Radar high-resolution range profile open set identification method and device based on clustering boundary detection

Abstract

The invention discloses a radar high-resolution range profile open set identification method and device based on clustering boundary detection. The method comprises the following steps: establishing a training and testing set of a radar high-resolution range profile; performing data preprocessing on the training and testing set; constructing a convolutional neural network; training a convolutional neural network by using the training set; extracting features of the training set; calculating the Euclidean length of the vector sum from the same-class feature center of the training set and the sample feature to the other features; the partial sample with the maximum vector and Euclidean length is a clustering boundary sample; extracting features of the test set; calculating the minimum Euclidean distance from the test feature to the feature center and the corresponding category; and calculating the Euclidean length of the vector sum from the test feature to the training feature of the minimum distance category, comparing the Euclidean length with the Euclidean length corresponding to the clustering boundary sample, and performing open set recognition. The method not only can be used for identifying and classifying the known category of targets in the library, but also can reject the unknown category of targets outside the library, thereby improving the target identification accuracy and the automation and intelligence level of the radar.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and particularly relates to a method and device for open set recognition of high-resolution range profiles of radar based on clustering boundary detection. Background Art

[0002] The range cell length R of the radar transmitted signal can be expressed as where c is the speed of light and B is the bandwidth of the radar transmitted signal. The large bandwidth of the radar transmitted signal provides high range resolution. When the wavelength of the wideband radar is much smaller than the target size, the target can be approximately represented by a scattering point model. At this time, the part of the target in each range cell will be replaced by a scattering center in this range cell. The echoes of the scattering centers in different range cells together constitute the high-resolution range profile (HRRP, high resolution range profile) of the target. The HRRP of the target contains information such as the geometric structure of the target and is of great value for target recognition and classification.

[0003] In recent years, the target recognition methods for radar high-resolution range profile data mainly focus on closed set recognition, which requires the data categories in the test sample set to be consistent with those in the training sample set. However, in practical applications, the radar not only captures the high-resolution range profiles of the targets in the library, but also captures the high-resolution range profiles of many unknown category targets outside the library. In this case, the existing closed set recognition algorithms cannot reject the unknown category data outside the library, but will misclassify it as a certain category in the library, which greatly reduces the target recognition accuracy of the radar.

[0004] Therefore, some researchers have started to study the open set recognition of radar high-resolution range profiles. 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 multimodal distribution of HRRP data in the high-dimensional feature space, thereby improving the recognition and rejection performance of radar HRRP. Zhang Xuefeng et al. proposed a multi-classifier fusion algorithm based on the maximum correlation classifier (Maximum correlation classifier, MCC), support vector machine (Support vector machine, SVM), and relevance vector machine (Relevance vector machine, RVM) to achieve the rejection and recognition functions of radar HRRP. However, both of the above two algorithms need to rely on specific forms of kernel functions to extract features, which limits the ability of the model to extract sufficiently separable features, thus affecting the target recognition accuracy and the intelligent level of the radar.

[0005] To overcome the performance disadvantages of feature extraction in 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 after Softmax normalization, reject targets outside the database by dividing the sample classification probability threshold. However, such loss functions make the sample feature distribution not compact enough, thus restricting the open-set performance of the model. Therefore, some works have proposed an open-set recognition model based on prototype learning, which can improve the open-set performance by making the intra-class features compact and expanding the inter-class differences. However, the way of dividing the decision boundary centered on the prototype points of each category in this method is not reasonable, which also limits the open-set performance of such models. Summary of the Invention

[0006] In order to solve the problem that the open-set performance of the model in the prior art is restricted, thus affecting the accuracy of target recognition and the intelligent level of the radar, the present invention provides a method and device for open-set recognition of radar high-resolution range profiles based on clustering boundary detection.

