Sample labeling method and device, electronic equipment and computer readable storage medium

Abstract

Embodiments of the present application disclose a sample labeling method and device, electronic equipment and computer readable storage medium. Compared with the existing automatic target identification method, the method has the ability to evaluate the prediction reliability of unknown new samples in an open environment, can timely discover prediction error samples and unknown new samples, and update the identification model through an efficient labeling update mechanism. Through the present application, the technical problems of low identification reliability and update efficiency of the existing automatic target identification method for incremental samples are solved, and the technical effects of improving the identification reliability of incremental samples and the update efficiency of unknown new samples in an open environment are achieved.

Description

Technical Field

[0001] This invention relates to the field of target recognition technology, and in particular to a sample annotation method, apparatus, electronic device, and computer-readable storage medium. Background Technology

[0002] Automatic target recognition utilizes image information acquired by remote sensing sensors to determine the type, model, and other attributes of targets, and has clear application requirements in military fields such as battlefield reconnaissance and precision strikes. The high recognition accuracy of current state-of-the-art machine learning methods in automatic target recognition systems is predicated on highly supervised training data and a stable deployment environment.

[0003] However, in open and dynamic environments, as sensors continuously acquire new data, the number and types of captured target samples gradually increase. Existing recognition models, due to insufficient prior knowledge, cannot accurately predict unknown new samples. To improve the intelligence level of target recognition systems in open-world applications, it is necessary to reliably identify newly captured targets and update new samples in a timely manner to obtain new recognition capabilities. The above-mentioned open-set incremental learning scenario mainly includes three tasks: (1) identifying new samples (including new classes that have not been trained); (2) labeling new samples that need to be updated; and (3) updating the recognition model.

[0004] The processing of incremental data in open environments mainly involves two aspects: recognition reliability and update efficiency. In machine learning and computer vision, many computational models and sample update frameworks for incremental learning have been proposed. For example, in recognition models, incremental learning methods not only transform traditional batch computation methods into incremental update modes but also apply model improvement methods such as knowledge distillation and gradient preservation. Research on evaluating distribution differences for the problem of identifying unknown new classes has received widespread attention, leading to the proposal of open-set recognition methods including probability decay models and anomaly detection. Regarding update mechanisms, priority should be given to labeling and updating new samples with high recognition uncertainty. For closed-set classification models, many metrics have been proposed to evaluate the uncertainty of new sample recognition, such as prediction error size, model confidence output, and the distance from the sample to the classification surface. However, because closed-set models only consider using existing class data to divide an infinite data space, they cannot assess the probability that a new sample belongs to an unknown new class, and may lose high-value targets of new types during the update process.

[0005] It is evident that existing automatic target recognition methods suffer from low reliability and update efficiency in recognizing incremental samples in open environments.

[0006] No effective solution has yet been proposed to address the above problems. Summary of the Invention

[0007] This invention provides a sample annotation method, apparatus, electronic device, and computer-readable storage medium to at least solve the technical problems of low reliability and low update efficiency of existing automatic target recognition methods for incremental sample recognition.

[0008] According to one aspect of the present invention, a sample labeling method is provided, comprising: training an inclusion decision model using training samples of a target, and determining a probability decay model by the inclusion decision model for an unknown new sample to be included in the training samples; calculating the probability that a test sample is included by the category of each training sample; and using the label of the training sample corresponding to the maximum probability value as the classification label of the test sample, wherein the inclusion decision model includes a probability decay model corresponding to each training sample; calculating an inclusion consistency value for an unknown new sample, wherein if the inclusion consistency value is greater than or equal to an inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label, and used to update the inclusion decision model; if the inclusion consistency value is less than the inclusion consistency threshold, calculating a spatial consistency value for the unknown new sample, wherein if the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new sample is manually labeled and used to update the inclusion decision model; and if the spatial consistency value is less than the spatial consistency threshold, the unknown new sample is saved to an unlabeled sample library for accumulation.

