Unmanned aerial vehicle open set identification method and system

Abstract

The invention discloses an unmanned aerial vehicle open set identification method and system, and the method comprises the steps: enabling a cross training set to be used for constructing an over-complete dictionary, and solving a reconstruction error of cross test set data through employing a sparse representation algorithm; selecting tail parameters and reconstruction errors, sorting each type of reconstruction errors, selecting trailing data and a threshold value, calculating an overflow value, and estimating an extreme value parameter by using maximum likelihood estimation according to the overflow value to obtain the extreme value parameter; solving reconstruction errors of the test set, taking the in-library target category corresponding to the minimum reconstruction error as a candidate category, and obtaining an optimal parameter according to the candidate category and the obtained extreme value parameter; the optimal parameters are an optimal tail parameter and an optimal threshold value; and based on the optimal parameters, the unmanned aerial vehicle target is identified. According to the invention, the specified type of unmanned aerial vehicle can be effectively identified and other types of low-slow small targets are rejected.

Description

Technical Field

[0001] The present application relates to the technical field of drone identification, and in particular to a drone open set identification method and system. Background Art

[0002] In practical applications, there is often a lack of prior information about non-cooperative, low, slow, and small targets, which makes it impossible to build a complete feature library when designing a classifier. The classifier cannot accurately reject non-cooperative targets, and traditional recognition methods can no longer meet the needs. The open set recognition problem accurately identifies known class targets and accurately rejects unknown class targets by solving the open set risk minimization problem, providing a good solution for the prevention and control of non-cooperative, low, slow, and small targets.

[0003] In the existing technical solution, the YOLOv4 network is improved. The concat (connection layer) layer fuses the feature maps output by the first CBL (a module consisting of convolution layer + batch normalization layer + activation function) and the third CBL, making network training easier to converge; in the CSP (cross stage partial) structure, one CBL is removed and a depth-separable convolution layer is added to the Res unit (residual component), which can reduce the computational complexity of a part of the network while ensuring the number of convolution layers.

[0004] The existing technical solutions include: collecting and labeling drone image data in non-cooperative scenarios; extracting image features from image data using a bottleneck feature extraction module; fusing the extracted image features using a feature fusion network; and identifying targets based on the feature fusion results. It uses bottleneck features to reduce feature dimensions, uses dilated convolution to reduce the amount of calculation, and implements target recognition based on feature fusion results. However, in actual applications, non-cooperative data is often difficult to obtain, and the quality and quantity of image data are difficult to guarantee.

[0005] A convolutional neural network was designed to identify the collected drone pictures and bird pictures, and an experimental comparison was made with the support vector machine algorithm. The recognition results of the convolutional neural network were better than those of the support vector machine. First, four convolutional neural networks were used to pre-train the data set to extract the deep features of the target, and then the target was transferred to learn and the classification model was constructed. After that, the ensemble learning method of relative majority voting and weighted average was used to integrate the classification model to obtain the transfer ensemble model to identify the data.

[0006] In summary, the current domestic and foreign patents and related journal papers rarely pay attention to the rejection of non-cooperative low, slow and small (flying low, slow speed, small target) targets, and there are even fewer reports on the problem of open set recognition of drones. Therefore, the research on open set recognition of drones is of great significance. Summary of the invention

[0007] In view of this, the present application provides a method and system for open set identification of drones, which aims at the problem of drone target identification under the condition of lack of prior information of non-cooperative targets. By introducing an open set identification method into the field of drone identification, open set identification of drone targets is achieved, which can effectively identify drones of specified types and reject other types of low, slow and small targets, where rejection refers to the rejection of categories in the library, that is, it is judged as a category not in the library.

[0008] The present application discloses a method for identifying an open set of drones, which includes: Step 1: Divide the UAV optical image data into a training set and a test set. In the training phase of the cross-training set, obtain the in-library target data for fitting the extreme value distribution, and divide the training set into a cross-training set and a cross-test set. Step 2: In order to obtain the reconstruction error information of the target data in the database, the cross-training set is used to build an over-complete dictionary, and the sparse representation algorithm is used to obtain the reconstruction error of the cross-test set data; Step 3: Select the tail parameters and the reconstruction error output from step 2, sort each type of reconstruction error, select the tail data and threshold value, calculate the excess value, and estimate the extreme value parameter based on the excess value using maximum likelihood estimation to obtain the extreme value parameter; Step 4: Obtain the reconstruction error of the test set, take the target category in the library corresponding to the minimum reconstruction error as the candidate category, and obtain the optimal parameters based on the candidate category and the obtained extreme value parameters; the optimal parameters are the optimal tail parameter and the optimal threshold; Step 5: Based on the optimal parameters, identify the drone target.

