Face recognition method for label data missing of user side

Abstract

The invention provides a face recognition method for label data missing of a user side. The face recognition method comprises the steps that an initial model is selected, and a sample selection strategy is determined. Based on a sample selection strategy, supplementary samples are screened out from the label-free sample set, and a supplementary sample set is obtained. And adding the supplementary sample set into the labeled sample set to obtain a first training set. The first training set comprises labeled samples and unlabeled samples, and the first training set is adopted to carry out mixed training on the initial model to obtain a face recognition model. In the semi-supervised training, firstly, a labeled sample is used for learning a mapping relation between input features and output labels, and then the sample type of an unlabeled sample is predicted. The unlabeled samples comprise a large amount of unstructured or unannotated data, and the method combines the labeled samples and the unlabeled samples, can solve the problem of falling into a local optimal solution due to independent adoption of active learning, accelerates the convergence process of the initial model in the absence of label data, and improves the efficiency and precision of training the initial model.

Description

Technical Field

[0001] The present invention relates to the field of face recognition technology, and in particular to a face recognition method for missing user-side label data. Background Art

[0002] Existing face recognition methods have certain limitations when label data is missing. Traditional face recognition methods include algorithms based on feature extraction and machine learning models. Algorithms based on feature extraction include LBPH algorithm and HOG algorithm, and machine learning models include CNN model. Feature extraction algorithms and machine learning models usually rely on complete label data sets for training. In the absence of label data, the accuracy of face recognition will be greatly reduced.

[0003] Although deep learning models have achieved good results in face recognition, they also rely on a large amount of labeled data. When the labeled data is insufficient, the performance of deep learning models will be significantly affected. In addition, some data completion methods, such as generative adversarial networks and autoencoders, although they can generate missing labeled data, these generated labels are often not accurate enough and may introduce noise, affecting the accuracy of recognition results. Although unsupervised learning and semi-supervised learning methods attempt to learn with less labeled or unlabeled data, these methods do not have high enough accuracy in face recognition in the absence of labels, and generally have poor robustness. In addition, deep learning and semi-supervised learning usually require a lot of computing resources and time, which limits the real-time application of deep learning and semi-supervised learning. In general, the existing technology not only relies on the effect of data completion when the labeled data is missing, but also faces problems such as high computational complexity, low data completion accuracy, and lack of a unified theoretical framework. These problems limit the promotion and application of existing face recognition algorithms in practical applications.

[0004] Therefore, a face recognition method with high computational efficiency is needed that can reduce the amount of data that needs to be labeled, select the most informative samples, and gradually improve the face recognition ability of the face recognition model. Summary of the invention

[0005] In order to overcome the problems existing in the related art, the purpose of the present invention is to provide a face recognition method for missing user-side label data, which can reduce the amount of data that needs to be labeled, select the most informative samples, and gradually improve the face recognition ability of the face recognition model, and has high computational efficiency.

[0006] A face recognition method for missing user-side label data, comprising: Select the initial model and determine the sample selection strategy; Based on the sample selection strategy, supplementary samples are screened out from the unlabeled sample set to obtain a supplementary sample set; Adding the supplementary sample set to the labeled sample set to obtain a first training set; the first training set includes labeled samples and unlabeled samples; Using the first training set to perform hybrid training on the initial model to obtain a face recognition model; wherein the hybrid training includes alternating iterative active learning and semi-supervised training; The face recognition model is used to perform face recognition on the user's face.

[0007] In a preferred technical solution of the present invention, the first training set is used to perform mixed training on the initial model to obtain a face recognition model, including: Using the first training set to perform semi-supervised training on the initial model to generate a first pseudo-label sample set; Labeling some pseudo-label samples in the first pseudo-label sample set to obtain a first updated pseudo-label sample set; Adding the first updated pseudo-label sample set to the first training set to obtain a second training set; The initial model is semi-supervised trained using the second training set to generate a second pseudo-label sample set.

[0008] In a preferred technical solution of the present invention, the selecting of the initial model comprises: A model is selected as an initial model from a neural network model, a decision tree model, a support vector machine model, and a logistic regression model.

[0009] In a preferred technical solution of the present invention, the step of determining the sample selection strategy includes: Select a sampling method from uncertainty sampling, diversity sampling and information entropy sampling as the sample selection strategy.