[0009] Optionally, training an inclusion decision model using training samples of the target, and determining a probability decay model by the inclusion decision model that indicates an unknown new sample is included in the training samples, includes: calculating the distance values ​​from the training samples to samples of other classes, sorting all the distance values, filtering out a predetermined number of distance values, and fitting the parameters κ of the Weibull probability density function using the maximum likelihood estimation method based on the predetermined number of distance values. i and λi i The Weibull probability density function is as follows: The probability decay model for the unknown new sample being included in the training sample is as follows: Ψ i =1-F(||x′-x) i ||;κ i , λ i ), where F is the cumulative distribution function of f, x i Let x be the training sample and x′ be the unknown new sample.

[0010] Optionally, the probability that a test sample is included in the category of each of the training samples is calculated using the following expression: Where l represents the category of the training sample, and Score l() represents the probability that the test sample is included in category l of the training sample, y i The training sample labels are used for this purpose.

[0011] Optionally, the expression used to calculate the inclusion consistency value of the unknown new sample is as follows: Wherein, Score1 is the maximum value of the class inclusion probability of all training samples, Score2 is the maximum value of the class inclusion probability of all training samples other than the predicted class of the unknown new sample, δ1 is the first constant, and δ2 is the second constant.

[0012] Optionally, calculating the spatial consistency value of the unknown new sample includes: representing the spatial consistency value using the k-nearest neighbor radius of the unknown new sample, wherein the expression for the k-nearest neighbor radius is as follows: Where m = 1,...,k,Inclu m Let α represent the inclusion consistency of the m-th nearest neighbor sample of the unknown new sample x′ in the unknown new sample set, where α∈[0,1] is the inclusion constraint coefficient, X represents the unknown new sample set, B represents the closed sphere, and r is the radius of the closed sphere.

[0013] Optionally, the method further includes: determining a suggested value for the spatial consistency threshold, wherein the suggested value for the spatial consistency threshold is the mean of the k-nearest neighbor radii of the training sample set.

[0014] Optionally, the suggested value of the spatial consistency threshold is expressed as follows: in, Here, B represents the closed sphere, r is the radius of the closed sphere, X1 represents the training sample set, x′ is the training sample in the training sample set, and m = 1,...,k.

[0015] According to another aspect of the present invention, a sample labeling apparatus is also provided, comprising: a first processing module, configured to train an inclusion decision model using training samples of a target, and determine a probability decay model by which an unknown new sample is included in the training samples using the inclusion decision model, calculate the probability that a test sample is included by the category of each training sample, and use the label of the training sample corresponding to the maximum probability value as the classification label of the test sample, wherein the inclusion decision model includes a probability decay model corresponding to each training sample; a second processing module, configured to calculate an inclusion consistency value of an unknown new sample, wherein if the inclusion consistency value is greater than or equal to an inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label, and the inclusion decision model is updated accordingly; a third processing module, configured to calculate a spatial consistency value of an unknown new sample if the inclusion consistency value is less than the inclusion consistency threshold, wherein if the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new sample is manually labeled and the inclusion decision model is updated accordingly; and a fourth processing module, configured to save the unknown new sample to an unlabeled sample library for accumulation if the spatial consistency value is less than the spatial consistency threshold.

[0016] According to another aspect of the present invention, an electronic device is also provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to perform the sample labeling method described in any one of the above embodiments.

[0017] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the sample annotation method described in any one of the above embodiments.

[0018] In this embodiment of the invention, a decision-making model is trained using training samples of the target, and the probability decay model for an unknown new sample being included in the training samples is determined by the decision-making model. The probability that a test sample is included in the category of each training sample is calculated, and the maximum probability value is used as the classification label of the test sample. The decision-making model includes a probability decay model corresponding to each training sample. The inclusion consistency value of the unknown new sample is calculated. If the inclusion consistency value is greater than or equal to the inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label, and this is used to update the inclusion decision-making model. If the inclusion consistency value is less than the inclusion consistency threshold, the spatial consistency value of the unknown new sample is calculated. If the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new sample is manually labeled and used to update the inclusion decision-making model. If the spatial consistency value is less than the spatial consistency threshold, the unknown new sample is saved to an unlabeled sample library for accumulation. In other words, compared with existing automatic target recognition methods, the embodiments of the present invention have the ability to evaluate the prediction reliability of unknown new samples in open environments, can promptly detect erroneous samples and unknown new samples, and update the recognition model through an efficient annotation and update mechanism. This solves the technical problem of low reliability and update efficiency of existing automatic target recognition methods for incremental samples, and achieves the technical effect of improving the recognition reliability of incremental samples and the update efficiency of unknown new samples in open environments. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0020] Figure 1 A flowchart of the sample annotation method provided in the embodiments of the present invention;

