional layer.

[0098] For the third convolutional layer:

[0099] The convolutional layer is configured to have C″ convolutional kernels. The C″ convolutional kernels of the third convolutional layer are set to K″, with the same size as the kernel window size in the second convolutional layer. The convolutional stride of the third convolutional layer is set to L″, which is equal to the convolutional stride L′ of the second convolutional layer. The kernel window size for downsampling in the third convolutional layer is set to m"×m", where 1 < m" < D, and m" is a positive integer greater than 0. The stride for downsampling in the third layer is set to I″, which is equal to the value of m″.

[0100] Set the activation function of the second convolutional layer to... This represents the second feature map output by the second convolutional layer. represents the convolution operation, and b″ represents the all-1 bias of the third convolutional layer.

[0101] For the fourth fully connected layer:

[0102] Set its randomly initialized weight matrix It is a B×U dimensional matrix. `floor()` represents rounding down, `D` represents the total number of range cells contained in a single radar high-resolution range imaging data set in the second training sample, `B ≥ D`, where `B` is a positive integer greater than 0; the activation function is set to... This represents the third feature map output by the third convolutional layer. represents a full 1 bias of the fourth full connection layer, and is Ux1 dimensional.

[0103] After the model of the convolutional neural network is constructed, a loss function of the convolutional neural network is further constructed, and an expression of the loss function is as follows:

[0104]

[0105] wherein Θ(x) and Θ(x k ) are output results of the convolutional neural network, x k represents the kth category of training sample data; O i (i=1,...,k,...,N) are the ith prototypes randomly initialized according to a Gaussian distribution, and there are N prototypes in total; d(Θ(x),O k ) is the Euclidean distance from Θ(x) to O k , and λ is a hyperparameter.

[0106] Step four, training the convolutional neural network by using the second training sample set to obtain the trained convolutional neural network, and the specific steps include:

[0107] Step 4-1, randomly dividing the sample data in the second training sample set into q batches, and the data in each batch is an nxD dimensional matrix data; wherein, floor() represents rounding down, and P represents the number of high-resolution distance images in the second training sample set.

[0108] Step 4-2, sequentially inputting the sample data in each batch into the convolutional neural network for processing to obtain the output result of the convolutional neural network. Specifically, the following 5 steps are included:

[0109] Step 4-2-1, after inputting the sample data into the convolutional neural network, performing convolution and down-sampling processing on the current input sample data by using the first convolutional layer to obtain the first feature map.

[0110] Specifically, the input sample data x is convolved with the C convolution kernels of the first convolutional layer by using the convolution step length L of the first convolutional layer to obtain C convolution results of the first convolutional layer, and the C convolution results are recorded as C feature maps y of the first convolutional layer:

[0111]

[0112] performing Gaussian normalization processing on the C feature maps y of the first convolutional layer to obtain C feature maps

[0113] ​Each feature map in the first convolutional layer is downsampled to obtain C feature maps after downsampling. That is, the first feature map, is represented as:

[0114]

[0115] in, This represents the C feature maps of the first convolutional layer after Gaussian normalization, taken within the m×m kernel window size of the first downsampling processing layer. The maximum value, Let C represent the C feature maps of the first convolutional layer after Gaussian normalization.

[0116] Step 4-2-2: Use the second convolutional layer to perform convolution and downsampling on the first feature map to obtain the second feature map.

[0117] Specifically, the C feature maps downsampled by the first convolutional layer are used with the convolution stride L′ of the second convolutional layer. The first feature map is convolved with the C′ convolutional kernels K′ of the second convolutional layer, respectively, to obtain the result of the C′ convolutions of the second convolutional layer, which is denoted as the C′ feature maps of the second convolutional layer.

[0118] C′ feature maps of the second convolutional layer Gaussian normalization is performed to obtain C′ feature maps of the second convolutional layer after Gaussian normalization.

[0119] right Each feature map in the first convolutional layer is downsampled to obtain C′ feature maps after downsampling in the second convolutional layer. That is, the second feature map, represented as:

[0120]

[0121] in, This represents taking C′ feature maps of the second convolutional layer after Gaussian normalization within the kernel window size m′×m′ of the second downsampling processing. The maximum value, Let C′ represent the feature maps of the second convolutional layer after Gaussian normalization.