[0007] In order to achieve the above object, the technical solution provided by the present invention is as follows:

[0008] The first aspect of the present invention provides a method for open-set recognition of radar high-resolution range profiles based on clustering boundary detection, including the following steps:

[0009] Step 1: Establish a sample set:

[0010] It includes a first training sample set and a first test sample set; wherein, the training sample set includes radar high-resolution range profiles of several known-category targets, and the test sample set includes radar high-resolution range profiles of several 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: Use the second training sample set to train the convolutional neural network to obtain a trained convolutional neural network;

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

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

[0016] Step 7: Sort the Euclidean lengths of the vectors of the N-category samples, and select several category samples with the largest numerical values as the detected clustering boundary samples;

[0017] Step 8: Use the trained convolutional neural network to extract the high-dimensional features of the second test sample set;

[0018] Step 9: Calculate the Euclidean distances from the high-dimensional features of the second test samples to the N feature centers of the second training sample set in Step 6, and determine the category represented by the minimum value among them;

[0019] Step 10: Use the clustering boundary detection result to perform open-set recognition on the high-dimensional features of the second test sample set to obtain the recognition result of the second test sample set.

[0020] Further, in the above Step 3, the convolutional neural network includes: three convolutional layers and one fully connected layer; the convolutional strides of the three convolutional layers are the same, each convolutional layer includes a number of convolutional kernels, and the size of each convolutional kernel is the same; the expression of the loss function of the convolutional neural network is:

[0021]

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

[0023] Further, the training process of the above Step 4 includes:

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

[0025] Input the sample data of each batch into the convolutional neural network for processing in turn to obtain the output result of the convolutional neural network;

[0026] Calculate 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 update the parameter values of the convolutional neural network by using the stochastic gradient method until the network converges to obtain the trained convolutional neural network.

[0027] Further, the preprocessing in the above step 1 includes: successively performing centroid alignment and normalization on the radar high-resolution range profiles to obtain a preprocessed second training sample set and a second test sample set.

[0028] Further, in the above steps 5 and 8, the process of using the trained convolutional neural network to extract high-dimensional features includes:

[0029] Performing convolution and downsampling on the input data of the current batch of the sample set using the first convolutional layer to obtain a first feature map;

[0030] Performing convolution and downsampling on the first feature map using the second convolutional layer to obtain a second feature map;

[0031] Performing convolution and downsampling on the second feature map using the third convolutional layer to obtain a third feature map;

[0032] Performing a non-linear transformation on the third feature map using the fourth fully connected layer 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 the above step 6, the expression of the feature center of the high-dimensional features of the second training sample set is:

[0035]

[0036] where represents the i-th sample in the k-th category in the second training sample set, and there are a total of N k samples in the k-th category;

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

[0038]

[0039] Its Euclidean length is expressed as:

[0040]

[0041] Further, in the above step 7, the number of clustering boundary samples selected is:

[0042] m k = floor(βN k )

[0043] where m kdenotes the number of samples detected as clustering boundary samples in the k-th category, and β is a hyperparameter representing the proportion of clustering boundary samples in the samples N k proportion.

[0044] Furthermore, in the above step nine, the trained convolutional neural network is used to extract the high-dimensional features of the second test sample set, and the Euclidean distance from the high-dimensional features of the second test sample to the feature center of the training sample set is calculated to determine the category represented by the minimum value:

[0045]

[0046] Furthermore, in the above step ten, the open-set recognition results include:

[0047] When holds, the test samples corresponding to the high-dimensional features of the second test sample set are unknown categories outside the database;

[0048] When does not hold, the category of the test samples corresponding to the high-dimensional features of the second test sample set is

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

[0050] The second aspect of the present invention provides a radar high-resolution range profile open-set recognition device based on clustering boundary detection, including:

[0051] A data acquisition module for establishing a first training sample set and a first test sample set; wherein, the training sample set includes radar high-resolution range profiles of targets of several known categories, and the test sample set includes radar high-resolution range profiles of targets of several known categories and radar high-resolution range profiles of targets of unknown categories outside the database;

[0052] A preprocessing module for preprocessing 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;

[0053] A model construction module for constructing a convolutional neural network;

[0054] A training module for training the convolutional neural network with the second training sample set to obtain a trained convolutional neural network;

[0055] A first extraction module for extracting the high-dimensional features of the second training sample set by using the trained convolutional neural network;

[0056] ​The clustering boundary detection module is used to calculate the feature center of the high-dimensional features of the second training sample set, as well as the Euclidean length of the vector sum of the high-dimensional features of each sample in the same category to the high-dimensional features of the remaining samples. By sorting these Euclidean lengths, a small part of the samples with the largest numerical values are the detected clustering boundary samples;