[0009] Furthermore, the step 1 comprises: Select one type of drone optical image data from the drone optical image data as the non-cooperative target data outside the database, and select the remaining types of drone optical image data as the target data inside the database; Divide the target data in the database into training sets and test sets in proportion, and put all non-cooperative target data outside the database into the test set; The target data in the database is divided into training set and test set in proportion, and all non-cooperative target data outside the database are included in the test set.

[0010] Furthermore, the step 2 comprises: In the training phase of the cross-training set, the reconstruction error information of the optical image data of the cth type of UAV in the target data library is ; represents the cross-training set tr, tr is the cross-training set subscript, the reconstruction error corresponding to the Nth target in the optical image data of the cth type UAV, and T represents transposition; Select the reconstruction error information of all c-th type UAV optical image data to build an over-complete dictionary ; Represents the reconstruction error information of the optical image data of the c-th drone obtained in the cross-training phase; the value range of c is 1 to C, and C is the total number of target categories in the library, which is a positive integer; Use sparse representation algorithm to obtain the reconstruction error of the cross test set , It represents the reconstruction error information obtained from the c-th type of UAV optical image data in the cross test set according to the over-complete dictionary D.

[0011] Furthermore, the step 3 comprises: Select tail parameters and reconstruction error , sort the c-th type reconstruction errors in order of size:

[0012] in, is the reconstruction error information of the target data in the c-th library, is the number of c-th category targets in the cross-test set, Sort the reconstruction errors of the c-th target from small to large, reconstruction error; Select tailing data and threshold , calculate the excess value ; Sort the reconstruction errors of the cth type of targets from small to large. The reconstruction error, For the The reconstruction error, is the threshold value, and its value is The reconstruction error, For the reconstruction error; Using maximum likelihood estimation, the extreme value parameter is estimated by the following formula to obtain the extreme value parameter estimate of the cth class: , :

[0013] In the formula, For parameters and The likelihood function of and is the extreme value parameter, is the number of c-th category targets in the cross-test set, Represents the likelihood function, i is a positive integer, and its value range is 1 to .

[0014] Furthermore, the step 4 comprises: The sparse representation algorithm is used to obtain the reconstruction error of the test set, and the target category in the library corresponding to the minimum reconstruction error is As a candidate category, obtain the target category in the library The corresponding extreme value parameters and , calculate the extreme value parameters and The confidence level of the corresponding extreme value distribution is compared with the set threshold to determine the category to which the test sample in the test set belongs and adjust the tail parameters , repeat steps 3 and 4 until the optimal parameters are obtained; the optimal parameters are the optimal tail parameters and the optimal threshold.

[0015] Furthermore, in step 4, the candidate categories are calculated , where Y is the test data, D is the overcomplete dictionary, and C is the total number of target categories in the database. To preserve the sparse coefficient matrix The coefficient elements of the training samples corresponding to category c in the , and the coefficient elements of the remaining categories in the target category in the corresponding library are set to zero, express norm; the sparse coefficient matrix has a corresponding coefficient element for each category in the target category in the library.

[0016] Further categories include: If the confidence is less than or equal to the set threshold, the category to which the test sample in the test set belongs is considered to be the target category in the library, otherwise it is an out-of-library target.

[0017] Furthermore, the step 5 comprises: The reconstruction error of the input unknown target information is calculated, and the target category in the library corresponding to the minimum reconstruction error is taken as the candidate class. The corresponding extreme value distribution confidence is calculated according to the extreme value parameters corresponding to the optimal tail parameters. Finally, the confidence is compared with the optimal threshold to determine whether the UAV target belongs to the target category in the library or the target outside the library.

[0018] Furthermore, before step 1, the method further includes: Collect optical image data of drones taken by sensor equipment.