[0021] Figure 2 A flowchart illustrating the incremental update mechanism provided in an optional embodiment of the present invention;

[0022] Figure 3(a) is a schematic diagram of optical photographs of BMP2, BTR70, T72, BTR60 and 2S1 provided in optional embodiments of the present invention;

[0023] Figure 3(b) is a schematic diagram of optical photographs of BRDM2, D7, T62, ZIL131 and ZSU23 / 4 provided in optional embodiments of the present invention;

[0024] Figure 3(c) is a schematic diagram of SAR images of BMP2, BTR70, T72, BTR60 and 2S1 provided in an optional embodiment of the present invention;

[0025] Figure 3(d) is a schematic diagram of SAR images of BRDM2, D7, T62, ZIL131 and ZSU23 / 4 provided in the optional embodiments of the present invention;

[0026] Figure 4(a) is a schematic diagram showing the change in recognition rate as the number of labeled training samples increases, according to an optional embodiment of the present invention;

[0027] Figure 4(b) is a magnified schematic diagram showing the change in recognition rate as the number of labeled training samples increases, according to an optional embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of a sample labeling device provided in an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this invention are used to distinguish different objects, rather than to limit a specific order.

[0031] According to one aspect of the present invention, a sample annotation method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0032] Figure 1 A flowchart of the sample annotation method provided in the embodiments of the present invention is shown below. Figure 1 As shown, the method includes the following steps:

[0033] Step S102: Train the inclusion decision model using the training samples of the target, and determine the probability decay model of the unknown new sample being included in the training samples by the inclusion decision model. Calculate the probability that the test sample is included in the category of each training sample, and use the training sample label corresponding to the maximum probability value as the classification label of the test sample. The inclusion decision model includes a probability decay model corresponding to each training sample.

[0034] Step S104: Calculate the inclusion consistency value of the unknown new sample. If the inclusion consistency value is greater than or equal to the inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label and used to update the inclusion decision model.

[0035] Step S106: If the inclusion consistency value is less than the inclusion consistency threshold, calculate the spatial consistency value of the unknown new sample. If the spatial consistency value is greater than or equal to the spatial consistency threshold, manually label the unknown new sample and use it to update the inclusion decision model.

[0036] In step S108, if the spatial consistency value is less than the spatial consistency threshold, the unknown new sample is saved to the unlabeled sample library for accumulation.

[0037] It should be noted that the consistency threshold and spatial consistency threshold mentioned above can be flexibly set according to the needs of the application scenario, and will not be elaborated on here.

[0038] In this embodiment of the invention, a decision-making model is trained using training samples of the target, and the probability decay model for an unknown new sample to be included in the training samples is determined by the inclusion decision model. The probability that a test sample is included in the category of each training sample is calculated, and the label of the training sample corresponding to the maximum probability value is used as the classification label of the test sample. The inclusion decision model includes a probability decay model corresponding to each training sample. The inclusion consistency value of the unknown new sample is calculated. If the inclusion consistency value is greater than or equal to the inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label, and the inclusion decision model is updated. If the inclusion consistency value is less than the inclusion consistency threshold, the spatial consistency value of the unknown new sample is calculated. If the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new sample is manually labeled and used to update the inclusion decision model. If the spatial consistency value is less than the spatial consistency threshold, the unknown new sample is saved to an unlabeled sample library for accumulation. In other words, compared with existing automatic target recognition methods, the embodiments of the present invention have the ability to evaluate the prediction reliability of unknown new samples in open environments, can promptly detect erroneous samples and unknown new samples, and update the recognition model through an efficient annotation and update mechanism. This solves the technical problem of low reliability and update efficiency of existing automatic target recognition methods for incremental samples, and achieves the technical effect of improving the recognition reliability of incremental samples and the update efficiency of unknown new samples in open environments.