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

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

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] 1. The radar high-resolution range profile open-set recognition method provided by the present invention uses a one-dimensional convolutional neural network to extract high-dimensional features of data. First, preprocessing such as energy normalization and alignment is performed on the data, which can mine high-level features of high-resolution range profile data, remove the amplitude sensitivity, translation sensitivity, and pose sensitivity of radar high-resolution range profile data, and has stronger robustness compared with traditional direct classification methods; second, it can combine the primary features of each layer to obtain higher-level features for recognition. Compared with traditional machine learning feature extraction methods, its network structure and end-to-end non-linear mapping method both contribute to improving the target recognition rate of the model.

[0061] 2. The present invention uses a loss function based on prototype learning to train the one-dimensional convolutional neural network. Compared with the conventional cross-entropy loss function, it can further improve the compactness of known category features in the feature space and expand the distribution difference between categories, thereby further improving the target recognition rate.

[0062] 3. The present invention calculates the Euclidean distance from the features of the sample to be measured to the features of all samples in the known category clustering, so as to globally detect the boundary samples of the known category feature clustering. The present invention can detect the boundary samples of the known category feature clustering point by point and determine the clustering boundary to specific sample points; on this basis, comparing the features of the test sample with the features of the detected clustering boundary samples, the specific category of the target to be measured can be judged. Compared with the previous methods based on distance measurement, this strategy can effectively improve the recognition rate of samples at the edge of the known category cluster and has strong recognition robustness.

[0063] 4. Starting from the basic principles of pattern recognition, the present invention proposes innovative technical points that are implemented in combination in two major aspects: feature extraction and decision boundary. Specifically, in the design of the feature extraction aspect, the present invention uses a loss function based on prototype learning to train a one-dimensional convolutional neural network, which helps to extract more separable high-dimensional features of the original data. In addition, in the design of the decision boundary aspect, the present invention uses an original clustering boundary detection algorithm to accurately detect the feature clustering boundary, thereby improving the classification accuracy of test samples.

[0064] 5. Wide range of applications: It can not only be used to identify and classify known-category targets in the library, but also reject unknown-category targets outside the library, improving the target recognition accuracy, and thus effectively enhancing the automation and intelligence level of the radar. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 is a schematic flow chart of a method for open-set recognition of high-resolution range profiles of a radar based on clustering boundary detection provided by an embodiment of the present invention;

[0066] Figure 2 is a schematic diagram of an apparatus for implementing the method of the present invention;

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

[0068] The following further describes the present invention in detail with reference to specific embodiments and the accompanying drawings, but the implementation manners of the present invention are not limited thereto.

[0069] Embodiment 1, as Figure 1 shown, a method for open-set recognition of high-resolution range profiles of a radar based on clustering boundary detection provided by the present invention includes the following steps:

[0070] Step 1. Establish a sample set:

[0071] It includes a first training sample set and a first test sample set. Among them, the training sample set includes radar high-resolution range profiles of several known-category targets, and the test sample set includes radar high-resolution range profiles of several known-category targets and radar high-resolution range profiles of unknown-category targets outside the database.

[0072] 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;

[0073] Step 3. Construct a convolutional neural network;

[0074] Step 4. Use the second training sample set to train the convolutional neural network to obtain a trained convolutional neural network;

[0075] Step 5: Use the trained convolutional neural network to extract the high-dimensional features of the second training sample set;

[0076] Step 6: Calculate the feature centers of the high-dimensional features of N samples of the same category in the second training sample set, the sum of the vectors from the high-dimensional features of each sample in the N samples of the same category to the high-dimensional features of the remaining samples, and find the Euclidean length of the sum of the N category vectors;

[0077] Step 7: Sort the Euclidean lengths of the sum of the vectors of the N category samples, and select several category samples with the largest numerical values as the detected clustering boundary samples;

[0078] Step 8: Use the trained convolutional neural network to extract the high-dimensional features of the second test sample set;

[0079] Step 9: Calculate the Euclidean distances from the high-dimensional features of the second test sample to the N feature centers of the second training sample set in Step 6, and determine the category represented by the minimum value;

[0080] Step 10: Calculate the Euclidean length of the sum of all vectors from 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 9, compare it with the Euclidean lengths of the sum of the vectors of several clustering boundary samples in Step 7, perform open-set recognition, and obtain the recognition result of the second test sample set.

