A High-Resolution Range Profile Open Set Recognition Method, Device, Equipment and Storage Medium

Abstract

The present application discloses a high-resolution range image open-set recognition method, device, equipment and storage medium, which is applied to the field of radar target recognition, and includes: obtaining the reconstruction error of a sample to be recognized on various category sub-dictionaries, and taking the category corresponding to the minimum reconstruction error as the candidate category; obtaining the mean value of the candidate category reconstruction error and the candidate category measure in the candidate category training stage; fitting the Weibull distribution to the candidate category measure to obtain the candidate category scale parameter and the candidate category shape parameter; taking the ratio of the minimum reconstruction error of the sample to be recognized to the mean value of the candidate category reconstruction error as the measure of the sample to be recognized; calculating the confidence level of the Weibull distribution according to the measure of the sample to be recognized, the candidate category shape parameter and the candidate category scale parameter; comparing the confidence level with a preset threshold to determine the category of the sample to be recognized. The present application uses a new measure and fits the Weibull distribution, without the need for manual setting of the reconstruction error tail and weight parameters; and the high-resolution range image open-set recognition performance of this method is also superior to traditional algorithms.

Description

Technical Field

[0001] The present application relates to the field of radar target recognition, and particularly to a high-resolution range profile open set recognition method, apparatus, device, and storage medium. Background Art

[0002] Traditional target recognition methods generally distinguish targets within a closed set and cannot effectively identify unknown categories outside the library. That is, the algorithm completes the determination of the target category under closed set conditions. When an out-of-library target sample is input, it will be judged as the target with the best match in the library, resulting in incorrect judgments. Research on open set recognition problems has been carried out in the field of computer vision, but there is less research in the field of radar target recognition.

[0003] Before the implementation of the present application, the application of common open set recognition algorithms in the field of computer vision, such as SROSR (SR, Sparse Representation, sparse representation; OSR, Open Set Recognition, radar target open set recognition), WSVM (Weibull-SVM, Weibull-corrected support vector machine), and 1-vs-Set (current class / 1 class vs. other classes) algorithms in the field of radar target recognition was studied. After comparison, it was found that the recognition performance of the SROSR algorithm is better than that of other algorithms. However, it was also found that this algorithm requires manual setting of tail and weight parameters, and it is difficult to obtain optimized parameters in practical applications.

[0004] Therefore, an open set recognition algorithm that does not require manual parameter setting while ensuring recognition accuracy is needed. Summary of the Invention

[0005] In view of this, the purpose of the present application is to provide a high-resolution range profile open set recognition method, apparatus, device, and storage medium, which solves the problems of inaccurate recognition of the closed set recognition method in an open environment and the need for manual parameter setting in the existing SROSR technology.

[0006] To solve the above technical problems, the present application provides a high-resolution range profile open set recognition method, including:

[0007] Obtain the reconstruction error of the sample to be recognized on each category sub-dictionary, where each category sub-dictionary is constructed using training samples of each category, and the sample to be recognized and the training samples of each category are high-resolution range profiles;

[0008] Take the category corresponding to the minimum value in the reconstruction errors as the candidate category, and take the minimum value as the reconstruction error of the sample to be recognized;

[0009] Obtain the candidate category reconstruction error mean and candidate category measure obtained during the training phase for the candidate category;

[0010] Fitting the Weber distribution to the candidate class measure to obtain the candidate class scale parameter and the candidate class shape parameter;

[0011] Calculating the ratio of the reconstruction error of the sample to be recognized to the mean value of the reconstruction errors of the candidate classes, and using the ratio as the measure of the sample to be recognized;

[0012] Calculating the confidence of the Weber distribution according to the measure of the sample to be recognized, the candidate class scale parameter, and the candidate class shape parameter;

[0013] Comparing the confidence with a preset threshold to determine the class of the sample to be recognized.

[0014] Optionally, obtaining the reconstruction error of the sample to be recognized on the sub-dictionaries of each class includes:

[0015] Obtaining the training samples of each class;

[0016] Using the training samples of each class to construct the sub-dictionaries of each class, and combining the sub-dictionaries of each class into an over-complete dictionary;

[0017] Calculating the first sparse coefficient corresponding to the sample to be recognized on the sub-dictionaries of each class by using a sparse algorithm c = 1, …, C, the is in the form of a column vector, and C represents the number of training sample classes; calculating the reconstruction error of the sample to be recognized on the sub-dictionaries of each class according to the first sparse coefficient, the sub-dictionaries of each class, and the sample to be recognized.

[0018] Optionally, taking the class corresponding to the minimum value in the reconstruction errors as the candidate class includes:

[0019] According to Determining the candidate class;

[0020] The y represents the sample to be recognized and is in vector form; the D represents the over-complete dictionary and is in matrix form; the represents the candidate class; represents the first sparse coefficients calculated for the sample to be recognized on the sub-dictionaries of each class and combined end to end to form a second sparse coefficient, which is in the form of a column vector; represents retaining the coefficient elements corresponding to the dictionary atoms of the c-th class in

[0021] Optionally, obtaining the mean value of the reconstruction errors of the candidate class obtained during the training phase includes:

[0022] Sparsely reconstruct each training sample in each training category using the sparse algorithm to obtain the reconstruction error of the training sample;

[0023] Calculate the mean reconstruction error of the training samples of each category based on the training sample category and the reconstruction error of the training sample;

[0024] Select the candidate class reconstruction error mean corresponding to the candidate class from the mean reconstruction errors of the training samples of each category.

[0025] Optionally, obtaining the candidate class measure obtained by the candidate class in the training stage includes:

[0026] Calculate the ratio of the reconstruction error of each category of training samples to the mean reconstruction error of the training samples of each category, and form a measure matrix;

[0027] Select the candidate class measure corresponding to the candidate class from the measure matrix.

[0028] Optionally, the sample to be identified and the training samples of each category are high-resolution range images, including:

[0029] Obtain the initial training sample data and the initial sample to be tested for each category of the high-resolution range image;

[0030] Perform incoherent averaging, 2-norm amplitude normalization, and power transformation on the initial sample data of each category and the initial sample to be tested within a preset angle range to obtain the training samples of each category and the sample to be identified respectively.