[0010] In a preferred technical solution of the present invention, the method of selecting supplementary samples from the unlabeled sample set based on the sample selection strategy to obtain the supplementary sample set includes: If the sample selection strategy is uncertainty sampling, then select the uncertainty index; Selecting supplementary samples from the unlabeled sample set according to the uncertainty indicator; All of the supplementary samples are regarded as a supplementary sample set.

[0011] In a preferred technical solution of the present invention, the step of selecting supplementary samples from the unlabeled sample set according to the uncertainty index includes: If the uncertainty indicator is entropy, the unlabeled samples are sorted in descending order according to the size of the entropy to obtain an entropy sequence; The unlabeled samples corresponding to the first entropy to the K1th entropy in the entropy sequence are used as supplementary samples.

[0012] In a preferred technical solution of the present invention, the step of selecting supplementary samples from the unlabeled sample set according to the uncertainty index includes: If the uncertainty indicator is confidence, the unlabeled samples are sorted in ascending order according to the size of the confidence to obtain a confidence sequence; The unlabeled samples corresponding to the first confidence to the K2th confidence in the confidence sequence are used as supplementary samples.

[0013] In a preferred technical solution of the present invention, the method of selecting supplementary samples from the unlabeled sample set based on the sample selection strategy to obtain the supplementary sample set includes: If the sample selection strategy is diversity sampling, clustering or maximizing margins are used to select supplementary samples from the unlabeled sample set; All of the supplementary samples are regarded as a supplementary sample set.

[0014] In a preferred technical solution of the present invention, after obtaining the first updated pseudo-label sample set, the method further includes: If the number of labeled pseudo-label samples is greater than or equal to the labeled sample threshold, or the face recognition accuracy of the initial model has converged during the training process, the training of the initial model is stopped to obtain a face recognition model.

[0015] In a preferred technical solution of the present invention, the semi-supervised training of the initial model using the first training set to generate a first pseudo-label sample set includes: Inputting the labeled samples in the first training set into an initial model so that the initial model learns the mapping relationship between input features and output labels; Inputting the unlabeled samples in the first training set into the initial model, predicting the unlabeled samples, and obtaining a prediction output; Compare the N predicted outputs with the corresponding actual outputs respectively to obtain label matching degrees; The model parameters of the initial model are adjusted according to the label matching degree.

[0016] The beneficial effects of the present invention are: The face recognition method for missing user-side label data provided by the present invention includes selecting an initial model and determining a sample selection strategy. Supplementary samples are screened out from an unlabeled sample set based on the sample selection strategy to obtain a supplementary sample set. The present invention is based on active learning and adopts a sample selection strategy to select the most informative unlabeled samples with high uncertainty of categories. The supplementary sample set is added to the labeled sample set to supplement the labeled sample set to obtain a first training set. The first training set contains both labeled samples and unlabeled samples. The initial model is mixedly trained using the first training set to obtain a face recognition model; wherein the mixed training includes alternating iterative active learning and semi-supervised training. In the semi-supervised training, the mapping relationship between the input features and the output labels is firstly learned using the labeled samples, and then the initial model predicts the unlabeled samples based on the mapping relationship to obtain the sample type of the unlabeled samples. The predicted output label category is compared with the actual output label category, and the model parameters of the initial model are continuously updated according to the comparison result. When the model parameters converge or the number of training times reaches the training times threshold, the face recognition model is obtained. Unlabeled samples contain a large amount of unstructured or unannotated data. By utilizing unlabeled data, semi-supervised learning can improve the accuracy of face recognition of the initial model while saving the time spent on labeling samples. The present invention combines labeled samples and unlabeled samples to solve the problem that active learning alone falls into a local optimal solution, resulting in low efficiency and accuracy of the learning process, as well as the problem that active learning requires a large number of correctly labeled samples. Active learning can focus on those samples that are the most difficult to classify or the most representative, thereby accelerating the convergence process of the initial model and improving the accuracy of the initial model when labeled data is scarce. In terms of semi-supervised learning, the present invention further optimizes the use of unlabeled data and improves the training effect of the model by combining the effective labels obtained from active learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the face recognition method for missing user-side tag data of the present invention; Figure 2 It is a flow chart of the present invention combining active learning and semi-supervised training to perform hybrid training on the initial model. DETAILED DESCRIPTION

[0018] The preferred embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0019] Example 1 like Figure 1 As shown, this embodiment provides a face recognition method for missing user-side label data, including: S1: Select the initial model and determine the sample selection strategy.