[0081] Example 2: An open-set recognition method for radar high-resolution range profiles based on clustering boundary detection, specifically including the following steps:

[0082] Step 1, Establish a sample set:

[0083] Step 1-1, Obtain P original radar high-resolution range profile data of N categories as the first training sample set, where N≥3 and P≥900;

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

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

[0086] Step 2: Preprocess the radar high-resolution range image sample data in the first training sample set and the first test sample set. In this embodiment, the preprocessing is to perform centroid alignment and normalization on the original data in the first training sample set and the first test sample set in sequence to obtain the preprocessed training sample set and test sample set, that is, the second training sample set and the second test sample set are obtained after preprocessing.

[0087] Specifically, denote the original data in the first training sample set or the first test sample set as x 0 , first, perform centroid alignment on the original data x 0 to obtain the data x' after centroid alignment 0 ; then, perform two-norm normalization on the data x' after centroid alignment 0 to obtain the sample data x after normalization processing, and its expression is:

[0088]

[0089] where 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 P×D and (Q + R)×D-dimensional matrices respectively, where D represents the total number of range cells included in one radar high-resolution range image original data.

[0090] Step 3: Construct a convolutional neural network:

[0091] In this embodiment, the set convolutional neural network includes three convolutional layers and one fully connected layer, denoted as the first convolutional layer, the second convolutional layer, the third convolutional layer, and the fourth fully connected layer respectively. Among them, each convolutional layer has the same convolutional stride, and each convolutional layer includes a number of convolutional kernels, and the size of each convolutional kernel is the same.

[0092] Specifically, for the first convolutional layer:

[0093] It is set to include C convolutional kernels, and the C convolutional kernels of the first convolutional layer are denoted as K, and the size of K is set to 1×w×1, where w represents the window of each convolutional kernel in the first convolutional layer, 1 < w < D; C is a positive integer greater than 0; the convolutional stride of the first convolutional layer is set to L; at the same time, the kernel window size of the downsampling process of the first convolutional layer is set to m×m, 1 < m < D, where D represents the total number of range cells included in one radar high-resolution range imaging data in the second training sample set, and m is a positive integer greater than 0; the stride of the downsampling process of the first convolutional layer is set to I, and the value of I is equal to that of m.

[0094] Set the activation function of the first convolutional layer to be x represents the preprocessed 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-ones bias of the third convolutional layer.

[0095] For the second convolutional layer:

[0096] It is set to include C' convolutional kernels. Denote the C' convolutional kernels of the second convolutional layer as K', and their size takes the same value as that of the convolutional kernels K of the first convolutional layer; the convolution stride of the second convolutional layer is denoted as L', where w ≤ L' ≤ D - w, and L' takes the same value as the convolution stride L of the first convolutional layer; at the same time, set the kernel window size of the downsampling process of the second convolutional layer to be m'×m', where 1 < m' < D and m' is a positive integer greater than 0; the stride of the second downsampling process is all I', and I' takes the same value as m'.

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

[0098] For the third convolutional layer:

[0099] It is set to include C″ convolutional kernels. Let the C″ convolutional kernels of the third convolutional layer be K″, and their size takes the same value as the size of each convolutional kernel window of the second convolutional layer; set the convolution stride of the third convolutional layer to be L″, which takes the same value as the convolution stride L' of the second convolutional layer; at the same time, set the kernel window size of the downsampling process of the third convolutional layer to be m"×m", where 1 < m" < D and m" is a positive integer greater than 0; the stride of the third downsampling process is all I″, and I″ takes the same value as m″.

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

[0101] For the fourth fully connected layer:

[0102] Set its randomly initialized weight matrix to be a B×U-dimensional matrix. floor() represents rounding down, D represents the total number of range cells included in a radar high-resolution range imaging data in the second training sample, B ≥ D, and B is a positive integer greater than 0; set the activation function to be represents the third feature map output by the third convolutional layer, represents the all-ones bias of the fourth fully-connected layer, and is of dimension U×1.