[0039] It should be noted that the application scenarios of the above methods include, but are not limited to, automatic target recognition of radar images in open environments. By using a plasticized decision model, a reliable decision-making and efficient update mechanism for unknown new samples can be established during the gradual acquisition of target samples, thereby expanding the scope of new target attributes and categories.

[0040] In one optional implementation, a decision-making model is trained using training samples of the target, and the probability decay model for determining the inclusion of unknown new samples in the training samples is determined by the decision-making model. This includes: calculating the distance values ​​from the training samples to samples of other classes, sorting all the distance values, selecting a predetermined number of distance values, and fitting the parameters κ of the Weibull probability density function using the maximum likelihood estimation method based on the predetermined number of distance values. i and λ i The Weibull probability density function is as follows:

[0041]

[0042] The probability decay model for an unknown new sample being included in the training samples is as follows:

[0043] Ψ i =1-F(||x′-x) i ||;κ i , λ i )

[0044] Where F is the cumulative distribution function of f, x i x is the training sample, and x′ is an unknown new sample.

[0045] It should be noted that the above-mentioned predetermined quantity can be determined based on the quantity of other categories of samples. For example, the above-mentioned predetermined quantity can be one-tenth of the quantity of other categories of samples.

[0046] In one alternative implementation, the probability that a test sample is included in the category of each training sample is calculated using the following expression:

[0047]

[0048] Where l represents the class of the training sample, and Score l () represents the probability that the test sample is contained in class l of the training samples, y i These are the labels for the training samples.

[0049] In one alternative implementation, the expression used to calculate the inclusion consistency value of the unknown new sample is as follows:

[0050]

[0051] Where Score1 is the maximum value of the class inclusion probability of all training samples, Score2 is the maximum value of the class inclusion probability of all training samples except for the predicted class of unknown new samples, δ1 is the first constant, and δ2 is the second constant.

[0052] It should be noted that the values ​​of the first and second constants mentioned above are different.

[0053] In one alternative implementation, calculating the spatial consistency value of the unknown new sample includes: representing the spatial consistency value using the k-nearest neighbor radius of the unknown new sample, wherein the expression for the k-nearest neighbor radius is as follows:

[0054]

[0055] Where m = 1,...,k,Inclu m Let represent the inclusion consistency of the m-th nearest neighbor sample in the unknown new sample set, α∈[0,1] is the inclusion constraint coefficient, X represents the unknown new sample set, B represents the closed sphere, r is the radius of the closed sphere, and ρ ～ Let k represent the radius of the nearest neighbor.

[0056] In an alternative implementation, the method further includes: determining a suggested value for a spatial consistency threshold, wherein the suggested value for the spatial consistency threshold is the mean of the k-nearest neighbor radii of the training sample set.

[0057] In an alternative implementation, the suggested value of the aforementioned spatial consistency threshold is expressed as follows:

[0058]

[0059] in, Here are the suggested values ​​for the spatial consistency threshold, where B represents the closed sphere, r is the radius of the closed sphere, X1 represents the training sample set, x′1 is the training sample in the training sample set, and m = 1,...,k.

[0060] The following is a detailed description of an incremental learning method based on model inclusive expansion provided by an optional embodiment of the present invention.

[0061] Figure 2 A flowchart illustrating the incremental update mechanism provided in an optional embodiment of the present invention is shown below. Figure 2As shown, the basic incremental recognition model, which includes a decision model, is an optional embodiment of the present invention and is updated with new samples. The output index generated by this model calculates the consistency score for unknown new samples. Samples with high scores are automatically labeled with predicted labels, while those with low scores correspond to misclassified samples and unknown new class samples. After evaluation through spatial consistency, samples with high spatial consistency are manually labeled, while those with low spatial consistency, corresponding to unknown isolated distribution samples, are sent to an unlabeled sample library for accumulation.

[0062] Furthermore, the specific implementation steps of the incremental learning method based on model inclusive expansion provided by this invention are as follows:

[0063] Step 1: Train the inclusion decision model using the initial training samples. The inclusion decision model is integrated from a probability decay model corresponding to each training sample. The modeling process for each training sample after feature extraction is as follows.