[0019] The present application also discloses a drone open set identification system, which implements the above-mentioned drone open set identification method, and comprises: A partitioning module is used to partition the optical image data of the UAV into a training set and a test set. In the training phase of the cross-training set, the target data in the library for fitting the extreme value distribution is obtained, and the training set is partitioned into a cross-training set and a cross-test set. The reconstruction module is used to obtain the reconstruction error information of the target data in the library, use the cross-training set to build an over-complete dictionary, and use the sparse representation algorithm to obtain the reconstruction error of the cross-test set data; A calculation module is used to select the tail parameter and the reconstruction error output by the reconstruction module, sort each type of reconstruction error, select the tail data and the threshold value, calculate the excess value, and estimate the extreme value parameter according to the excess value using the maximum likelihood estimation to obtain the extreme value parameter; The optimization module is used to obtain the reconstruction error of the test set, and the target category in the library corresponding to the minimum reconstruction error is used as the candidate category. According to the candidate category and the obtained extreme value parameters, the optimal parameters are obtained; the optimal parameters are the optimal tail parameters and the optimal threshold; The recognition module is used to identify the drone target based on the optimal parameters.

[0020] Due to the adoption of the above technical solution, the present application has the following advantages: The present application provides a method and system for open set identification of drones. For the problem of drone target identification under the condition of lack of prior information of non-cooperative targets, the open set identification method is introduced into the field of drone identification to achieve open set identification of drone targets. After experimental verification, the method of the present application can effectively reject non-cooperative low, slow and small targets, and can reject flying birds. For the problem of drone target identification under the condition of lack of prior information of non-cooperative targets, the open set identification method is introduced into the field of drone identification to achieve open set identification of drone targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0022] Figure 1 A schematic diagram of a flow chart of a method for open set identification of drones according to an embodiment of the present application; FIG2( a ) is a schematic diagram of the accuracy of an embodiment of the present application; FIG. 2( b ) is a schematic diagram of the F value of an embodiment of the present application. DETAILED DESCRIPTION

[0023] The present application is further described in conjunction with the accompanying drawings and embodiments, and the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present application.

[0024] Traditional UAV image methods do not consider non-cooperative targets when training classifiers. In applications, it is often difficult to identify low, slow, and small targets outside the feature library. Figure 1 The present application provides an embodiment of an open set recognition method for drones. Compared with the traditional drone image recognition method, the present application mainly realizes open set recognition of drone targets by introducing an open set recognition method into the drone recognition field.

[0025] The present application embodiment includes the following steps: S1: Divide the optical image data of the UAV into a training set and a test set. In the training phase of the cross-training set, obtain the in-library target data for fitting the extreme value distribution, and divide the training set into a cross-training set and a cross-test set. S2: In order to obtain the reconstruction error information of the target data in the database, the cross-training set is used to build an over-complete dictionary, and the sparse representation algorithm is used to obtain the reconstruction error of the cross-test set data; S3: Select the tail parameters and the reconstruction error output by S2, sort each type of reconstruction error, select the tail data and threshold value, calculate the excess value, and estimate the extreme value parameters based on the excess value using maximum likelihood estimation to obtain the extreme value parameters; S4: Obtain the reconstruction error of the test set, take the target category in the library corresponding to the minimum reconstruction error as the candidate category, and obtain the optimal parameters based on the candidate category and the obtained extreme value parameters; the optimal parameters are the optimal tail parameters and the optimal threshold; S5: Use the optimal parameters obtained in S4 to build an open set recognition classifier to identify drone targets.

[0026] Optionally, S1 includes: Select one type of drone optical image data from the drone optical image data as the non-cooperative target data outside the database, and select the remaining types of drone optical image data as the target data inside the database; The target data in the database is divided into training set and test set in proportion, and all non-cooperative target data outside the database are included in the test set.

[0027] Specifically, the data set is divided into: labels are set for different types of drone data, and divided into training sets and test sets. Labels are manually divided into different categories for image data of different types of drones before training, and different categories correspond to different labels. In order to simulate the conditions of non-cooperative targets in actual applications, one type of drone data is selected as non-cooperative target data outside the library, and this type of drone data is not included in the training set. The remaining types of drones are used as targets in the library. The target data in the library are divided into training sets and test sets at a ratio of 80% and 20% respectively, and all non-cooperative drone data outside the library are included in the test set. The training set is divided according to the proportion, for example, the cross-training set data accounts for 80% of the training set, and the cross-test set data accounts for 20% of the training set; Training set division: In the training stage, the target data in the library for fitting the extreme value distribution needs to be obtained, and the training set data is divided into cross-training sets and cross-test sets, accounting for 80% and 20% of the training set data respectively.