[0103] After constructing the model of the convolutional neural network, it further includes constructing the loss function of the convolutional neural network, and its expression is:

[0104]

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

[0106] Step Four, use the second training sample set to train the convolutional neural network to obtain the trained convolutional neural network. The specific steps include:

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

[0108] Step 4-2, sequentially input the sample data of each batch into the convolutional neural network for processing to obtain the output results of the convolutional neural network. Specifically, it includes the following 5 steps:

[0109] Step 4-2-1, after inputting the sample data into the convolutional neural network, use the first convolutional layer to perform convolution and downsampling on the currently input sample data to obtain the first feature map.

[0110] Specifically, use the convolution stride L of the first convolutional layer to perform convolution on the input sample data x with the C convolutional kernels of the first convolutional layer respectively to obtain the C convolution results of the first convolutional layer, and denote them as the C feature maps y of the first convolutional layer:

[0111]

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

[0113] For each feature map in perform downsampling processing respectively to obtain C feature maps after the downsampling processing of the first convolutional layer That is, the first feature map, denoted as:

[0114]

[0115] where represents taking the maximum value of the C feature maps of the first convolutional layer after Gaussian normalization within the kernel window size m×m of the first downsampling processing and represents 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 processing on the first feature map to obtain the second feature map.

[0117] Specifically, use the convolution stride L′ of the second convolutional layer to convolve the C feature maps (that is, the first feature map) after the downsampling processing of the first convolutional layer with the C′ convolutional kernels K′ of the second convolutional layer respectively to obtain the C′ convolution results of the second convolutional layer, and denote them as the C′ feature maps of the second convolutional layer

[0118] Perform Gaussian normalization processing on the C′ feature maps of the second convolutional layer to obtain the C′ feature maps of the second convolutional layer after Gaussian normalization

[0119] For each feature map in perform downsampling processing respectively, and further obtain the C′ feature maps after the downsampling processing of the second convolutional layer That is, the second feature map, denoted as:

[0120]

[0121] where represents taking the maximum value of the C′ feature maps of the second convolutional layer after Gaussian normalization within the kernel window size m′×m′ of the second downsampling processing and represents the C′ 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 processing on the second feature map to obtain the third feature map.

[0123] Specifically, use the convolution stride \(L'\) of the third convolutional layer to perform convolution on the \(C'\) feature maps after downsampling by the second convolutional layer (i.e., the second feature map) and the \(C''\) convolutional kernels \(K''\) of the third convolutional layer respectively, to obtain \(C''\) convolution results of the third convolutional layer, and denote them as the \(C''\) feature maps of the third convolutional layer

[0124] Perform Gaussian normalization on the \(C''\) feature maps of the third convolutional layer to obtain the \(C''\) feature maps of the third convolutional layer after Gaussian normalization

[0125] For each feature map in, perform downsampling respectively, and then obtain the \(C''\) feature maps after downsampling by the third convolutional layer i.e., the third feature map, expressed as:

[0126]

[0127] where represents taking the maximum value of the \(C''\) feature maps of the third convolutional layer after Gaussian normalization within the kernel window size \(m''\times m''\) of the third downsampling process and represents the \(C''\) feature maps of the third convolutional layer after Gaussian normalization.

[0128] Step 4-2-4, use the fourth fully connected layer to perform non-linear transformation on the third feature map to obtain the processing result of the current sample data Its expression is:

[0129]

[0130] where represents the randomly initialized weight matrix of the fourth fully connected layer, and

[0131] represents the all-ones bias of the fourth fully connected layer.

[0132] Step 4-2-5, repeat the above steps 4-2-1 to 4-2-4 until the processing of all batches of input data is completed, to obtain the output result of the convolutional neural network, and this output result is high-dimensional features.

[0132]

[0133] Step 4-3, calculate 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 proposed in step three, and use the stochastic gradient method to update the parameter values of the convolutional neural network until the network converges, to obtain the trained convolutional neural network.