[0031] The present application also provides a high-resolution range image open-set recognition device, including:

[0032] A first acquisition module, configured to acquire the reconstruction error of the sample to be identified on each category of sub-dictionaries, where each category of sub-dictionaries is the sub-dictionaries constructed using the training samples of each category, and the sample to be identified and the training samples of each category are high-resolution range images;

[0033] A candidate class determination module, configured to use the category corresponding to the minimum value in the reconstruction error as the candidate class, and use the minimum value as the reconstruction error of the sample to be identified;

[0034] A second acquisition module, configured to acquire the candidate class reconstruction error mean and the candidate class measure obtained by the candidate class in the training stage;

[0035] A fitting module, configured to fit the Weber distribution to the candidate class measure to obtain the candidate class scale parameter and the candidate class shape parameter;

[0036] A measure calculation module for samples to be recognized, which is used to calculate the ratio of the reconstruction error of the sample to be recognized to the average reconstruction error of the candidate classes, and take the ratio as the measure of the sample to be recognized;

[0037] A confidence calculation module, which is used to calculate the confidence of the Weibull distribution according to the measure of the sample to be recognized, the scale parameter of the candidate class, and the shape parameter of the candidate class;

[0038] A class determination module, which is used to compare the confidence with a preset threshold to determine the class of the sample to be recognized.

[0039] Optionally, the first acquisition module includes:

[0040] An acquisition unit, which is used to acquire the training samples of each class;

[0041] A construction combination unit, which is used to construct the sub-dictionaries of each class by using the training samples of each class, and combine the sub-dictionaries of each class into an over-complete dictionary;

[0042] A sparse coefficient calculation unit, which is used to calculate the first sparse coefficient corresponding to the sample to be recognized on the sub-dictionaries of each class by using a sparse algorithm c = 1, …, C, the is in the form of a column vector, and C represents the number of training sample classes;

[0043] A reconstruction error calculation unit for each class, which is used to calculate the reconstruction error of the sample to be recognized on the sub-dictionaries of each class according to the first sparse coefficient, the sub-dictionaries of each class, and the sample to be recognized.

[0044] The present application also provides a high-resolution range profile open-set recognition device, including:

[0045] A memory, which is used to store a computer program;

[0046] A processor, which is used to implement the steps of the above high-resolution range profile open-set recognition method when executing the computer program.

[0047] The present application also provides a storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above high-resolution range profile open-set recognition method are implemented.

[0048] It can be seen that in this application, the reconstruction errors of the sample to be recognized on various category sub-dictionaries are obtained. The various category sub-dictionaries are the various category sub-dictionaries constructed using the training samples of various categories. The sample to be recognized and the training samples of various categories are high-resolution range images. The category corresponding to the minimum value in the reconstruction errors is used as the candidate category, and the minimum value is used as the reconstruction error of the sample to be recognized. The candidate category reconstruction error mean and the candidate category measure obtained during the training stage for the candidate category are acquired. The Weber distribution is fitted to the candidate category measure to obtain the candidate category scale parameter and the candidate category shape parameter. The ratio of the reconstruction error of the sample to be recognized to the candidate category reconstruction error mean is calculated, and the ratio is used as the measure of the sample to be recognized. The confidence level of the Weber distribution is calculated based on the measure of the sample to be recognized, the candidate category scale parameter, and the candidate category shape parameter. The confidence level is compared with a preset threshold to determine the category of the sample to be recognized. By using the reconstruction error mean ratio as a new measure and fitting the Weber distribution, this application eliminates the need for manual setting of the reconstruction error tail and weight parameters. Moreover, the performance of this method in open-set recognition of high-resolution range images is also superior to traditional algorithms.

[0049] In addition, this application also provides a high-resolution range image open-set recognition device, equipment, and storage medium, which also have the above-mentioned beneficial effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0051] Figure 1 It is a flowchart of a high-resolution range image open-set recognition method provided by an embodiment of this application;

[0052] Figure 2 It is a graph of the accuracy rate results of the open-set recognition algorithm under different opening degrees provided by an embodiment of this application;

[0053] Figure 3 It is a graph of the F-value results of the open-set recognition algorithm under different opening degrees provided by an embodiment of this application;

[0054] Figure 4 It is a schematic structural diagram of a high-resolution range image open-set recognition device provided by an embodiment of this application;

[0055] Figure 5 It is a schematic structural diagram of a high-resolution range image open-set recognition device provided by an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application more clear, the following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of them. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0057] Traditional target recognition methods generally complete the determination of target categories under closed-set conditions. When out-of-library target samples are input, they will be judged as the target category with the best match in the library, resulting in judgment errors. In practical applications, radar automatic target recognition systems often face non-cooperative targets, and the target types are complex and diverse. It is impossible to establish a complete target recognition library during the training stage, and the target recognition method under closed-set conditions has difficulty meeting actual requirements. Therefore, the research on open-set recognition of radar targets is of great significance. How to enhance the rejection ability of the algorithm for out-of-library targets while ensuring the classification accuracy of in-library targets is one of the key problems that radar automatic target recognition algorithms need to solve.

[0058] Open-set recognition was first proposed in the field of computer vision and has produced relatively rich research results. For example: ① Recognition methods based on sparse representation, such as the SROSR algorithm, which uses the extreme value theorem to fit the tails of the reconstruction errors of the matching and non-matching classes to identify and reject targets; ② Recognition methods based on support vector machines. For example, the 1-vs-SetMachine algorithm optimizes the risks of experience and open space by generalizing or specializing two planes; the Weibull-corrected support vector machine algorithm corrects scores by combining the extreme value theory and support vector machines; the PI-SVM algorithm models the decision boundary through the extreme value theory for open-set recognition; ③ Recognition methods based on marginal distributions. For example, the Extreme Value Machine algorithm fits the marginal distribution characteristics of targets through the extreme value theory for open-set recognition; ④ Recognition methods based on distance metrics, such as the boundary detection algorithm based on K-nearest neighbor distribution; ⑤ Recognition methods based on deep neural networks. For example, replacing the SoftMax classifier in the deep neural network with the OpenMax classifier, such as the deep open classifier, C2AE (class-conditional autoencoder) algorithm, and generative adversarial network; ⑥ Recognition methods based on multi-algorithm fusion. For example, the combined algorithm of support vector machines and convolutional neural networks, the combined algorithm of the extreme value theory and random forest methods, the combined algorithm of threshold analysis and reciprocal point learning, the combined algorithm of convolutional neural networks and random forest, etc.

[0059] Some scholars have applied the open-set recognition method in the field of computer vision to radar target recognition, and the results show that the open-set recognition method has better classification accuracy and rejection performance compared with traditional rejection classification methods.