[0028] Optionally, S2 includes: In the cross-training stage, the reconstruction error information of the optical image data of the cth type of UAV in the target data library is: ; represents the cross-training set tr, tr is the cross-training set subscript, the reconstruction error corresponding to the Nth target in the optical image data of the cth type UAV, and T represents transposition; Select the reconstruction error information of all c-th type UAV optical image data to build an over-complete dictionary ; represents the reconstruction error information of the optical image data of the c-th type of UAV obtained during the cross-training phase; Use sparse representation algorithm to obtain the reconstruction error of the cross test set , It represents the reconstruction error information obtained from the optical image data of the cth type of UAV in the cross test set according to the overcomplete dictionary D. The reconstruction error is the difference between the original data and the reconstructed data in the cross test set. The reconstructed data is the product of the overcomplete dictionary and the sparse coefficient, that is, ;in, is the optical image data of the cth type of drone in the cross-test set data, D is an over-complete dictionary, is the sparse coefficient, the value range of c is 1 to C, C is the total number of target categories in the library, which is a positive integer.

[0029] Optionally, S3 includes: Select tail parameters and reconstruction error , sort the c-th type reconstruction errors in order of size:

[0030] in, is the reconstruction error information of the target data in the c-th library, is the number of c-th category targets in the cross-test set, Sort the reconstruction errors of the c-th target from small to large, reconstruction error; Select tailing data and threshold , calculate the excess value ; Sort the reconstruction errors of the cth type of targets from small to large. The reconstruction error, For the The reconstruction error, is the threshold value, and its value is The reconstruction error, For the reconstruction error; Using maximum likelihood estimation, the extreme value parameter is estimated by the following formula to obtain the extreme value parameter estimate of the cth class: , :

[0031] In the formula, For parameters and The likelihood function of and is the extreme value parameter, is the number of c-th category targets in the cross-test set, Represents the likelihood function, i is a positive integer, and its value range is 1 to .

[0032] Optionally, S4 includes: The sparse representation algorithm is used to obtain the reconstruction error of the test set, and the target category in the library corresponding to the minimum reconstruction error is As a candidate category, obtain the target category in the library The corresponding extreme value parameters and (can be obtained according to the formula for estimating the extreme value parameter in step S3), calculate the extreme value parameter and The confidence of the corresponding extreme value distribution (the corresponding extreme value distribution can be directly obtained according to the extreme value parameter, and then the confidence of the test set data in the test phase for the extreme value distribution is calculated) and used as the discriminant score , the judgment score With the set threshold Compare and determine the category to which the test samples in the test set belong. Specifically,

[0033] If the discrimination score Greater than the set threshold , then the category to which the test samples in the test set belong is the target outside the library (newclass), otherwise it belongs to the target category within the library; Adjust tail parameters , repeat S3 and S4 until the optimal parameters are obtained; the optimal parameters are the optimal tail parameter and the optimal threshold.

[0034] Optionally, in S4, calculate the candidate categories , where Y is the test data, D is the overcomplete dictionary, and C is the total number of target categories in the database. To preserve the sparse coefficient matrix The coefficient elements of the training samples corresponding to category c in the , and the coefficient elements of the remaining categories in the target category in the corresponding library are set to zero, express norm; the sparse coefficient matrix has a corresponding coefficient element for each category in the target category in the library.

[0035] Optionally, S5 includes: The reconstruction error of the input unknown target information is calculated, and the target category in the library corresponding to the minimum reconstruction error is taken as the candidate class. The corresponding extreme value distribution confidence is calculated according to the extreme value parameters corresponding to the optimal tail parameters. Finally, the confidence is compared with the optimal threshold to determine whether the UAV target belongs to the target category in the library or the target outside the library.

[0036] Optionally, before S1, it also includes: Collect optical image data of drones taken by sensor equipment.