[0133] Specifically, substitute the output result obtained in the above step 4-2, i.e., , into the loss function expression of the above convolutional neural network to obtain the value of the loss function, and use the existing stochastic gradient method to update the parameter values of the convolutional neural network until the network converges, thus obtaining the trained convolutional neural network. The stochastic gradient method here belongs to the mature existing technology and will not be specifically introduced in this embodiment. As for the stochastic gradient method here, it belongs to the mature existing technology and will not be specifically introduced in this embodiment.

[0134] In this embodiment, by adopting a multi-layer convolutional neural network structure and performing preprocessing of two-norm normalization and alignment on the data, the high-level features of the high-resolution range image data can be mined.

[0135] Step Five: Use the trained convolutional neural network to extract the high-dimensional features Θ(x train ) of the second training sample set x train .

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

[0137] Step Six: Calculate the feature center of the high-dimensional features of N samples of the same category in the second training sample set, as well as the sum of the vectors from the high-dimensional features of each sample in the N samples of the same category to the high-dimensional features of the remaining samples, and then calculate the Euclidean length of the sum of the N category vectors.

[0138] Specifically, the calculation expression for the feature center (where k represents the k-th category) of the high-dimensional features of the sample data of each category in the second training sample set is:

[0139]

[0140] where, represents the i-th sample in the k-th category in the second training sample set, and there are a total of N k samples in this category.

[0141] The sum of the vectors from the high-dimensional features 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] Its Euclidean length can be expressed as:

[0144]

[0145] Step Seven: After sorting the Euclidean lengths of the N category sample vectors , select a part of the samples with the largest numerical values as the detected clustering boundary samples:

[0146] The selection quantity is:

[0147] m k = floor(βN k )

[0148] where m k represents the number of samples detected as clustering boundary samples in the k-th category, and β is a hyperparameter representing the proportion of clustering boundary samples in the samples N k . In this embodiment, the boundary detection result for clustering the target features of known categories in the library is introduced in the decision-making stage. By calculating the Euclidean length of the vector sum of the features of the target to be measured to the known category features and comparing it with the detected boundary samples of the known category feature clustering, the specific category of the target to be measured is determined. Compared with the previous methods based on classification probability and distance metric, this strategy can effectively improve the recognition rate of samples at the clustering boundary of known categories and has strong recognition robustness.

[0149] Step eight, use the trained convolutional neural network to extract the high-dimensional features Θ(x test ) of the second test sample set x test :

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

[0151] Step nine, calculate the Euclidean distances from the high-dimensional features of the second test sample to the N feature centers of the second training sample set in step six, and determine the category represented by the minimum value among them:

[0152]

[0153] Step ten, use the clustering boundary detection result to perform open-set recognition on the high-dimensional features of the second test sample set to obtain the recognition result of the second test sample set.

[0154] Specifically, calculate the Euclidean length of the vector sum of all the high-dimensional features Θ(x test ) of the second test sample to the high-dimensional features of the second training sample of the known category in step nine and compare it with the Euclidean length of the vector sum of several clustering boundary samples in step seven for open-set recognition:

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

[0156] When does not hold, the category of the test sample corresponding to the high-dimensional features of the second test sample set is

[0157] where is the m-th in the k-th categoryk The Euclidean length of the vector sum of the highest-dimensional features of one largest sample to the remaining sample high-dimensional features.

[0158] Example 3, see Figure 2 The embodiment of the present invention also provides a device for an open-set recognition method of radar high-resolution range profiles based on clustering boundary detection, which is characterized by including:

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

[0160] A preprocessing module, 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] A model construction module, configured to construct a convolutional neural network;

[0162] A training module, configured to train the convolutional neural network using the second training sample set to obtain a trained convolutional neural network;

[0163] A first extraction module, configured to extract the high-dimensional features of the second training sample set using the trained convolutional neural network;

[0164] A clustering boundary detection module, configured to calculate the feature center of the high-dimensional features of the second training sample set, and the Euclidean length of the vector sum of the high-dimensional features of each sample in the same category to the remaining sample high-dimensional features. By sorting these Euclidean lengths, a part of the samples with the largest numerical values are the detected clustering boundary samples;

[0165] A second extraction module, configured to extract the high-dimensional features of the second test sample set using the trained convolutional neural network;

[0166] A target recognition module, configured to perform open-set recognition on the high-dimensional features of the second test sample set using the clustering boundary samples to obtain the recognition result of the second test sample set.