[0020] S2: Based on the sample selection strategy, supplementary samples are screened out from the unlabeled sample set to obtain a supplementary sample set.

[0021] S3: Add the supplementary sample set to the labeled sample set to obtain a first training set; the first training set includes labeled samples and unlabeled samples.

[0022] S4: Using the first training set to perform hybrid training on the initial model to obtain a face recognition model; wherein the hybrid training includes alternating iterative active learning and semi-supervised training.

[0023] S5: Performing face recognition on the user's face using the face recognition model.

[0024] The user end of the present invention can be a face recognition attendance machine or a face recognition security monitoring system, which is not limited here.

[0025] The selecting of the initial model comprises: A model is selected as an initial model from a neural network model, a decision tree model, a support vector machine model, and a logistic regression model.

[0026] Determining the sample selection strategy includes: Select a sampling method from uncertainty sampling, diversity sampling and information entropy sampling as the sample selection strategy.

[0027] The initial model may also be a support vector machine model, a k-nearest neighbor model, a naive Bayes model, a random forest model, a linear regression model or a logistic regression model. This embodiment takes the initial model as a neural network model as an example.

[0028] The method of selecting supplementary samples from the unlabeled sample set based on the sample selection strategy to obtain a supplementary sample set includes: S21: If the sample selection strategy is uncertainty sampling, select the uncertainty indicator.

[0029] S22: Filter out supplementary samples from the unlabeled sample set according to the uncertainty indicator.

[0030] S23: All the supplementary samples are taken as a supplementary sample set.

[0031] Uncertainty sampling is to select samples whose classification results are uncertain or wrong under the current model, use some indicators to measure sample uncertainty such as entropy and confidence to sort the samples in ascending or descending order, and then select a part of the samples ranked at the top as the training set. Common methods of uncertainty sampling include minimizing margins, information entropy, and entropy reduction.

[0032] The purpose of diversity sampling is to select samples with different characteristics from the data set as much as possible. If the sample selection strategy is diversity sampling, clustering or maximizing the margin is used to screen out supplementary samples from the unlabeled sample set, and all supplementary samples are used as the supplementary sample set.

[0033] Information entropy sampling: Information entropy sampling is a method based on uncertainty sampling, which selects the most valuable samples by calculating the information entropy of each sample. The classification entropy of each sample under the current model is calculated. The larger the entropy value, the more uncertain the classification result of the sample, and the more necessary it is to add the sample to the labeled sample set.

[0034] The step of selecting supplementary samples from the unlabeled sample set according to the uncertainty indicator includes: S221: If the uncertainty indicator is entropy, the unlabeled samples are sorted in descending order according to the size of the entropy to obtain an entropy sequence.

[0035] S222: Use the unlabeled samples corresponding to the first entropy to the K1th entropy in the entropy sequence as supplementary samples.

[0036] Entropy is a function that describes the state of a system. Entropy is used to measure the degree of disorder in a system. The larger the entropy value, the more uncertain the sample classification result is, and the more it needs to be added to the training set. Therefore, the unlabeled samples are sorted in descending order according to the size of the entropy to obtain the entropy sequence. The higher the entropy in the entropy sequence, the more additional information the unlabeled sample corresponding to the entropy can bring. The unlabeled samples corresponding to the first entropy to the K1th entropy in the entropy sequence are used as supplementary samples.

[0037] Preferably, the value of K1 is set according to the total number of entropies in the entropy sequence, and the total number of entropies in the entropy sequence is multiplied by 20%. If the product is a decimal, it is rounded up. For example, the entropy sequence has a total of 100 entropies, that is, the number of unlabeled samples is 100, and the unlabeled samples corresponding to the first 20 entropies in the entropy sequence are used as supplementary samples.

[0038] The step of selecting supplementary samples from the unlabeled sample set according to the uncertainty indicator includes: If the uncertainty indicator is confidence, the unlabeled samples are sorted in ascending order according to the size of the confidence to obtain a confidence sequence; The unlabeled samples corresponding to the first confidence to the K2th confidence in the confidence sequence are used as supplementary samples.

[0039] The lower the confidence, the greater the additional information the corresponding unlabeled sample can bring, and the more likely it is that the unlabeled sample should be added to the labeled sample set. Therefore, the unlabeled samples are sorted in ascending order by confidence. The unlabeled samples corresponding to the first K2 confidences in the confidence sequence have lower confidences, and the K2 unlabeled samples with lower confidences are used as supplementary samples.