[0064] For training sample x i , tag y i Calculate the distance values ​​from the sample to other classes, sort these distance values, and take the smallest τ distance values. τ is typically one-tenth the number of samples from other classes. Using these τ distance values ​​as a basis, fit the parameters κ of the Weibull probability density function using the maximum likelihood estimation method. i and λ i The Weibull probability density function is as follows:

[0065]

[0066] Then the unknown new sample x′ is contained in x i The probability decay model is as follows:

[0067] Ψ i =1-F(||x′-x) i ||;κ i , λ i )

[0068] Where F is the cumulative distribution function of f. The probability that a test sample belongs to class l of the training samples is expressed as follows:

[0069]

[0070] During testing, the probability that a test sample is included in the category of each training sample is calculated, and the category label corresponding to the maximum probability value is the classification label of the test sample.

[0071] Step 2: Calculate the inclusion consistency value for the unknown new sample. Assume l p For the predicted category of an unknown new sample x′, the Inclusion Consistency calculation formula is as follows:

[0072]

[0073]

[0074]

[0075]

[0076] Here, Score1 represents the maximum class inclusion probability of all training samples, and Score2 represents the maximum class inclusion probability of all training samples other than the predicted class. δ1 and δ2 are constants. δ1 can be set to 0.001 to set the inclusion consistency of new samples with low inclusion probabilities to 0, and δ2 can be set to 0.01 to avoid unreasonable inclusion consistency calculations caused by large differences in the orders of magnitude between Score1 and Score2. In practice, samples with higher inclusion consistency values ​​correspond to samples with higher recognition reliability, and it is recommended to directly label and update them using classification labels.

[0077] Step 3: The spatial consistency value of the unknown new sample cannot be calculated directly. Calculate the k-nearest neighbor radius ρ of the unknown new sample. ～ Indirectly representing the spatial consistency value, ρ ～ The larger the value, the smaller the spatial consistency value. The calculation formula is as follows:

[0078]

[0079] Where m = 1,...,k,Inclu m Let represent the inclusion consistency value of x′ in the m-th nearest neighbor sample of the unknown new sample set, α∈[0,1] is the inclusion constraint coefficient, and X represents the unknown new sample set. B represents the closed sphere, and its expression is as follows:

[0080] B(z, r)={z′:|zz′|≤r}

[0081] To improve generalization and update performance when manually labeling new samples, a randomization operation is introduced, which involves randomly selecting new samples with high spatial consistency for manual labeling. The proposed value for the spatial consistency threshold can be derived from the k-nearest neighbor radius ρ of the training sample set. t ~Given, for example, the mean of the k nearest neighbor radii of the training sample, ρ t The calculation formula is as follows:

[0082]

[0083] Thus, we have obtained the predicted labels for the unknown new samples, including both inclusion consistency and spatial consistency. Since inclusion consistency constrains spatial consistency, in practical applications, the two consistency metrics can be used separately. For applications requiring an assessment of recognition reliability, inclusion consistency is used. When selecting new samples for manual annotation updates, spatial consistency is used to select unknown samples.

[0084] In summary, compared with existing automatic target recognition methods, this invention has the ability to evaluate the reliability of predicting unknown new samples in open environments, can promptly detect mispredicted samples and unknown new samples, and can update the recognition model through an efficient annotation update mechanism.

[0085] Taking MSTAR ten-category target data as an example, the optional embodiments of the present invention are described in detail by simulating recognition scenarios and incremental learning scenarios in an open environment.