[0122] Step 4-2-3: Use the third convolutional layer to perform convolution and downsampling on the second feature map to obtain the third feature map.

[0123] Specifically, the C′ feature maps downsampled by the second convolutional layer are used with the convolution stride L′ of the third convolutional layer. The second feature map is convolved with the C″ convolutional kernels K″ of the third convolutional layer to obtain the result of the C″ convolutions of the third convolutional layer, which is denoted as the C″ feature maps of the third convolutional layer.

[0124] C″ feature maps of the third convolutional layer Gaussian normalization is performed to obtain C″ feature maps of the third convolutional layer after Gaussian normalization.

[0125] right Each feature map in the process is downsampled to obtain C″ feature maps after downsampling in the third convolutional layer. That is, the third feature map, represented as:

[0126]

[0127] in, This indicates that C″ feature maps of the third convolutional layer are taken within the kernel window size m″×m″ after Gaussian normalization during the third downsampling process. The maximum value, Let C″ represent the feature maps of the third convolutional layer after Gaussian normalization.

[0128] Step 4-2-4: Use the fourth fully connected layer to perform a nonlinear transformation on the third feature map to obtain the processing result of the current sample data. Its expression is:

[0129]

[0130] in, This represents the randomly initialized weight matrix of the fourth fully connected layer. This indicates an all-1 bias for the fourth fully connected layer.

[0131] Step 4-2-5: Repeat steps 4-2-1 to 4-2-4 until all batches of input data have been processed, and obtain the output of the convolutional neural network. This output is a high-dimensional feature.

[0132] Step 4-3: Calculate the value of the loss function based on the output of the convolutional neural network and the loss function of the convolutional neural network proposed in Step 3, and update the parameter values ​​of the convolutional neural network using the stochastic gradient method until the network converges, thus obtaining the trained convolutional neural network.

[0133] Specifically, the high-dimensional feature Θ(x The output result of the convolutional neural network is substituted into the loss function expression of the convolutional neural network, the value of the loss function is obtained, and the existing stochastic gradient method is used to update the parameter value of the convolutional neural network until the network converges, thereby obtaining the trained convolutional neural network. The stochastic gradient method herein belongs to mature prior art, and will not be specifically introduced in the embodiment.

[0134] The embodiment can mine high-level features of high-resolution range image data by using a multi-layer convolutional neural network structure and performing two-norm normalization and alignment preprocessing on data.

[0135] Step five, extracting high-dimensional features Θ(x train ) of the second training sample set x train ) by using the trained convolutional neural network.

[0136] The extraction process is the same as steps 4-1 to 4-2 in step four.

[0137] Step six, calculating the feature center of the high-dimensional features of N samples of the same category in the second training sample set and the sum of the vector of the high-dimensional feature of each sample in N categories to the high-dimensional features of the remaining samples, and calculating the Euclidean length of the sum of N category vectors.

[0138] Specifically, the feature center of the high-dimensional features of the sample data of each category of the second training sample set is calculated as follows:

[0139]

[0140] wherein, x represents the i-th sample in the k-th category in the second training sample set, and there are N k samples in this category.

[0141] The sum of the vector of the high-dimensional feature of the i-th sample in the k-th category to the high-dimensional features of the remaining samples can be expressed as:

[0142]

[0143] The Euclidean length can be expressed as:

[0144]

[0145] Step seven, sorting the Euclidean lengths of the N category sample vector sums, and selecting a part of samples with the largest values as detected cluster boundary samples:

[0146] The number of selected samples is:​​

[0147] m k = floor(βN k )

[0148] where m k represents the number of cluster boundary samples in the kth category, β is a hyperparameter, and represents the proportion of cluster boundary samples in the sample N k . The embodiment introduces the boundary detection result of the known category target feature cluster in the decision stage, compares the Euclidean length of the vector sum of the to-be-detected target feature to the known category feature with the detected boundary sample of the known category feature cluster, and determines the specific category of the to-be-detected target. Compared with the previous method based on classification probability and distance measurement, this strategy can effectively improve the recognition rate of samples at the boundary of the known category cluster and has strong recognition robustness.

[0149] Step eight, extracting the high-dimensional feature Θ(x test ) of the second test sample set x test using the trained convolutional neural network:

[0150] The extraction process is the same as steps 4-1 to 4-2 in step four.