[0040] Preferably, before selecting the supplementary samples from the unlabeled sample set based on the sample selection strategy, it also includes constructing an unlabeled sample set, the unlabeled sample set including real unlabeled samples and simulated unlabeled samples. The simulated unlabeled samples are obtained by performing a nonlinear transformation on the real unlabeled samples, and performing a mean operation on P real unlabeled samples to obtain an average face, where P ≥ 2. The average face is nonlinearly transformed using the following formula: ; Among them, K i For the ith simulated unlabeled sample, AVG i is the ith average face, and MAX is the maximum pixel value that is set. The maximum pixel value can be 255 or other values, which are not limited here.

[0041] By performing nonlinear transformation on the average face, we can highlight the face area with moderate pixel values, such as the face area with an average value of 128, while suppressing the face area with too low pixel values ​​and too high pixel values, so that the simulated unlabeled samples have facial features that the real unlabeled samples do not contain, thereby supplementing the real unlabeled samples. The simulated unlabeled samples are added to the real unlabeled samples to obtain the unlabeled sample set.

[0042] In security monitoring, face recognition technology is often used for automated identity authentication, abnormal behavior detection, and intrusion warning. Due to the large amount of surveillance video data and the high proportion of unlabeled data, traditional supervised learning methods may face the problem of scarce labeled data. At this time, active learning and semi-supervised training can effectively alleviate this problem.

[0043] This embodiment uses a small amount of labeled data, such as facial images of people with known identities, to train the initial model, and then uses active learning to select the most uncertain samples for labeling. When the monitoring system encounters an unseen face, active learning will select the most difficult samples to classify, such as facial images at some specific angles and under specific lighting conditions, and then manually label these high-uncertainty samples. As the amount of monitoring data increases, semi-supervised learning can use a large number of unlabeled facial images to train the initial model without manual labeling. By extracting facial features from unlabeled samples, the recognition effect can be improved and the reliance on a large number of labeled samples can be reduced.

[0044] The face recognition method for missing user-side label data provided in this embodiment includes selecting an initial model and determining a sample selection strategy. Supplementary samples are screened out from the unlabeled sample set based on the sample selection strategy to obtain a supplementary sample set. The present invention is based on active learning and adopts a sample selection strategy to select the most informative unlabeled samples with high uncertainty of the category. The supplementary sample set is added to the labeled sample set to supplement the labeled sample set to obtain a first training set. The first training set contains both labeled samples and unlabeled samples. The initial model is mixedly trained using the first training set to obtain a face recognition model; wherein the mixed training includes alternating iterative active learning and semi-supervised training. In semi-supervised training, the mapping relationship between input features and output labels is first learned using labeled samples, and then the initial model predicts the unlabeled samples based on the mapping relationship to obtain the sample type of the unlabeled samples. The predicted output label category is compared with the actual output label category, and the model parameters of the initial model are continuously updated according to the comparison result. When the model parameters converge or the number of training times reaches the training times threshold, the face recognition model is obtained. Unlabeled samples contain a large amount of unstructured or unannotated data. By utilizing unlabeled data, semi-supervised learning can improve the accuracy of face recognition of the initial model while saving the time spent on labeling samples. The present invention combines labeled samples and unlabeled samples to solve the problem that active learning alone falls into a local optimal solution, resulting in low efficiency and accuracy of the learning process, as well as the problem that active learning requires a large number of correctly labeled samples. Active learning can focus on those samples that are the most difficult to classify or the most representative, thereby accelerating the convergence process of the initial model and improving the accuracy of the initial model when labeled data is scarce. In terms of semi-supervised learning, the present invention further optimizes the use of unlabeled data and improves the training effect of the model by combining the effective labels obtained from active learning.

[0045] Example 2 This embodiment provides a face recognition method for missing user-side tag data. This embodiment only describes the differences from Embodiment 1, such as Figure 2As shown, the first training set is used to perform mixed training on the initial model to obtain a face recognition model, including: S41: Performing semi-supervised training on the initial model using the first training set to generate a first pseudo-label sample set.

[0046] S42: Labeling some pseudo-label samples in the first pseudo-label sample set to obtain a first updated pseudo-label sample set.

[0047] S43: Add the first updated pseudo-label sample set to the first training set to obtain a second training set.