[0060] Drawing on the research results of open-set recognition in the field of computer vision, this embodiment of the application compares the applications of the SROSR, WSVM, and 1-vs-Set algorithms in radar target recognition. The results show that the SROSR algorithm has better recognition performance. However, SROSR requires some parameters to be set manually. Specifically, under the open-set condition, SROSR calculates the sum of the matching-class reconstruction error and the non-matching-class reconstruction error of the training samples of each category, fits the generalized Pareto distribution to the tail of the reconstruction error according to the Extreme Value Theory (EVT), calculates the confidence level fitted by the reconstruction error of the sample to be recognized, and makes a category decision based on the threshold. According to the extreme value parameters obtained by the SROSR algorithm in the training stage The confidence levels corresponding to the respective tail distributions are obtained as the probability scores score for matching m and the probability scores score for non-matching nm . Among them is the scale parameter for the matching class, is the shape parameter for the matching class, is the scale parameter for the non-matching class, is the shape parameter for the non-matching class. The two probability scores respectively represent the probability scores that the target to be judged belongs to the candidate class and does not belong to the non-candidate classes in the training classes, and are used to comprehensively evaluate the probability that the target belongs to the candidate class. By calculating the weighted sum of the two probability scores, the final discrimination score is obtained as shown in the following formula:

[0061] score = score m + w·score nm ;

[0062] where w is the weight parameter. The non-matching class score score nm is affected by the openness. The weight w is negatively correlated with it, that is, it is positively correlated with the closed-set degree 1 - openness

[0063] The openness is used to describe the difference between the test stage and the training stage categories, and is defined as:

[0064]

[0065] where N TR represents the number of training classes, N TE represents the number of test classes, and N TG represents the number of target classes

[0066] In view of the problems that the SROSR algorithm requires manual setting of tail parameters, matching class weights, non-matching class weights, etc., this application proposes an improved sparse representation open set recognition algorithm based on the mean ratio of reconstruction errors, and uses this algorithm for high-resolution range image open set recognition. Before formally introducing the content of the invention of this application, some theorems used in this application are introduced first:

[0067] In practice, it is often difficult to accurately describe data with known distributions. For some extreme events, the probability model assignment is usually 0. The extreme value theory can infer the distribution of extreme events that may be observed under the condition that the original data distribution is unknown.

[0068] Theorem 1: Fisher-Tippett Theorem.

[0069] Suppose X n,n is the maximum value of the sequence of independent and identically distributed random variables {X1,..., X n}. If there exists a sequence of constants a n > 0, such that the asymptotic distribution of (X n,n - b n ) / a n is non-degenerate, then G must belong to the Gumbel type (extreme value type I), Frechet-Pareto type (extreme value type II), or Reversed Weibull type (extreme value type III) distribution.

[0070]

[0071] When x is bounded, G follows the Reversed Weibull distribution. When applied to target recognition, for the bounded output value of the recognition function, it follows the Reversed Weibull distribution when there is an upper bound, and the Weibull distribution when there is a corresponding lower bound.

[0072] The above theorem can be represented by the General Extreme Value distribution (GEV):

[0073]

[0074] where γ represents the shape parameter, and its value depends on the distribution of the original data. γ < 0 corresponds to the Weibull type distribution, and γ can be estimated by the Pickands-Balkema-de Haan theorem.

[0075] Theorem 2: Pickands-Balkema-de Haan Theorem.

[0076] For a sufficiently large threshold \(t\), the distribution exceeding the threshold \(t\) is approximately a Generalized Pareto Distribution (GPD):

[0077]

[0078] where \(\sigma\) and \(\gamma\) correspond to the scale parameter and the shape parameter respectively. Denote the excess function \(Y\) j \(= X\) j \(- t\), where \(X\) j \(> t\), \(j = 1,2,\cdots,N\) t represents the subscript corresponding to the \(j\)-th excess value in the sequence \(\{X_1,\cdots,X\) n \(\}\), \(I(\cdot)\) is the indicator function, and the log-likelihood function corresponding to this distribution is:

[0079]

[0080] where, \(i = 1,\cdots,N\) t . Based on the likelihood function of the above formula, \(\) and \(\) can be estimated and

[0081] Considering that the mean of the reconstruction errors of the same category can characterize the overall characteristics of this category to a certain extent, this application can use the mean of the reconstruction errors as the clustering center of this category. Based on this, in the embodiments of this application, the ratio of the reconstruction error to the mean is used as a measure to replace the reconstruction error tail in SROSR for distribution fitting, avoiding the problem of selecting the tail parameters, matching class, and non-matching class score weights. For details, please refer to Figure 1 , Figure 1 is a flowchart of a high-resolution range profile open-set recognition method provided by the embodiments of this application. The method may include:

[0082] S101: Obtain the reconstruction errors of the sample to be recognized on the sub-dictionaries of each category, where the sub-dictionaries of each category are sub-dictionaries of each category constructed using the training samples of each category, and the sample to be recognized and the training samples of each category are high-resolution range profiles.

[0083] In this embodiment, when the number of categories of the training samples is \(C\), \(C\) reconstruction errors corresponding to the sample to be recognized on the sub-dictionaries of each category are obtained. This embodiment does not limit the calculation of the reconstruction errors of the sample to be recognized. As long as \(C\) reconstruction errors can be calculated based on the sample to be recognized. For example, the compressive sensing method or the sparse algorithm.

[0084] The calculation of the reconstruction errors in this embodiment may specifically include the following steps:

[0085] Step 21: Obtain the training samples of each category;

[0086] Step 22: Construct sub-dictionaries for each category using various category training samples, and combine the sub-dictionaries of each category into an over-complete dictionary;

[0087] Step 23: Use the sparse algorithm to calculate the first sparse coefficients corresponding to the sample to be recognized on the sub-dictionaries of each category c = 1,..., C, in the form of column vectors, where C represents the number of categories of training samples;

[0088] Step 24: Calculate the reconstruction error of the sample to be recognized on the sub-dictionaries of each category according to the first sparse coefficients, the sub-dictionaries of each category, and the sample to be recognized.

[0089] The over-complete dictionary D in this embodiment is constructed according to the training samples. This embodiment does not limit the specific construction method of the over-complete dictionary, as long as the dictionary is over-complete. This embodiment uses the sparse algorithm to obtain the sparse coefficients of the sample to be recognized, that is, the first sparse coefficients and according to the over-complete dictionary D, the first sparse coefficients combined into column vectors calculate the reconstruction errors of C samples to be recognized. The sparse representation algorithm (SR) can obtain the sparse reconstruction of the signal, and the reconstruction errors for different target categories often vary greatly. Generally speaking, the sparse reconstruction of the same category can obtain a smaller reconstruction error than the reconstruction of different categories. Therefore, the reconstruction error can be used to measure the distance relationship between the sample to be recognized and the C target categories.

[0090] The process of obtaining the reconstruction error includes: using the sparse algorithm to obtain the sparse coefficients of the sample, reconstructing according to the sparse coefficients and the over-complete dictionary to obtain the reconstructed data, and comparing the reconstructed data with the original sample data to obtain the reconstruction error.

[0091] S102: Take the category corresponding to the minimum value in the reconstruction errors as the candidate category, and take the minimum value as the reconstruction error of the sample to be recognized.