[0037] The present application also provides an embodiment of a drone open set identification system, which implements the drone open set identification method described in the above embodiment, and includes: A partitioning module is used to partition the optical image data of the UAV into a training set and a test set. In the training phase of the cross-training set, the target data in the library for fitting the extreme value distribution is obtained, and the training set is partitioned into a cross-training set and a cross-test set. The reconstruction module is used to obtain the reconstruction error information of the target data in the library, use the cross-training set to build an over-complete dictionary, and use the sparse representation algorithm to obtain the reconstruction error of the cross-test set data; A calculation module is used to select the tail parameter and the reconstruction error output by the reconstruction module, sort each type of reconstruction error, select the tail data and the threshold value, calculate the excess value, and estimate the extreme value parameter according to the excess value using the maximum likelihood estimation to obtain the extreme value parameter; The optimization module is used to obtain the reconstruction error of the test set, and the target category in the library corresponding to the minimum reconstruction error is used as the candidate category. According to the candidate category and the obtained extreme value parameters, the optimal parameters are obtained; the optimal parameters are the optimal tail parameters and the optimal threshold; The recognition module is used to identify the drone target based on the optimal parameters.

[0038] The extreme value theory in this application (for details, please refer to the detailed description of extreme value theory below) expands the traditional target recognition method to open set recognition, so that the classifier has the ability to distinguish non-cooperative targets that lack prior information, greatly improving the rejection accuracy of non-cooperative low, slow and small targets.

[0039] In Figure 2(a) and Figure 2(b), SR-OSR is a sparse representation algorithm, WSVM is a Weibull calibration model, and 1-vs-set is a 1-class support vector machine. , where TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative, respectively. ,in , .

[0040] Extreme Value Theory: Assume that the sequence of independent and identically distributed random variables set up The Fisher-Tippett theorem[[i]] states that if there is a constant sequence , so that The asymptotic distribution of is non-degenerate, that is: (1) In the formula, the subscript n is the sequence length.

[0041] Then, G must belong to the Frechet-Pareto, Gumbel or Weibull distribution.

[0042] The above theorem can be expressed by the generalized extreme value distribution (GEV): (2) in is the extreme value index, which depends on the distribution law of the original data, and x is the random variable of the distribution function It can be estimated using the Peaks-Over-threshold (POT) method, which relies on the Pickands-Balkema-de Haan theorem, namely: For sufficiently large threshold t, the distribution of excess values Approximately obeys the generalized Pareto distribution (GPD): (3) In the formula is the extreme value parameter By selecting a threshold value t, we can use the observations exceeding t to estimate the generalized Pareto distribution. in Indicates the subscript corresponding to the jth overflow value in the X sequence, , is the indicator function.

[0043] The corresponding logarithmic maximum likelihood function is: (4) in .like , then the likelihood function is: (5) According to the likelihood function, the parameter estimates can be obtained .

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present application rather than to limit it. Although the present application has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present application can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present application should be included in the scope of protection of the claims of the present application.

Claims

1. A drone open set identification method, characterized in that: include: Step 1: Divide the UAV optical image data into a training set and a test set. In the training phase of the cross-training set, obtain the in-library target data for fitting the extreme value distribution, and divide the training set into a cross-training set and a cross-test set. Step 2: In order to obtain the reconstruction error information of the target data in the database, the cross-training set is used to build an over-complete dictionary, and the sparse representation algorithm is used to obtain the reconstruction error of the cross-test set data; Step 3: Select the tail parameters and the reconstruction error output from step 2, sort each type of reconstruction error, select the tail data and threshold value, calculate the excess value, and estimate the extreme value parameter based on the excess value using maximum likelihood estimation to obtain the extreme value parameter; Step 4: Obtain the reconstruction error of the test set, take the target category in the library corresponding to the minimum reconstruction error as the candidate category, and obtain the optimal parameters based on the candidate category and the obtained extreme value parameters; the optimal parameters are the optimal tail parameter and the optimal threshold; Step 5: Based on the optimal parameters, identify the drone target.

2. The open set identification method of drones according to claim 1, characterized in that: The step 1 comprises: Select one type of drone optical image data from the drone optical image data as the non-cooperative target data outside the database, and select the remaining types of drone optical image data as the target data inside the database; Divide the target data in the database into training sets and test sets in proportion, and put all non-cooperative target data outside the database into the test set; The target data in the database is divided into training set and test set in proportion, and all non-cooperative target data outside the database are included in the test set.

3. The open set identification method of drones according to claim 1, characterized in that: The step 2 comprises: In the training phase of the cross-training set, the reconstruction error information of the optical image data of the cth type of UAV in the target data library is ; represents the cross-training set tr, tr is the cross-training set subscript, the reconstruction error corresponding to the Nth target in the optical image data of the cth type UAV, and T represents transposition; Select the reconstruction error information of all c-th type UAV optical image data to build an over-complete dictionary ; Represents the reconstruction error information of the optical image data of the c-th drone obtained in the cross-training phase; the value range of c is 1 to C, and C is the total number of target categories in the library, which is a positive integer; Use sparse representation algorithm to obtain the reconstruction error of the cross test set , It represents the reconstruction error information obtained from the c-th type of UAV optical image data in the cross test set according to the over-complete dictionary D.