[0167] The radar high-resolution range profile open-set recognition device provided in this embodiment can implement the radar high-resolution range profile open-set recognition method provided in the above-mentioned Embodiment 1, and the detailed process will not be elaborated here.

[0168] The beneficial effects of the present invention will be verified and described below through simulation experiments.

[0169] 1. Simulation conditions

[0170] The hardware platform for the simulation experiment of this embodiment is as follows:

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

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

[0173] The data used in this simulation experiment is the measured high-resolution range images of 13 types of aircraft. The models of these 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 used as unknown categories to make training sample sets and test sample sets. Among them, the training sample set has a total of 21,000 samples, with approximately 3,000 samples for each in-library category; the test sample set includes 7,000 samples of the 7 known in-library categories and 6,000 samples of the 6 unknown out-of-library categories, with 1,000 samples for each category.

[0175] Before conducting the experiment, all the original data is preprocessed according to Step 2 in Embodiment 2 above, and then the open-set recognition experiment is carried out using the present invention.

[0176] 2. Simulation Content and Result Analysis

[0177] This simulation experiment compares the method of the present invention with some traditional two-stage rejection recognition methods, as well as the SoftMax threshold method and the GCPL prototype method.

[0178] Traditional two-stage rejection recognition methods mainly use methods such as SVDD, OCSVM, and I-Forest to reject out-of-library targets, and then use methods such as SVM to further classify the targets that have been judged as in-library. The SoftMax threshold method determines the category of the target to be measured according to the size of the final classification output of the convolutional neural network. If the final network classification output is greater than the defined threshold, it is judged as an in-library category; otherwise, it is judged as an out-of-library category. The GCPL prototype method determines the category of the target to be measured according to the distance between the feature of the sample to be measured and the nearest prototype center. If this distance is greater than the defined threshold, it is judged as an out-of-library category; otherwise, it is judged as an in-library category.

[0179] This simulation experiment uses the area AUC under the Receiver Operating Characteristic (ROC) curve to evaluate the rejection ability of different methods for out-of-library targets. The larger the value of AUC, the stronger the rejection ability for out-of-library targets.

[0180] See Figure 3 , it can be seen that in this simulation experiment, the present invention has the strongest rejection ability for out-of-library targets. Followed by the GCPL prototype method and the SoftMax threshold method, and the three traditional machine learning methods have general rejection abilities for out-of-library targets.

[0181] Since there are many data categories used in this simulation experiment, the Macro Average F1-Score is used to comprehensively evaluate the open-set recognition ability of different methods. The larger the value of F1-Score, the stronger the open-set recognition ability. The results of the simulation experiment are shown in the following table. It can be seen that in this simulation experiment, the present invention has the strongest comprehensive open-set recognition ability, which is significantly better than other 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 present invention 0.822 0.785

[0183] In summary, it can be seen that the present invention has achieved the best results both in terms of the rejection ability for out-of-library targets and the comprehensive consideration of the open-set recognition ability, which proves that the method of the present invention has made significant progress.

[0184] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A radar high-resolution range profile open set recognition method based on cluster boundary detection, characterized by: The following steps are involved: Step 1: Create a sample set: It includes a first training sample set and a first test sample set; wherein the training sample set includes radar high-resolution range images of several targets of known categories, and the test sample set includes radar high-resolution range images of several targets of known categories and radar high-resolution range images of targets of unknown categories outside the database; Step 2: 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 3: Construct a convolutional neural network; Step 4: training the convolutional neural network using the second training sample set to obtain a trained convolutional neural network; Step 5: Using the trained convolutional neural network to extract high-dimensional features of the second training sample set; 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 vectors from the high-dimensional features of each sample of the same category to the high-dimensional features of the remaining samples, and calculate the Euclidean length of the sum of the N category vectors; Step 7: Sort the Euclidean lengths of the vector sums of the N category samples, and select the category samples with the largest values ​​as the detected cluster boundary samples; Step 8: Using the trained convolutional neural network to extract high-dimensional features of the second test sample set; Step 9: Calculate the Euclidean distance between the high-dimensional feature of the second test sample and the N feature centers of the second training sample set in step 6, and determine the category represented by the minimum value; Step 10: Use the cluster boundary detection result to perform open set recognition on the high-dimensional features of the second test sample set to obtain a 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 3, the convolutional neural network includes: three convolutional layers and one fully connected layer; the convolutional steps of the three convolutional layers are the same, each convolutional layer includes a number of convolutional kernels, and the size of each convolutional kernel is the same; the loss function of the convolutional neural network is expressed as: Among them, Θ(x) and Θ(x k ) is the output result of the convolutional neural network, O i (i=1,…,k,…,N) is the i-th prototype randomly initialized according to Gaussian distribution; d(Θ(x),O k ) is Θ(x) to O k The Euclidean distance of , λ is a hyper parameter.