[0092] In this embodiment, the minimum reconstruction error value among the C reconstruction errors corresponding to the sample to be recognized is taken as the reconstruction error of the sample to be recognized, and the category corresponding to the minimum reconstruction error value is taken as the candidate category of the sample to be recognized.

[0093] When determining the candidate category of the sample to be recognized as described above, it may specifically include the following steps:

[0094] According to determine the candidate category; y represents the sample to be recognized, in vector form; D represents the over-complete dictionary, in matrix form; represents the candidate category; Denote the first sparse coefficients calculated for the sample to be recognized on each category sub-dictionary and concatenate them head-to-tail to form the second sparse coefficient, which is in the form of a column vector; Denote to retain the coefficient elements of the dictionary atoms corresponding to the c-th category in

[0095] S103: Obtain the candidate class reconstruction error mean and candidate class measure obtained during the training phase for the candidate class.

[0096] When training the training samples in this embodiment, the reconstruction errors of the training samples of all categories are calculated, the reconstruction error means of each category are calculated from the reconstruction errors, and the ratio of the reconstruction error of each training sample to the mean is used as the measure corresponding to the training samples of each category. The candidate class reconstruction error mean is selected from the reconstruction error means of each category, and the candidate class measure is selected from the measures corresponding to each category.

[0097] The above-mentioned obtaining the candidate class reconstruction error mean calculated during the training phase for the candidate class may specifically include the following steps:

[0098] Step 41: Use the sparse algorithm to perform sparse reconstruction on each training sample in each training category to obtain the training sample reconstruction error;

[0099] Step 42: Calculate the reconstruction error mean of the training samples of each category according to the training sample category and the training sample reconstruction error;

[0100] Step 43: Select the candidate class reconstruction error mean corresponding to the candidate class from the reconstruction error means of the training samples of each category.

[0101] The training samples in this embodiment and the label set L; Denote the M-dimensional real number field; the number of training sample categories is C; the over-complete dictionary D c Denote the sub-dictionary corresponding to the c-th category; Denote the total number of dictionary atoms, N c is the number of sub-dictionary atoms in the c-th category, Denote the M×N-dimensional real number field.

[0102] This embodiment performs sparse reconstruction on all training samples of C categories. For the i-th training sample from the c-th category target the reconstruction error The calculation formula is:

[0103]

[0104]

[0105] Among them, OMP(·) represents the Orthogonal Matching Pursuit algorithm; represents the reconstruction error matrix of the training samples of the c-th class of targets, and represents the number of training samples of the c-th class of targets.

[0106] Calculate the mean reconstruction error of the training samples of each category respectively. For example, the mean reconstruction error of the c-th class is expressed as:

[0107]

[0108] Among them, represents the mean reconstruction error of all training samples in the c-th class, that is, the mean reconstruction error of the c-th class. The mean reconstruction error of the training samples of each category can be calculated according to the above formula.

[0109] The mean reconstruction error of the candidate class is calculated according to the above formula.

[0110] The above-mentioned candidate class measure obtained by calculation in the training stage may include the following steps, specifically including:

[0111] Step 51: Calculate the ratio of the reconstruction error of each category of training samples to the mean reconstruction error of each category, and form a measure matrix;

[0112] Step 52: Select the candidate class measure corresponding to the candidate class from the measure matrix.

[0113] In this embodiment, the candidate class measure can be first determined from the measure matrix, and the Weber distribution is directly fitted to the candidate class measure to obtain the candidate class scale parameter and the candidate class shape parameter. Or it is also possible to fit the Weber distribution to each measure in the measure matrix in the training stage to obtain the scale parameter and shape parameter corresponding to each category, and then select the candidate class scale parameter and the candidate class shape parameter from the scale parameters and shape parameters of each category. For example, calculate the ratio of the reconstruction error of the i-th training sample of the c-th class to the mean reconstruction error of the c-th class as the measure of the i-th training sample of the c-th class, which is expressed as:

[0114]

[0115] Then construct the measure matrix J of the c-th class c :

[0116]

[0117] According to Theorem 1, fit the Weibull distribution to the measure matrix J of the c-th class c to obtain the scale parameter corresponding to the c-th class and the shape parameter

[0118]

[0119] S104: Fit the Weber distribution to the candidate class measure to obtain the candidate class scale parameter and the candidate class shape parameter.

[0120] Candidate class measure Candidate class scale parameter corresponding to the Weber distribution And the candidate class shape parameter

[0121]

[0122] S105: Calculate the ratio of the reconstruction error of the sample to be recognized to the mean of the reconstruction errors of the candidate classes, and use the ratio as the measure of the sample to be recognized.

[0123] In this embodiment, the ratio of the reconstruction error of the sample to be recognized to the mean of the reconstruction errors of the candidate classes is calculated, and the formula is as follows:

[0124]

[0125] Among them, r te Represents the reconstruction error of the sample to be recognized; Represents the measure of the sample to be recognized.

[0126] S106: Calculate the confidence of the fitted Weber distribution according to the measure of the sample to be recognized, the candidate class scale parameter, and the candidate class shape parameter.

[0127] In this embodiment, according to Theorem 2, the measure of the sample to be recognized Candidate class measure Candidate class scale parameter corresponding to the Weber distribution And the candidate class shape parameter Calculate the fitted Weber distribution, and calculate the confidence score, and the formula is as follows:

[0128]

[0129] S107: Compare the confidence with a preset threshold to determine the category of the sample to be recognized.

[0130] In this embodiment, by comparing and judging the confidence with a preset threshold, it is determined that the sample to be recognized is a candidate class in the in-library category Or an out-of-library category.

[0131] The specific discrimination formula is as follows:

[0132]

[0133] Among them, δ represents the threshold; openset represents the out-of-library category; Class y Represents the category of the sample to be recognized.

[0134] At present, radar automatic target recognition (RATR) can be mainly divided into two categories: target recognition based on high range resolution profile (HRRP) and target recognition based on synthetic aperture radar (SAR) image. The target recognition method based on HRRP has the advantages of simple data acquisition, low platform requirements, small storage and calculation amount, etc., and has received extensive attention. HRRP obtains the target echo by a broadband radar, which reflects the distribution of the target structure in the range dimension and plays an important role in target discrimination. The above samples to be recognized and training samples are both sample data of high range resolution images.

[0135] The data processing process of the above samples to be recognized and various category training samples of high range resolution images can specifically include the following steps:

[0136] Step 61: Obtain the initial various category training sample data and the initial sample to be tested of the high range resolution image;

[0137] Step 62: Perform non-coherent averaging, 2-norm amplitude normalization and power transformation on the initial various category sample data and the initial sample to be tested within a preset angle range to obtain various category training samples and samples to be recognized respectively.