[0048] S44: Using the second training set to perform semi-supervised training on the initial model to generate a second pseudo-label sample set.

[0049] In traditional supervised learning, the algorithm uses only labeled data with known outputs or labels. However, in many real-world scenarios, labeled data may be scarce or expensive to obtain, so only a large amount of unlabeled data can be used. Semi-supervised learning is a type of machine learning that uses both labeled and unlabeled data to train a neural network model.

[0050] In the active learning stage, a small number of supplementary samples in the first training set are labeled. In the semi-supervised learning stage, the labeled samples are first used to learn the mapping relationship between input features and output labels, and then the unlabeled samples are predicted based on the mapping relationship to generate pseudo-labeled samples. The labels of the pseudo-labeled samples are compared with the actual labels of the pseudo-labeled samples, and the model parameters of the initial model are adjusted according to the comparison results to optimize the performance of the initial model.

[0051] A first pseudo-label sample set is generated by using semi-supervised learning, and the first pseudo-label sample set includes a plurality of pseudo-label samples. The pseudo-label samples in the first pseudo-label sample set are annotated by active learning to obtain a first updated pseudo-label sample set.

[0052] After obtaining the first updated pseudo-label sample set, the method further includes: If the number of labeled pseudo-label samples is greater than or equal to the labeled sample threshold, or the face recognition accuracy of the initial model has converged during the training process, active learning is stopped.

[0053] If the number of annotated pseudo-label samples is less than the annotated sample threshold and the face recognition accuracy of the initial model has not converged, continue active learning. Execute steps S43-S44. After step S44, annotate some pseudo-label samples in the second pseudo-label sample set to obtain a second updated pseudo-label sample set. Add the second updated pseudo-label sample set to the second training set to obtain a third training set. Use the third training set to perform semi-supervised training on the initial model to generate a third pseudo-label sample set. Repeat the above process to continuously generate pseudo-label sample sets and training sets until the number of annotated pseudo-label samples is greater than or equal to the annotated sample threshold, or the face recognition accuracy of the initial model converges, stop training the initial model, and obtain a face recognition model.

[0054] The step of using the first training set to perform semi-supervised training on the initial model to generate a first pseudo-label sample set includes: S411: Input the labeled samples in the first training set into an initial model, so that the initial model learns the mapping relationship between input features and output labels.

[0055] S412: Input the unlabeled samples in the first training set into the initial model, predict the unlabeled samples, and obtain prediction output.

[0056] S413: Compare the N predicted outputs with the corresponding actual outputs respectively to obtain label matching degrees.

[0057] S414: Adjusting model parameters of the initial model according to the label matching degree.

[0058] In the process of semi-supervised learning, if the pseudo-labels predicted in the initial stage are inaccurate, it may affect the subsequent training and cause the model performance to decline. In addition, the generation of pseudo-labels often lacks an effective quality control mechanism, which easily leads to the transfer of erroneous labeled data to the next round of training, ultimately affecting the recognition accuracy. Therefore, the present invention adopts an iterative active learning and semi-supervised training fusion strategy to dynamically adjust the sample selection strategy while continuously supplementing pseudo-labels, thereby gradually improving the generalization ability and robustness of the model.

[0059] The lower the label matching degree, the more times the model parameters are adjusted, and the larger the amplitude of each adjustment of the model parameters. The higher the label matching degree, the fewer times the model parameters are adjusted, and the smaller the amplitude of each adjustment of the model parameters. During the training process of the initial model, if the face recognition accuracy of the initial model has converged, the training of the initial model is stopped to obtain the face recognition model.

[0060] This embodiment uses the first training set to perform semi-supervised training on the initial model to generate a first pseudo-label sample set, including inputting the labeled samples in the first training set into the initial model so that the initial model learns the mapping relationship between the input features and the output labels. The unlabeled samples in the first training set are input into the initial model, and the unlabeled samples are predicted to obtain the predicted output. The N predicted outputs are compared with the corresponding actual outputs to obtain the label matching degree, and the model parameters of the initial model are adjusted according to the label matching degree. The present invention adopts an iterative active learning and semi-supervised training fusion strategy, which can dynamically adjust the sample selection strategy while continuously supplementing pseudo-labels, thereby gradually improving the generalization performance and robustness of the model. Compared with the traditional single method, the present invention can more effectively deal with the problem of insufficient labeled samples, especially in scenarios with large amounts of data or high annotation costs, and shows good application potential. In addition, the framework proposed in the present invention can continuously update the initial model so that the initial model can adapt to different types of data.