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 4 includes: The sample data of the second training sample set is randomly divided into q batches, and the data of each batch is 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 the sample data of each batch into the convolutional neural network in sequence for processing to obtain the output result of the convolutional neural network; 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 values ​​of the convolutional neural network are updated by using the stochastic gradient method until the network converges to obtain a trained convolutional neural network.

4. The radar high-resolution range profile open set recognition method based on cluster boundary detection according to claim 3 is characterized by: The preprocessing of step 1 includes: performing centroid alignment and normalization processing on the radar high-resolution range image in sequence to obtain a preprocessed second training sample set and a 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 step 5 and step 8, the process of extracting high-dimensional features using the trained convolutional neural network includes: The first convolution layer is used to perform convolution and down-sampling processing on the input sample set data of the current batch to obtain the first feature map; Using a second convolutional layer to perform convolution and down-sampling processing on the first feature map to obtain a second feature map; Using a third convolutional layer to perform convolution and down-sampling processing on the second feature map to obtain a third feature map; Using the fourth fully connected layer to perform nonlinear transformation processing on the third feature map to obtain an output result of the current sample set data; Repeat the above steps until multiple batches of sample data are processed and the output results of the convolutional neural network are 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 step 6, the expression of the feature center of the high-dimensional feature of the second training sample set is: in, represents the i-th sample in the k-th category in the second training sample set, and the k-th category has a total of N k samples; The vector sum of the high-dimensional features of the i-th sample in the k-th category to the high-dimensional features of the remaining samples is expressed as: Its European length is expressed as:

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

8. The radar high-resolution range profile open set recognition method based on cluster boundary detection according to claim 7, characterized in that: In step nine, the trained convolutional neural network is used to extract the high-dimensional features of the second test sample set, and the Euclidean distance from the high-dimensional features of the second test sample to the feature center of the training sample set is calculated 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 10, the open set recognition result includes: when When it is established, the test samples corresponding to the high-dimensional features of the second test sample set are unknown categories outside the database; when If it is not true, the category of the test sample corresponding to the high-dimensional feature of the second test sample set is in, is the mth in the kth category k The Euclidean length of the vector sum from the largest sample high-dimensional feature to the remaining sample high-dimensional features.

10. The radar high-resolution range profile open set recognition device based on cluster boundary detection according to claim 1, characterized in that: include: A data acquisition module, used 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 several targets of known categories, and the test sample set includes radar high-resolution range images of several targets of known categories and radar high-resolution range images of targets of unknown categories outside the database; A preprocessing module, used for preprocessing 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; Model building module, used to build convolutional neural networks; A training module, used for training the convolutional neural network using the second training sample set to obtain a trained convolutional neural network; A first extraction module, configured to extract high-dimensional features of the second training sample set using the trained convolutional neural network; A cluster boundary detection module is used to calculate the feature center of the high-dimensional features of the second training sample set and the Euclidean length of the vector sum of the high-dimensional features of each sample in the same category to the high-dimensional features of other samples, and by sorting these Euclidean lengths, a part of the samples with the largest values ​​are the detected cluster boundary samples; A second extraction module, used to extract high-dimensional features of the second test sample set using a trained convolutional neural network; The target recognition module is used to perform open set recognition on the high-dimensional features of the second test sample set using the cluster boundary samples to obtain a recognition result of the second test sample set.

Images