[0086] The samples used in the experiment were MSTAR ten-class target slices. The training samples were targets with a pitch angle of 17°, and the test samples were targets with a pitch angle of 15°. The number of MSTAR target samples and the slice size used are shown in Table 1. Figure 3(a) is a schematic diagram of the optical photographs of BMP2, BTR70, T72, BTR60 and 2S1 provided in the optional embodiment of the present invention. As shown in Figure 3(a), the optical photographs of BMP2, BTR70, T72, BTR60 and 2S1 are shown from left to right. Figure 3(b) is a schematic diagram of the optical photographs of BRDM2, D7, T62, ZIL131 and ZSU23 / 4 provided in the optional embodiment of the present invention. As shown in Figure 3(b), the optical photographs of BRDM2, D7, T62, ZIL131 and ZSU23 / 4 are shown from left to right. Figure 3(c) is a schematic diagram of SAR images of BMP2, BTR70, T72, BTR60, and 2S1 provided in an optional embodiment of the present invention. As shown in Figure 3(c), the SAR images of BMP2, BTR70, T72, BTR60, and 2S1 are shown from left to right. Figure 3(d) is a schematic diagram of SAR images of BRDM2, D7, T62, ZIL131, and ZSU23 / 4 provided in an optional embodiment of the present invention. As shown in Figure 3(d), the SAR images of BRDM2, D7, T62, ZIL131, and ZSU23 / 4 are shown from left to right. Target slice readout images and corresponding optical photographs are shown in Figures 3(a) and 3(b). Figure 3(c) and 3(d) All slices were cut to 128×128 from the center, and the feature extraction method used was nonnegative matrix factorization.

[0087] Table 1 shows the number of MSTAR target samples and slice sizes used.

[0088]

[0089] First, the recognition scenario in an open environment is simulated, as shown in Table 2. A portion of the training samples from the three target classes are used for training. Inclusion consistency is calculated for all ten test samples to evaluate recognition reliability. All new samples are sorted by inclusion consistency from highest to lowest, and then all sorts are divided into 10 sorting intervals. In each interval, the proportion of correctly classified samples and the proportion of samples classified as unknown new classes are calculated. The remaining, undisplayed intervals represent the proportion of misclassified samples belonging to existing classes. The performance evaluation of recognition reliability after merging duplicate intervals is shown in Table 3. It can be seen that in intervals with high inclusion consistency, most samples are correctly classified, while new class samples and misclassified samples are mostly located in intervals with low inclusion consistency. Thus, the automatic target recognition system can automatically label new samples and identify unknown new classes through inclusion consistency evaluation.

[0090] Table 2. Settings for recognition scenarios in a simulated open environment.

[0091]

[0092] Table 3. Evaluation of the reliability of identification after merging duplicate intervals.

[0093]

[0094] Subsequently, this paper sets up an incremental learning experiment in an open environment to verify the performance of annotation update efficiency. The first 40% of the training samples in the first three classes are used as the initial samples. After training the decision model, the spatial consistency value of the remaining training samples is calculated. Half of the mean k-nearest neighbor radius of the initial samples is set as the reference value for manual annotation. The classification performance of the ten test samples changes as the number of manually annotated samples increases. Figure 4(a) is a schematic diagram of the change in recognition rate as the number of labeled training samples increases, provided by the optional embodiment of the present invention. Figure 4(b) is a schematic diagram of the enlarged partial curve of the change in recognition rate as the number of labeled training samples increases, provided by the optional embodiment of the present invention. The comparison methods are the trust score method and the random selection method. It can be seen that the proposed method IEosIL achieves the best recognition performance.

[0095] According to another aspect of the present invention, a sample labeling device is also provided. Figure 5 This is a schematic diagram of the sample labeling device provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the sample labeling device includes: a first processing module 52, a second processing module 54, a third processing module 56, and a fourth processing module 58. The sample labeling device will now be described in detail.

[0096] The first processing module 52 is used to train an inclusion decision model using the training samples of the target, and to determine the probability decay model of unknown new samples being included in the training samples by the inclusion decision model, calculate the probability that the test sample is included in the category of each training sample, and use the training sample label corresponding to the maximum probability value as the classification label of the test sample. The inclusion decision model includes a probability decay model corresponding to each training sample.

[0097] The second processing module 54 is connected to the first processing module 52 and is used to calculate the inclusion consistency value of the unknown new sample. If the inclusion consistency value is greater than or equal to the inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label and used to update the inclusion decision model.

[0098] The third processing module 56 is connected to the second processing module 54 and is used to calculate the spatial consistency value of the unknown new sample when the included consistency value is less than the included consistency threshold. When the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new sample is manually labeled and used to update the included decision model.

[0099] The fourth processing module 58, connected to the third processing module 56, is used to save unknown new samples to the unlabeled sample library for accumulation when the spatial consistency value is less than the spatial consistency threshold.