[0151] Step nine, calculating the Euclidean distance of the high-dimensional feature of the second test sample to the N feature centers of the second training sample set in step six, and determining the category represented by the minimum value:

[0152]

[0153] Step ten, using the cluster boundary detection result to perform open set recognition on the high-dimensional feature of the second test sample set, and obtaining the recognition result of the second test sample set.

[0154] Specifically, calculating the Euclidean length of the vector sum of the high-dimensional feature Θ(x test ) to the high-dimensional features of the second training samples of the known categories in step nine and comparing the Euclidean length of the vector sum of the number of cluster boundary samples in step seven to perform open set recognition:

[0155] When is established, the test sample corresponding to the high-dimensional feature of the second test sample set is an unknown category outside the database;

[0156] When is not established, the category of the test sample corresponding to the high-dimensional feature of the second test sample set is

[0157] wherein, is the mthk Euclidean length of vector sum of high-dimensional features of each of the largest samples to high-dimensional features of the rest of the samples.

[0158] Embodiment 3, see Figure 2 The embodiment of the application also provides a radar high-resolution range profile open set identification method based on cluster boundary detection.

[0159] The data acquisition module is configured to establish a first training sample set and a first test sample set; wherein the training sample set comprises radar high-resolution range profiles of targets of known categories, and the test sample set comprises radar high-resolution range profiles of targets of known categories and radar high-resolution range profiles of targets of unknown categories outside the database.

[0160] The preprocessing module is configured to preprocess the radar high-resolution range profiles in the training sample set and the test sample set to obtain a second training sample set and a second test sample set.

[0161] The model construction module is configured to construct a convolutional neural network.

[0162] The training module is configured to train the convolutional neural network by using the second training sample set to obtain a trained convolutional neural network.

[0163] The first extraction module is configured to extract high-dimensional features of the second training sample set by using the trained convolutional neural network.

[0164] The cluster boundary detection module is configured to calculate feature centers of the high-dimensional features of the second training sample set, and Euclidean lengths of vector sums of high-dimensional features of each sample in the same category to high-dimensional features of the rest of the samples, and sort the Euclidean lengths, so that a part of samples with the largest numerical values are the detected cluster boundary samples.

[0165] The second extraction module is configured to extract high-dimensional features of the second test sample set by using the trained convolutional neural network.

[0166] The target identification module is configured to perform open set identification on the high-dimensional features of the second test sample set by using the cluster boundary samples to obtain an identification result of the second test sample set.

[0167] The radar high-resolution range profile open set identification device provided by the embodiment can implement the radar high-resolution range profile open set identification method provided by the first embodiment, and the detailed process is not described here.

[0168] The beneficial effects of the application are verified and described below through simulation tests.

[0169] 1. Simulation conditions

[0170] The hardware platform of the simulation experiment of the embodiment is:

[0171] Processor: Intel(R) Core(TM) i9-14900KF, main frequency is 3.20GHz, memory is 32GB.

[0172] The software platform of the simulation experiment of the embodiment is: Windows 11 operating system and python 3.12.

[0173] The data used in the simulation test this time is the high-resolution range image measured data of 13 types of aircraft. The 13 types of aircraft are An-26, Cessna, Yak-42, A319, A320, A330-2, A330-3, B737-8, CRJ-900, A321, A350-941, B737-7 and B747-89L.

[0174] From the above 13 types of aircraft, 7 types are randomly selected as known categories, and the remaining 6 types are unknown categories to make a training sample set and a test sample set. Among them, the training sample set has a total of 21000 samples, and each in-library category sample is about 3000; the test sample set includes 7000 samples of the known 7 categories in the library and 6000 samples of the unknown 6 categories outside the library, and each category sample is 1000.

[0175] Before the experiment, all the original data are preprocessed according to the step 2 in the above embodiment 2, and then the open set recognition experiment is performed using the present application.

[0176] 2. Simulation content and result analysis

[0177] The method of the present application and some traditional two-stage rejection recognition methods, as well as SoftMax threshold method and GCPL prototype method, are compared in this simulation experiment.