[0138] However, in this embodiment, considering that HRRP has pose sensitivity, translation sensitivity and intensity sensitivity, it is necessary to preprocess the data. The specific preprocessing operation is: for HRRP data, use non-coherent averaging within a certain angle range (without scatter point cross-range cell migration) to relax pose sensitivity; 2-norm amplitude normalization to eliminate intensity sensitivity; and enhance the distinguishability of the data through power transformation of the data. The preprocessed data is used as the data of the training samples.

[0139] This embodiment does not limit the training sample data set. For example, the measured data set of HRRP can be directly adopted; or if the measured data set of HRRP is lacking, the HRRP data can be generated by inverting the SAR measured data. For example, using the MSTAR (Moving and Stationary Target Acquisition and Recognition) public data set, the SAR data in MSTAR is inverted to generate HRRP data for experimental verification. The specific inversion process is: perform two-dimensional inverse Fourier transform on the SAR image; then perform deconvolution operation on the obtained data to eliminate the influence of windowing and delete zero elements; then perform two-dimensional Fourier transform on the data and segment the target area; finally, perform inverse Fourier transform on the obtained data in the azimuth dimension to obtain the inverted HRRP data.

[0140] Applying the open-set recognition method for radar targets provided by the embodiments of the present application, by obtaining the reconstruction errors of the samples to be recognized on the sub-dictionaries of various categories, where the sub-dictionaries of various categories are sub-dictionaries of various categories constructed using the training samples of various categories, and the samples to be recognized and the training samples of various categories are high-resolution range profiles; taking the category corresponding to the minimum value in the reconstruction errors as the candidate category, and taking the minimum value as the reconstruction error of the sample to be recognized; obtaining the candidate category reconstruction error mean and candidate category measure obtained in the training stage for the candidate category; fitting the Weber distribution to the candidate category measure to obtain the candidate category scale parameter and candidate category shape parameter; calculating the ratio of the reconstruction error of the sample to be recognized to the candidate category reconstruction error mean, and taking the ratio as the measure of the sample to be recognized; calculating the confidence level of the Weber distribution according to the measure of the sample to be recognized, the candidate category scale parameter, and the candidate category shape parameter; comparing the confidence level with a preset threshold to determine the category of the sample to be recognized. The present application uses the reconstruction error mean ratio as a new measure and fits the Weber distribution, and the performance of the open-set recognition of high-resolution range profiles by this method is also better than that of traditional algorithms.

[0141] Experimental verification is carried out using the method proposed in this embodiment as follows:

[0142] Due to the lack of an HRRP measured data set, the MSTAR (Moving and Stationary Target Acquisition and Recognition) public data set is used, and the SAR data is inversely transformed to generate HRRP data for the experiment. The MSTAR data set is a standard database for SAR target recognition. It contains 10 types of ground targets, namely BMP2, BTR70, T72, BTR60, 2S1, BRDM2, D7, T62, ZIL, and ZSU. The radar operates in the X band, uses spotlight imaging, the image resolution is 0.3m×0.3m, the coverage angle of each image is about 3°, and the data acquisition pitch angles are 15° and 17° respectively. In this experiment, the data with a pitch angle of 17° is selected for training, and the data with a pitch angle of 15° is selected for testing.

[0143] The process of inversely transforming the SAR image to generate HRRP data includes:

[0144] (1) Perform a two-dimensional inverse Fourier transform on the SAR image;

[0145] (2) Perform a deconvolution operation to eliminate the influence of windowing and delete zero elements;

[0146] (3) Perform a two-dimensional Fourier transform and segment the target area;

[0147] (4) Perform an inverse Fourier transform in the azimuth dimension;

[0148] (5) To maintain the unity of data, taking the center of the range dimension of the SAR image as the origin, 64 range cells are taken on each side (left and right).

[0149] Since HRRP has pose sensitivity, translation sensitivity, and intensity sensitivity, it is necessary to preprocess the inverted HRRP data. For the inverted HRRP data, incoherent averaging within a certain angular range is used to relax its pose sensitivity, taking the 2-norm amplitude normalization to eliminate the intensity sensitivity. By performing a power transformation on the data, the discriminability of the data can be enhanced.

[0150] In this embodiment, the overcomplete dictionary is constructed manually, and training samples are used to construct the overcomplete dictionary. For each SAR image with a pitch angle of 17°, the HRRP data generated is extracted for incoherent averaging to generate 1 piece of HRRP data. Sub-dictionaries of each category are constructed in the order of azimuth angle, and the sub-dictionaries are combined into an overcomplete dictionary. To increase the number of test samples, during the test phase, for each SAR image with a pitch angle of 15°, HRRP is generated by extracting incoherent averaging at an azimuth interval of 0.3° for testing.

[0151] The experiment first analyzes the performance of 3 OSR algorithms, namely SROSR, WSVM, and 1-vs-Set. The first 3 types of targets in MSTAR, BMP2, BTR70, and T72, are used as in-library targets, and BTR60 is added as an out-of-library target during the test phase to test the recognition performance in the presence of out-of-library targets. The experiment uses accuracy and F-value as evaluation indicators. Among them, accuracy reflects the accuracy of the model classification, and the F-value (the statistic value of the F-test) comprehensively evaluates the recall rate and precision rate of the classifier, and is a comprehensive indicator for evaluating open-set recognition algorithms.

[0152]

[0153]

[0154] Among them, TP, TN, FP, and FN represent true positive class, true negative class, false positive class, and false negative class respectively. Recall rate Precision rate The performance results of different algorithms are compared as shown in Table 1.

[0155] Table 1

[0156] Algorithm Accuracy F-value SROSR 0.672 0.759 WSVM 0.459 0.536 1-vs-Set 0.39 0.44

[0157] As can be seen from Table 1, both the accuracy rate and F value of the SROSR algorithm are higher than those of other algorithms. Among them, the accuracy rate of the SROSR algorithm is 21.3% higher than that of the WSVM and 28.2% higher than that of the 1-vs-Set algorithm. The F value of the SROSR is 22.3% higher than that of the WSVM and 31.9% higher than that of the 1-vs-Set algorithm. Through comprehensive comparison, the performance of the SROSR algorithm is relatively better. Therefore, the proposed method will be mainly compared with the SROSR algorithm in the following experiments.

[0158] The parameter settings of the SROSR algorithm and the proposed method are shown in Table 2. SROSR adopts weighted summation processing for both the matching class and non-matching class, and weights need to be set. The proposed method only reconstructs the error distance from the center distance measure of the matching class to fit the Weibull distribution, so no weights and tail parameters are set.