4. The open set identification method of drones according to claim 3, characterized in that: The step 3 comprises: Select tail parameters and reconstruction error , sort the c-th type reconstruction errors in order of size: in, is the reconstruction error information of the target data in the c-th library, is the number of c-th category targets in the cross-test set, Sort the reconstruction errors of the c-th target from small to large, reconstruction error; Select tailing data and threshold , calculate the excess value ; Sort the reconstruction errors of the cth type of targets from small to large. The reconstruction error, For the The reconstruction error, is the threshold value, and its value is The reconstruction error, For the reconstruction error; Using maximum likelihood estimation, the extreme value parameter is estimated by the following formula to obtain the extreme value parameter estimate of the cth class: , : In the formula, For parameters and The likelihood function of and is the extreme value parameter, is the number of c-th category targets in the cross-test set, Represents the likelihood function, i is a positive integer, and its value range is 1 to .

5. The open set identification method of drones according to claim 3, characterized in that: The step 4 comprises: The sparse representation algorithm is used to obtain the reconstruction error of the test set, and the target category in the library corresponding to the minimum reconstruction error is As a candidate category, obtain the target category in the library The corresponding extreme value parameters and , calculate the extreme value parameters and The confidence level of the corresponding extreme value distribution is compared with the set threshold to determine the category to which the test sample in the test set belongs and adjust the tail parameters , repeat steps 3 and 4 until the optimal parameters are obtained; the optimal parameters are the optimal tail parameters and the optimal threshold.

6. The open set identification method of drones according to claim 5, characterized in that: In step 4, the candidate categories are calculated , where Y is the test data, D is the overcomplete dictionary, and C is the total number of target categories in the database. To preserve the sparse coefficient matrix The coefficient elements of the training samples corresponding to category c in the , and the coefficient elements of the remaining categories in the target category in the corresponding library are set to zero, express norm; the sparse coefficient matrix has a corresponding coefficient element for each category in the target category in the library.

7. The open set identification method of drones according to claim 5, characterized in that: The step of comparing the confidence level with a set threshold to determine the category to which the test sample in the test set belongs includes: If the confidence is less than or equal to the set threshold, the category to which the test sample in the test set belongs is considered to be the target category in the library, otherwise it is an out-of-library target.

8. The open set identification method of drones according to claim 1, characterized in that: The step 5 comprises: The reconstruction error of the input unknown target information is calculated, and the target category in the library corresponding to the minimum reconstruction error is taken as the candidate class. The corresponding extreme value distribution confidence is calculated according to the extreme value parameters corresponding to the optimal tail parameters. Finally, the confidence is compared with the optimal threshold to determine whether the UAV target belongs to the target category in the library or the target outside the library.

9. The open set identification method of drones according to claim 1, characterized in that: Before step 1, the method further includes: Collect optical image data of drones taken by sensor equipment.

10. A drone open set identification system, implementing the drone open set identification method according to any one of claims 1 to 9, characterized in that: include: A partitioning module is used to partition the optical image data of the UAV into a training set and a test set. In the training phase of the cross-training set, the target data in the library for fitting the extreme value distribution is obtained, and the training set is partitioned into a cross-training set and a cross-test set. The reconstruction module is used to obtain the reconstruction error information of the target data in the library, use the cross-training set to build an over-complete dictionary, and use the sparse representation algorithm to obtain the reconstruction error of the cross-test set data; A calculation module is used to select the tail parameter and the reconstruction error output by the reconstruction module, sort each type of reconstruction error, select the tail data and the threshold value, calculate the excess value, and estimate the extreme value parameter according to the excess value using the maximum likelihood estimation to obtain the extreme value parameter; The optimization module is used to obtain the reconstruction error of the test set, and the target category in the library corresponding to the minimum reconstruction error is used as the candidate category. According to the candidate category and the obtained extreme value parameters, the optimal parameters are obtained; the optimal parameters are the optimal tail parameters and the optimal threshold; The recognition module is used to identify the drone target based on the optimal parameters.

Images