[0100] In this embodiment of the invention, the sample labeling device trains an inclusion decision model using training samples of the target, and the inclusion decision model determines a probability decay model for the inclusion of unknown new samples in the training samples. It calculates the probability that a test sample is included in the category of each training sample, and uses the training sample label corresponding to the maximum probability value as the classification label of the test sample. The inclusion decision model includes a probability decay model corresponding to each training sample. The device calculates the inclusion consistency value of unknown new samples. If the inclusion consistency value is greater than or equal to an inclusion consistency threshold, the unknown new samples are labeled and updated using the classification label, which is then used to update the inclusion decision model. If the inclusion consistency value is less than the inclusion consistency threshold, the device calculates the spatial consistency value of unknown new samples. If the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new samples are manually labeled and used to update the inclusion decision model. If the spatial consistency value is less than the spatial consistency threshold, the unknown new samples are saved to an unlabeled sample library for later accumulation. In other words, compared with existing automatic target recognition methods, the embodiments of the present invention have the ability to evaluate the prediction reliability of unknown new samples in open environments, can promptly detect erroneous samples and unknown new samples, and update the recognition model through an efficient annotation and update mechanism. This solves the technical problem of low reliability and update efficiency of existing automatic target recognition methods for incremental samples, and achieves the technical effect of improving the recognition reliability of incremental samples and the update efficiency of unknown new samples in open environments.

[0101] It should be noted that the first processing module 52, the second processing module 54, the third processing module 56 and the fourth processing module 58 mentioned above correspond to steps S102 to S108 in the method embodiment. The examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the content disclosed in the above method embodiment.

[0102] In an optional implementation, the first processing module 52 includes: a first processing subunit, configured to calculate the distance values ​​from training samples to other class samples, sort all the distance values, filter out a predetermined number of distance values, and fit the parameters κ of the Weibull probability density function using the maximum likelihood estimation method based on the predetermined number of distance values. i and λ i The Weibull probability density function is as follows:

[0103]

[0104] The probability decay model for an unknown new sample being included in the training samples is as follows:

[0105] Ψ i =1-F(||x′-x) i ||;κi , λ i )

[0106] Where F is the cumulative distribution function of f, x i x is the training sample, and x′ is an unknown new sample.

[0107] In an optional implementation, the first processing module 52 includes a second processing subunit, used to calculate the probability that a test sample is included in the category of each training sample using the following expression:

[0108]

[0109] Where l represents the class of the training sample, and Score l () represents the probability that the test sample is contained in class l of the training samples, y i These are the labels for the training samples.

[0110] In an optional implementation, the second processing module 54 includes a third processing subunit, used to calculate the consistency value of the unknown new sample using the following expression:

[0111]

[0112] Where Score1 is the maximum value of the class inclusion probability of all training samples, Score2 is the maximum value of the class inclusion probability of all training samples except for the predicted class of unknown new samples, δ1 is the first constant, and δ2 is the second constant.

[0113] In an optional implementation, the third processing module 56 includes a fourth processing subunit, used to represent the spatial consistency value using the k-nearest neighbor radius of the unknown new sample, wherein the expression for the k-nearest neighbor radius is as follows:

[0114]

[0115] Where m = 1,...,k,Inclu m Let α represent the inclusion consistency of the m-th nearest neighbor sample in the unknown new sample set, where α∈[0,1] is the inclusion constraint coefficient, X represents the unknown new sample set, B represents the closed sphere, and r is the radius of the closed sphere.

[0116] In an optional embodiment, the above apparatus further includes: a determining module for determining a proposed value for a spatial consistency threshold, wherein the proposed value for the spatial consistency threshold is the mean of the k-nearest neighbor radii of the training sample set.

[0117] In an alternative implementation, the suggested value of the aforementioned spatial consistency threshold is expressed as follows:

[0118]

[0119] in, Here are the suggested values ​​for the spatial consistency threshold, where B represents the closed sphere, r is the radius of the closed sphere, X1 represents the training sample set, x′1 is the training sample in the training sample set, and m = 1,...,k.

[0120] According to another aspect of the present invention, an electronic device is also provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to perform the sample labeling method of any of the above.