[0159] Table 2

[0160] Algorithm Tail size Threshold Weight This method - 0.33 - SROSR 0.5 0.15 0.3

[0161] (1) Recognition performance with unknown classes.

[0162] In the experiment, BMP2, BTR70, and T72 were selected as in-library targets, and BTR60 was used as an out-of-library target for the experiment. The confusion matrices of the SROSR algorithm and the proposed method with unknown classes are shown in Tables 3 and 4 respectively.

[0163] Table 3

[0164]

[0165] Table 4

[0166]

[0167] As can be seen from Tables 3 and 4, the proposed method has improved both the in-library target recognition rate and the out-of-library target rejection rate compared with the SROSR. Among them, the in-library target recognition rate has increased by 6.2%, and the out-of-library target rejection rate has increased by 2.6%, enhancing the discrimination ability of in-library targets and the rejection ability of unknown classes.

[0168] The experiment further verifies the recognition performance under 10 types of test targets. BMP2, BTR70, and T72 were used as known in-library targets, and BTR60, 2S1, BRDM2, D7, T62, ZIL, and ZSU target data were added as the test data set on the basis of the in-library targets. The confusion matrices of the SROSR algorithm and the proposed method in this scenario are shown in Tables 5 and 6.

[0169] Table 5

[0170]

[0171] Table 6

[0172]

[0173] As can be seen from Table 5 and Table 6, in the typical case of training 3 categories and testing 10 categories, this method achieves better results compared with the SROSR algorithm. Among them, for the judgment of in-library targets, the performance of the SROSR algorithm is similar to that of this method, but for the discrimination of out-of-library targets, the overall out-of-library target rejection rate of this method has increased by 15.9%. The different performance indicators of the two algorithms are shown in Table 7.

[0174] Table 7

[0175] Algorithm Accuracy F-value This method 0.754 0.624 SROSR 0.637 0.532

[0176] According to Table 7, this method is superior to the SROSR algorithm in both accuracy and F-value. Accuracy reflects the proportion of true classes among all judgments as positive classes. It can be seen that both algorithms have a certain proportion of misjudgments for out-of-library samples, but the accuracy of this method has increased by 12.3% compared with the SROSR algorithm. The F-value comprehensively evaluates the performance of the algorithm, and this method has increased by 9.2% in the F-value compared with the SROSR algorithm. In summary, this method has achieved better open-set recognition and classification performance than the SROSR algorithm in typical cases.

[0177] (2) Recognition performance under different degrees of openness.

[0178] The experiment further verifies the recognition results of the algorithm under different degrees of openness. Taking BMP2, BTR70, and T72 at a 17° pitch angle as in-library target data, and based on the 3-class target data at a 15° pitch angle, the target data of BTR60, 2S1, BRDM2, D7, T62, ZIL, and ZSU at a 15° pitch angle are sequentially added as the test data set according to the different numbers of test categories. The test categories are 4 to 10 classes in sequence. The degrees of openness corresponding to different test categories are shown in Table 8, and its experimental results are as Figure 2 and Figure 3 shown.

[0179] Table 8

[0180]

[0181] Figure 2 is a graph of the accuracy results of an open-set recognition algorithm under different degrees of openness provided by an embodiment of the present application, Figure 3 is a graph of the F-value results of an open-set recognition algorithm under different degrees of openness provided by an embodiment of the present application. From Figure 2 and Figure 3It can be seen that as the degree of openness increases, the accuracy of the algorithm is continuously improving. The change in accuracy is related to the recognition rate of the classifier for this category. In open-set recognition, it is reflected in the rejection of newly added unknown category samples. If the classifier has strong discrimination ability for newly added categories, the accuracy will increase significantly; otherwise, it will decrease. The F-value comprehensively evaluates the comprehensive recognition ability of the algorithm for in-library and out-of-library categories. When the number of newly added categories increases, while the algorithm rejects out-of-library targets, it also increases the difficulty of recognizing in-library categories. Therefore, the F-value shows a downward trend, but the F-value of this method still reaches the optimal among the four open-set recognition algorithms. Considering the two performance indicators of accuracy and F-value comprehensively, this method has better recognition performance and rejection ability compared with other open-set recognition algorithms.

[0182] From the above comprehensive experiments, it can be seen that this method has good recognition performance and rejection ability in simulation experiments under different scenarios.

[0183] In this embodiment, the mean ratio of reconstruction errors is used as a measure, and the extreme value theory is used to fit the measure distribution, without the need for manual setting of the reconstruction error tail and weight parameters. Experimental verification has been carried out on the HRRP data inverted by MSTAR. The results show that this method has certain improvements in performance indicators such as accuracy and F-value compared with the classical SROSR algorithm, and also has the best performance compared with other mainstream OSR methods, and has good rejection ability for out-of-library targets.

[0184] Next, the high-resolution range profile open-set recognition device provided by the embodiments of the present application will be introduced. The high-resolution range profile open-set recognition device described below can be correspondingly referred to the high-resolution range profile open-set recognition method described above.

[0185] Specifically, please refer to Figure 4 , Figure 4 which is a schematic structural diagram of a high-resolution range profile open-set recognition device provided by an embodiment of the present application, and may include:

[0186] A first acquisition module 100, configured to acquire the reconstruction errors of the samples to be recognized on the sub-dictionaries of each category, where the sub-dictionaries of each category are the sub-dictionaries of each category constructed by using the training samples of each category, and the samples to be recognized and the training samples of each category are high-resolution range profiles;

[0187] A candidate class determination module 200, configured to use the category corresponding to the minimum value in the reconstruction errors as the candidate class, and use the minimum value as the reconstruction error of the sample to be recognized;

[0188] A second acquisition module 300, configured to acquire the mean value of the candidate class reconstruction errors and the candidate class measure obtained in the training stage of the candidate class;

[0189] The fitting module 400 is configured to fit the Weber distribution to the candidate class measure to obtain a candidate class scale parameter and a candidate class shape parameter;

[0190] The measure calculation module 500 for the sample to be recognized is configured to calculate the ratio of the reconstruction error of the sample to be recognized to the mean value of the reconstruction errors of the candidate classes, and use the ratio as the measure of the sample to be recognized;

[0191] The confidence calculation module 600 is configured to calculate the confidence of the Weber distribution according to the measure of the sample to be recognized, the candidate class scale parameter, and the candidate class shape parameter;

[0192] The class determination module 700 is configured to compare the confidence with a preset threshold to determine the class of the sample to be recognized.