[0061] Example 3 A computer device is also provided in an embodiment of the present application, which may be a server, wherein the computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection.

[0062] This embodiment also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the face recognition method for missing user-side tag data described in any one of Embodiments 1 to 2 is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0063] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0064] The above description is only a preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A face recognition method for missing user-side label data, characterized in that: include: Select the initial model and determine the sample selection strategy; Based on the sample selection strategy, supplementary samples are screened out from the unlabeled sample set to obtain a supplementary sample set; Adding the supplementary sample set to the labeled sample set to obtain a first training set; the first training set includes labeled samples and unlabeled samples; Using the first training set to perform hybrid training on the initial model to obtain a face recognition model; wherein the hybrid training includes alternating iterative active learning and semi-supervised training; The face recognition model is used to perform face recognition on the user's face.

2. The face recognition method for missing user-side tag data according to claim 1, characterized in that: The step of using the first training set to perform mixed training on the initial model to obtain a face recognition model includes: Using the first training set to perform semi-supervised training on the initial model to generate a first pseudo-label sample set; Labeling some pseudo-label samples in the first pseudo-label sample set to obtain a first updated pseudo-label sample set; Adding the first updated pseudo-label sample set to the first training set to obtain a second training set; The initial model is semi-supervised trained using the second training set to generate a second pseudo-label sample set.

3. The face recognition method for missing user-side tag data according to claim 1, characterized in that: The selecting of the initial model comprises: A model is selected as an initial model from a neural network model, a decision tree model, a support vector machine model, and a logistic regression model.

4. The face recognition method for missing user-side tag data according to claim 1, characterized in that: Determining the sample selection strategy includes: Select a sampling method from uncertainty sampling, diversity sampling and information entropy sampling as the sample selection strategy.

5. The face recognition method for missing user-side tag data according to claim 1, characterized in that: The method of selecting supplementary samples from the unlabeled sample set based on the sample selection strategy to obtain a supplementary sample set includes: If the sample selection strategy is uncertainty sampling, then select the uncertainty index; Selecting supplementary samples from the unlabeled sample set according to the uncertainty indicator; All of the supplementary samples are regarded as a supplementary sample set.

6. The face recognition method for missing user-side tag data according to claim 5, characterized in that: The step of selecting supplementary samples from the unlabeled sample set according to the uncertainty indicator includes: If the uncertainty indicator is entropy, the unlabeled samples are sorted in descending order according to the size of the entropy to obtain an entropy sequence; The unlabeled samples corresponding to the first entropy to the K1th entropy in the entropy sequence are used as supplementary samples.

7. The face recognition method for missing user-side tag data according to claim 5, characterized in that: The step of selecting supplementary samples from the unlabeled sample set according to the uncertainty indicator includes: If the uncertainty indicator is confidence, the unlabeled samples are sorted in ascending order according to the size of the confidence to obtain a confidence sequence; The unlabeled samples corresponding to the first confidence to the K2th confidence in the confidence sequence are used as supplementary samples.

8. The face recognition method for missing user-side tag data according to claim 1, characterized in that: The method of selecting supplementary samples from the unlabeled sample set based on the sample selection strategy to obtain a supplementary sample set includes: If the sample selection strategy is diversity sampling, clustering or maximizing margins are used to select supplementary samples from the unlabeled sample set; All of the supplementary samples are regarded as a supplementary sample set.

9. The face recognition method for missing user-side tag data according to claim 2, characterized in that: After obtaining the first updated pseudo-label sample set, the method further includes: If the number of labeled pseudo-label samples is greater than or equal to the labeled sample threshold, or the face recognition accuracy of the initial model has converged during the training process, the training of the initial model is stopped to obtain a face recognition model.

10. The face recognition method for missing user-side tag data according to claim 2, characterized in that: The step of using the first training set to perform semi-supervised training on the initial model to generate a first pseudo-label sample set includes: Inputting the labeled samples in the first training set into an initial model so that the initial model learns the mapping relationship between input features and output labels; Inputting the unlabeled samples in the first training set into the initial model, predicting the unlabeled samples, and obtaining a prediction output; Compare the N predicted outputs with the corresponding actual outputs respectively to obtain label matching degrees; The model parameters of the initial model are adjusted according to the label matching degree.

Images