[0121] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the sample annotation method of any of the above.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A sample annotation method, characterized in that, include: The inclusion decision model is trained using training samples of the target, and the inclusion decision model determines the probability decay model of an unknown new sample being included in the training samples. The probability of a test sample being included in the category of each training sample is calculated, and the label of the training sample corresponding to the maximum probability value is used as the classification label of the test sample. The inclusion decision model includes a probability decay model corresponding to each training sample. The training samples are targets with a pitch angle of 17°, and the test samples are targets with a pitch angle of 15°. Both the training samples and the test samples are MSTAR ten-class target slices. Calculate the inclusion consistency value of the unknown new sample. If the inclusion consistency value is greater than or equal to the inclusion consistency threshold, then the unknown new sample is labeled and updated using the classification label, and this is used to update the inclusion decision model. The expression used to calculate the inclusion consistency value of the unknown new sample is as follows: in, The maximum value of the probability for each class in all training samples. The maximum value of the class probability for all training samples, excluding the predicted class of the unknown new sample. It is the first constant. It is the second constant; If the included consistency value is less than the included consistency threshold, then the spatial consistency value of the unknown new sample is calculated. If the spatial consistency value is greater than or equal to the spatial consistency threshold, then the unknown new sample is manually labeled and used to update the included decision model. Calculating the spatial consistency value of the unknown new sample includes: Using the unknown new sample The nearest neighbor radius represents the spatial consistency value, wherein, The expression for the nearest neighbor radius is as follows: in, , Indicates the unknown new sample The first in the unknown new sample set Consistency of inclusion among nearest neighbor samples To include constraint coefficients, This represents the set of unknown new samples. Represents the closed ball. Let be the radius of the closed sphere; If the spatial consistency value is less than the spatial consistency threshold, the unknown new sample is saved to the unlabeled sample library for accumulation.

2. The method according to claim 1, characterized in that, Training a decision-making model using training samples of the target, and determining a probability decay model by the decision-making model that includes unknown new samples in the training samples, includes: Calculate the distance values ​​from the training samples to samples of other classes, sort all the distance values, select a predetermined number of distance values, and use the maximum likelihood estimation method to fit the parameters of the Weibull probability density function based on the predetermined number of distance values. and The Weibull probability density function is as follows: The probability decay model for the unknown new sample being included in the training sample is as follows: in, for The cumulative distribution function, The training samples are as follows. This refers to the unknown new sample.

3. The method according to claim 1, characterized in that, The expression used to calculate the probability that a test sample is included in the category of each of the training samples is as follows: in, The category of the training samples. The categories that the test samples contain within the training samples The probability, The training sample labels are used for this purpose.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Determine a suggested value for the spatial consistency threshold, wherein the suggested value for the spatial consistency threshold is a subset of the training sample set. The mean of the nearest neighbor radii.

5. The method according to claim 4, characterized in that, The suggested expression for the spatial consistency threshold is as follows: in, The suggested value for the spatial consistency threshold is... Represents the closed ball. Let be the radius of the closed sphere. This represents the training sample set. The training samples in the training sample set .

6. A sample labeling device, characterized in that, include: The first processing module is used to train an inclusion decision model using training samples of the target, and to determine a probability decay model by which an unknown new sample is included in the training samples using the inclusion decision model, calculate the probability that a test sample is included in the category of each training sample, and use the label of the training sample corresponding to the maximum probability value as the classification label of the test sample, wherein the inclusion decision model includes a probability decay model corresponding to each training sample. The second processing module is used to calculate the inclusion consistency value of the unknown new sample. If the inclusion consistency value is greater than or equal to the inclusion consistency threshold, the unknown new sample is labeled and updated using the classification label, and the inclusion decision model is updated accordingly. The third processing module is used to calculate the spatial consistency value of the unknown new sample when the included consistency value is less than the included consistency threshold, wherein when the spatial consistency value is greater than or equal to the spatial consistency threshold, the unknown new sample is manually labeled and used to update the included decision model. The fourth processing module is used to save the unknown new sample to the unlabeled sample library for accumulation when the spatial consistency value is less than the spatial consistency threshold.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to execute the sample annotation method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the sample annotation method of any one of claims 1 to 5.

Images