[0193] Based on the above embodiments, the first acquisition module 100 may include:

[0194] An acquisition unit configured to acquire the training samples of each class;

[0195] A construction combination unit configured to construct the sub-dictionaries of each class by using the training samples of each class, and combine the sub-dictionaries of each class into an over-complete dictionary;

[0196] A sparse coefficient calculation unit configured to calculate, by using a sparse algorithm, a first sparse coefficient corresponding to the sample to be recognized on the sub-dictionaries of each class c = 1, …, C, the is in the form of a column vector, and C represents the number of training sample classes;

[0197] A reconstruction error calculation unit for each class is configured to calculate the reconstruction error of the sample to be recognized on the sub-dictionaries of each class according to the first sparse coefficient, the sub-dictionaries of each class, and the sample to be recognized.

[0198] Based on the above embodiments, the candidate class determination module 200 may include:

[0199] A candidate class determination unit configured to, according to determine the candidate class; y represents the sample to be recognized and is in vector form; D represents the over-complete dictionary and is in matrix form; the represents the candidate class; represents the first sparse coefficient calculated by the sample to be recognized on the sub-dictionaries of each class and combined end to end into a second sparse coefficient, which is in the form of a column vector; represents retaining the coefficient elements corresponding to the dictionary atoms of the c-th class in, and setting other elements to zero at the same time.

[0200] Based on the above embodiments, obtaining the average reconstruction error of the candidate class obtained during the training phase in the second obtaining module 300 may include:

[0201] A sparse reconstruction unit for performing sparse reconstruction on training samples in each training category on their corresponding sub-dictionaries using the sparse algorithm to obtain training sample reconstruction errors;

[0202] A reconstruction error mean calculation unit for each category of training samples, configured to calculate the average reconstruction error of each category of training samples based on the training sample category and the training sample reconstruction error;

[0203] A first selection unit for selecting the candidate class reconstruction error mean corresponding to the candidate class from the average reconstruction errors of each category of training samples.

[0204] Based on the above embodiments, obtaining the candidate class measure obtained during the training phase in the second obtaining module 300 may include:

[0205] A measure calculation unit for calculating the ratio of the reconstruction error of each category of training samples to the average reconstruction error of each category of training samples and forming a measure matrix;

[0206] A second selection unit for selecting the candidate class measure corresponding to the candidate class from the measure matrix.

[0207] Based on the above embodiments, where the to-be-identified sample and each category of training samples are high-resolution range images, it may include:

[0208] An initial sample data acquisition unit for acquiring initial training sample data and an initial to-be-tested sample of each category of the high-resolution range image;

[0209] A preprocessing unit for performing incoherent averaging, 2-norm amplitude normalization, and power transformation on the initial sample data of each category and the initial to-be-tested sample within a preset angle range to obtain each category of training samples and the to-be-identified sample respectively.

[0210] It should be noted that the modules and units in the above open-set recognition device for radar targets can be changed in order before and after without affecting the logic.

[0211] Applying the high-resolution range image open-set recognition device provided by the embodiments of the present application, through the first acquisition module 100, which is used to acquire the reconstruction errors of the sample to be recognized on various category sub-dictionaries. The various category sub-dictionaries are the various category sub-dictionaries constructed using various category training samples. The sample to be recognized and the various category training samples are high-resolution range images; the candidate category determination module 200 is used to take the category corresponding to the minimum value in the reconstruction errors as the candidate category, and take the minimum value as the reconstruction error of the sample to be recognized; the second acquisition module 300 is used to acquire the candidate category reconstruction error mean and the candidate category measure obtained during the training stage of the candidate category; the fitting module 400 is used to fit the Weber distribution to the candidate category measure to obtain the candidate category scale parameter and the candidate category shape parameter; the sample to be recognized measure calculation module 500 is used to calculate the ratio of the reconstruction error of the sample to be recognized to the candidate category reconstruction error mean, and take the ratio as the measure of the sample to be recognized; the confidence calculation module 600 is used to calculate the confidence of the Weber distribution according to the measure of the sample to be recognized, the candidate category scale parameter, and the candidate category shape parameter; the judgment module 700 is used to compare the confidence with a preset threshold to determine the category of the sample to be recognized. By using the reconstruction error mean ratio as a new measure and fitting the Weber distribution, the present application does not require manual setting of the reconstruction error tail and weight parameters; and the performance of the high-resolution range image open-set recognition method of this method is also better than that of traditional algorithms.

[0212] The high-resolution range image open-set recognition device provided by the embodiments of the present application will be introduced below. The high-resolution range image open-set recognition device described below can be mutually corresponded and referred to the high-resolution range image open-set recognition method described above.

[0213] Please refer to Figure 5 , Figure 5 which is a schematic structural diagram of a high-resolution range image open-set recognition device provided by an embodiment of the present application, and may include:

[0214] A memory 10 for storing computer programs;

[0215] A processor 20 for executing the computer program to implement the above-mentioned high-resolution range image open-set recognition method.

[0216] The memory 10, the processor 20, and the communication interface 31 complete mutual communication through the communication bus 32.

[0217] In the embodiments of the present application, the memory 10 is used to store one or more programs. The program may include program codes, and the program codes include computer operation instructions. In the embodiments of the present application, the memory 10 may store programs for implementing the following functions:

[0218] Obtain the reconstruction errors of the sample to be recognized on the sub-dictionaries of each category. The sub-dictionaries of each category are the sub-dictionaries of each category constructed using the training samples of each category. The sample to be recognized and the training samples of each category are high-resolution range images;

[0219] Take the category corresponding to the minimum value in the reconstruction errors as the candidate category, and take the minimum value as the reconstruction error of the sample to be recognized;

[0220] Obtain the average reconstruction error of the candidate category and the measure of the candidate category obtained during the training phase;

[0221] Fit the Weber distribution to the measure of the candidate category to obtain the scale parameter and shape parameter of the candidate category;

[0222] Calculate the ratio of the reconstruction error of the sample to be recognized to the average reconstruction error of the candidate category, and take the ratio as the measure of the sample to be recognized;

[0223] Calculate the confidence level of the Weber distribution according to the measure of the sample to be recognized, the scale parameter of the candidate category, and the shape parameter of the candidate category;

[0224] Compare the confidence level with a preset threshold to determine the category of the sample to be recognized.

[0225] In a possible implementation, the memory 10 may include a program storage area and a data storage area. Among them, the program storage area may store an operating system and application programs required for at least one function, etc.; the data storage area may store the data created during use.

[0226] In addition, the memory 10 may include a read-only memory and a random access memory, and provide instructions and data to the processor. A part of the memory may also include NVRAM. The memory stores an operating system and operation instructions, executable modules or data structures, or subsets thereof, or extended sets thereof. Among them, the operation instructions may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic tasks and processing hardware-based tasks.

[0227] The processor 20 may be a central processing unit (CPU), an application-specific integrated circuit, a digital signal processor, a field programmable gate array, or other programmable logic devices. The processor 20 may be a microprocessor or any conventional processor, etc. The processor 20 may call the program stored in the memory 10.