[0178] The traditional two-stage rejection recognition method mainly uses SVDD, OCSVM, I-Forest and other methods to reject the library outside target, and then uses SVM and other methods to further classify the target judged as in the library. The SoftMax threshold method determines the category of the to-be-tested target according to the size of the final classification output of the convolutional neural network. If the final network classification output is greater than the specified threshold, it is judged as a library-in category, otherwise it is judged as a library-out category. The GCPL prototype method determines the category of the to-be-tested target according to the distance from the feature of the to-be-tested sample to the nearest prototype center. If the distance is greater than the specified threshold, it is judged as a library-out category, otherwise it is judged as a library-in category.

[0179] The area AUC under the receiver operating characteristic curve (ROC) is used to evaluate the rejection ability of different methods on the out-of-library target, wherein the greater the value of AUC, the stronger the rejection ability on the out-of-library target.

[0180] Referring to Figure 3 It can be seen that in the simulation experiment, the rejection ability of the application on the out-of-library target is the strongest, followed by the GCPL prototype method and the SoftMax threshold method, and the rejection ability of the three traditional machine learning methods on the out-of-library target is general.

[0181] Since the data used in the simulation experiment has many categories, Macro Average F1-Score is used to comprehensively evaluate the open set recognition ability of different methods, wherein the greater the value of F1-Score, the stronger the open set recognition ability. The simulation experiment results are shown in the following table. It can be seen that in the simulation experiment, the comprehensive open set recognition ability of the application is the strongest, and is significantly better than the other several methods.

[0182] Method AUC F1-Score SVDD+SVM 0.514 0.577 OCSVM+SVM 0.511 0.576 I-Forest+SVM 0.586 0.582 SoftMax 0.772 0.756 GCPL 0.782 0.774 The invention 0.822 0.785

[0183] As can be seen from the above, the application achieves the optimal result in terms of the rejection ability on the out-of-library target and the comprehensive open set recognition ability, and proves that the method of the application has significant progress.

[0184] The above content is a further detailed description of the application in combination with the specific preferred embodiments, and cannot be regarded as the limitation of the specific implementation of the application. For ordinary skilled persons in the technical field to which the application belongs, some simple deductions or substitutions can be made without departing from the concept of the application, and all of them should be regarded as falling within the protection scope of the application.

Claims

1. A radar high resolution range profile open set identification method based on cluster boundary detection, characterized in that: The method comprises the following steps: Step one, establishing a sample set: including a first training sample set and a first test sample set; wherein the training sample set comprises radar high-resolution range images of known category targets, and the test sample set comprises radar high-resolution range images of known category targets and radar high-resolution range images of unknown category targets outside the database; Step two, preprocessing the radar high-resolution range images in the first training sample set and the first test sample set to obtain a second training sample set and a second test sample set; Step three, constructing a convolutional neural network; Step four, training the convolutional neural network using the second training sample set to obtain a trained convolutional neural network; Step five, extracting high-dimensional features of the second training sample set using the trained convolutional neural network; Step six, calculating the feature centers of the high-dimensional features of N samples of the same category in the second training sample set, and the sum of the high-dimensional features of each sample in N samples of the same category to the vectors of the high-dimensional features of the remaining samples, and calculating the Euclidean lengths of the sum of the N category vectors; Step seven, sorting the Euclidean lengths of the sum of the vectors of N category samples, and selecting a number of category samples with the largest values as detected clustering boundary samples; Step eight, extracting high-dimensional features of the second test sample set using the trained convolutional neural network; Step nine, calculating the Euclidean distances of the high-dimensional features of the second test sample to the N feature centers of the second training sample set in step six, and determining the category represented by the minimum value; Step ten, performing open set recognition on the high-dimensional features of the second test sample set using the clustering boundary detection result to obtain the recognition result of the second test sample set.

2. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 1, characterized in that: In step three, the convolutional neural network comprises three convolutional layers and one fully connected layer; the convolutional steps of the three convolutional layers are the same, each convolutional layer comprises a plurality of convolutional kernels, and each convolutional kernel has the same size; and the expression of the loss function of the convolutional neural network is: wherein Θ(x) and Θ(x k ) are output results of the convolutional neural network, O i (i = 1, …, k, …, N) is the i-th prototype randomly initialized according to a Gaussian distribution; d(Θ(x), O k ) is the Euclidean distance from Θ(x) to O k , and λ is a hyperparameter.

3. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 2, characterized in that: The training process of step four comprises: The sample data of the second training sample set is randomly divided into q batches, and the data of each batch is an n×D-dimensional matrix data; wherein, floor() represents rounding down, and P represents the number of high-resolution range images in the second training sample set. inputting each batch of sample data into the convolutional neural network for processing to obtain the output result of the convolutional neural network; calculating the value of the loss function according to the output result of the convolutional neural network and the loss function of the convolutional neural network, and updating the parameter value of the convolutional neural network using the stochastic gradient method until the network converges to obtain the trained convolutional neural network.

4. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 3, characterized in that: The preprocessing of step one comprises sequentially performing center alignment and normalization processing on the radar high-resolution range image to obtain the preprocessed second training sample set and second test sample set.

5. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 4, characterized in that: In steps five and eight, the process of extracting high-dimensional features using the trained convolutional neural network comprises: performing convolution and down-sampling processing on the input current batch of sample set data using the first layer convolutional layer to obtain a first feature map; performing convolution and down-sampling processing on the first feature map using the second layer convolutional layer to obtain a second feature map; performing convolution and down-sampling processing on the second feature map using the third layer convolutional layer to obtain a third feature map; The fourth full connection layer is used for performing nonlinear transformation on the third feature map, so as to obtain an output result of the current sample set data. The above steps are repeated until the processing of the multiple batches of sample data is completed, and an output result of the convolutional neural network is obtained.

6. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 5, characterized in that: In the step six, an expression of the feature center of the high-dimensional feature of the second training sample set is: wherein, represents the i-th sample in the k-th class in the second training sample set, the k-th class has a total of N k samples; The vector sum expression of the high-dimensional feature of the i-th sample in the k-th category to the high-dimensional features of the rest of the samples is: The Euclidean length expression is:

7. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 6, characterized in that: In the step seven, the number of selected clustering boundary samples is: m k = floor(βN k ) where m k represents the number of samples detected as cluster boundary samples in the kth category, and β is a hyperparameter representing the proportion of cluster boundary samples in the total number of samples N k .

8. The radar high resolution range profile open set identification method based on cluster boundary detection according to claim 7, characterized in that: In the step nine, the high-dimensional features of the second test sample set are extracted by using the trained convolutional neural network, and the Euclidean distances of the high-dimensional features of the second test sample to the feature centers of the training sample set are calculated, so as to determine the category represented by the minimum value:

9. The radar high resolution range profile open set recognition method based on cluster boundary detection according to claim 8, characterized in that: In the step ten, the open set recognition result includes: When is established, then the high-dimensional features corresponding to the test sample of the second test sample set are unknown categories outside the database. When is not true, then the class of the test sample corresponding to the high-dimensional feature of the second test sample set is wherein, is the Euclidean length of the vector sum of the high-dimensional features of the mth largest sample in the kth class to the high-dimensional features of the remaining samples. k is the Euclidean length of the vector sum of the high-dimensional features of the mth largest sample in the kth class to the high-dimensional features of the remaining samples. 10.A radar high resolution range profile open set identification device based on cluster boundary detection, characterized in that: including: The data acquisition module is configured to establish a first training sample set and a first test sample set; wherein the training sample set includes radar high-resolution range images of targets of a plurality of known categories, and the test sample set includes radar high-resolution range images of targets of a plurality of known categories and radar high-resolution range images of targets of unknown categories other than the database; The preprocessing module is configured to preprocess the radar high-resolution range images in the training sample set and the test sample set, so as to obtain a second training sample set and a second test sample set; The model construction module is configured to construct a convolutional neural network; The training module is configured to train the convolutional neural network by using the second training sample set, so as to obtain a trained convolutional neural network; The first extraction module is configured to extract high-dimensional features of the second training sample set by using the trained convolutional neural network; The clustering boundary detection module is configured to calculate the feature center of the high-dimensional features of the second training sample set, and the Euclidean lengths of the vector sum of the high-dimensional features of each sample in the same category to the high-dimensional features of the rest of the samples, and sort the Euclidean lengths, so that a part of samples with the maximum values are detected as the clustering boundary samples; The second extraction module is configured to extract high-dimensional features of the second test sample set by using the trained convolutional neural network; The target recognition module is configured to perform open set recognition on the high-dimensional features of the second test sample set by using the clustering boundary samples, so as to obtain a recognition result of the second test sample set.

Images