[0228] The communication interface 31 may be an interface of a communication module for connecting to other devices or systems.

[0229] Of course, it should be noted that Figure 5The structure shown does not constitute a limitation on the high-resolution range image open-set recognition device in the embodiments of the present application. In practical applications, the high-resolution range image open-set recognition device may include more or fewer components than those Figure 5 shown, or combine certain components.

[0230] Next, the storage medium provided by the embodiments of the present application will be introduced. The storage medium described below can be correspondingly referred to in relation to the high-resolution range image open-set recognition method described above.

[0231] The present application also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned high-resolution range image open-set recognition method are implemented.

[0232] The storage medium may include various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs.

[0233] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0234] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in the form of hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0235] Finally, it should also be noted that in this article, relationships such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "including", "comprising", or any other variant is intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article, or device.

[0236] The above has introduced in detail a high-resolution range image open set recognition method, device, equipment and storage medium provided by the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A high-resolution range profile open set recognition method, characterized in that, Including: Obtaining the reconstruction error of the sample to be recognized on each category sub-dictionary, where each category sub-dictionary is the category sub-dictionary constructed using training samples of each category, and the sample to be recognized and the training samples of each category are high-resolution range images; Taking the category corresponding to the minimum value in the reconstruction errors as the candidate category, and taking the minimum value as the reconstruction error of the sample to be recognized; Obtaining the mean value of the candidate category reconstruction error and the candidate category measure obtained in the training stage for the candidate category; Fitting a Weibull distribution to the candidate category measure to obtain the candidate category scale parameter and the candidate category shape parameter; Calculating the ratio of the reconstruction error of the sample to be recognized to the mean value of the candidate category reconstruction error, and taking the ratio as the measure of the sample to be recognized; Calculating the confidence level of the Weibull distribution according to the measure of the sample to be recognized, the candidate category scale parameter, and the candidate category shape parameter; Comparing the confidence level with a preset threshold to determine the category of the sample to be recognized; The obtaining the reconstruction error of the sample to be recognized on each category sub-dictionary includes: Obtaining the training samples of each category; Constructing each category sub-dictionary using the training samples of each category, and combining each category sub-dictionary into an over-complete dictionary; Calculating, by using a sparse algorithm, a first sparse coefficient corresponding to the to-be-recognized sample on each category sub-dictionary , , where the is in the form of a column vector, and the represents the number of training sample categories; Calculating the reconstruction error of the sample to be recognized on each category sub-dictionary according to the first sparse coefficient, each category sub-dictionary, and the sample to be recognized.

2. The high-resolution range image open set recognition method according to claim 1, wherein The taking the category corresponding to the minimum value in the reconstruction errors as the candidate category includes: According to determine the candidate class; The said represents the sample to be recognized, in vector form; the represents the overcomplete dictionary, in matrix form; the represents the candidate class; represents the first sparse coefficients calculated by the sample to be recognized on the sub-dictionaries of each category and combined end to end to form the second sparse coefficient, in column vector form; represents to retain the coefficient elements corresponding to the dictionary atoms of the category in, and set other elements to zero at the same time.

3. The high-resolution range image open set recognition method according to claim 1, characterized in that The obtaining the mean value of the candidate category reconstruction error obtained in the training stage for the candidate category includes: Performing sparse reconstruction on the training samples in each training category on their corresponding category sub-dictionaries using the sparse algorithm to obtain the training sample reconstruction errors; Calculating the mean value of the reconstruction errors of the training samples of each category according to the training sample category and the training sample reconstruction errors; Selecting the candidate category reconstruction error mean value corresponding to the candidate category from the mean values of the reconstruction errors of the training samples of each category.

4. The high-resolution range profile open set recognition method according to claim 1, wherein The obtaining the candidate category measure obtained in the training stage for the candidate category includes: Calculating the ratio of the reconstruction error of each category training sample to the mean value of the reconstruction errors of the training samples of each category, and forming a measure matrix; Selecting the candidate category measure corresponding to the candidate category from the measure matrix.

5. The high-resolution range image open set recognition method according to claim 1, characterized in that The sample to be recognized and the training samples of each category are high-resolution range images, including: Obtaining the initial training sample data of each category and the initial sample to be tested for the high-resolution range image; Performing incoherent averaging, 2-norm amplitude normalization, and power transformation on the initial sample data of each category and the initial sample to be tested within a preset angle range to obtain the training samples of each category and the sample to be recognized respectively.

6. An open set recognition device for high-resolution range profiles, characterized in that, Including: A first obtaining module, configured to obtain the reconstruction error of the sample to be recognized on each category sub-dictionary, where each category sub-dictionary is the category sub-dictionary constructed using the training samples of each category, and the sample to be recognized and the training samples of each category are high-resolution range images; A candidate category determination module, configured to take the category corresponding to the minimum value in the reconstruction errors as the candidate category, and take the minimum value as the reconstruction error of the sample to be recognized; A second acquisition module, configured to acquire the mean value of the candidate class reconstruction error and the candidate class measure obtained by the candidate class during the training phase; A fitting module, configured to fit the Weber distribution to the candidate class measure to obtain a candidate class scale parameter and a candidate class shape parameter; A measure calculation module for the sample to be recognized, configured to calculate the ratio of the reconstruction error of the sample to be recognized to the mean value of the candidate class reconstruction error, and use the ratio as the measure of the sample to be recognized; A confidence calculation module, configured to calculate the confidence of the Weber distribution according to the measure of the sample to be recognized, the candidate class scale parameter, and the candidate class shape parameter; A category determination module, configured to compare the confidence with a preset threshold to determine the category of the sample to be recognized; The first acquisition module includes: An acquisition unit, configured to acquire the training samples of each category; A construction combination unit, configured to construct the sub-dictionaries of each category by using the training samples of each category, and combine the sub-dictionaries of each category into an over-complete dictionary; A sparse coefficient calculation unit, configured to calculate, by using a sparse algorithm, a first sparse coefficient corresponding to the sample to be recognized on each category sub-dictionary , , the is in the form of a column vector, and the represents the number of training sample categories; Reconstruction error calculation units for each category, configured to calculate the reconstruction error of the sample to be recognized on the sub-dictionaries of each category according to the first sparse coefficient, the sub-dictionaries of each category, and the sample to be recognized.

7. A high-resolution range profile open set recognition device, characterized in that It includes: A memory, configured to store a computer program; A processor, configured to implement the steps of the high-resolution range image open-set recognition method according to any one of claims 1 to 5 when executing the computer program.

8. A storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is executed by a processor, the steps of the high-resolution range image open-set recognition method according to any one of claims 1 to 5 are implemented.

Images