Training method and device of palm recognition model and access control recognition method and device

Abstract

The invention discloses a palm recognition model training method and device and an access control recognition method and device, and the training method comprises the steps: extracting student features from batch samples through a pre-trained student model, extracting teacher features from the batch samples through a teacher model, and carrying out the multi-dimensional analysis of the student features and the teacher features, and calculating a center core alignment multi-dimensional fusion loss value, and calculating a total loss value based on the total loss function integrated with the center core alignment multi-dimensional fusion loss value so as to update model parameters of the student model until the student model converges to obtain a palm recognition model, and setting the learning rate to be 10 <-4 >. Therefore, inter-feature multi-dimensional analysis including center core alignment similarity analysis is performed on the student features and the teacher features, so that the total loss value can consider multi-dimensional factors including center core alignment, the generalization ability of the model is further improved, the error rate of palm feature recognition of the palm recognition model is further reduced, and the recognition efficiency of the palm recognition model is improved. The recognition precision of the palm recognition model is improved.

Description

Technical Field

[0001] This application relates to the field of security technology, and more specifically, to a training method and apparatus for a palm recognition model, and an access control recognition method and apparatus. Background Technology

[0002] Palm feature recognition has become a core identity verification solution for security scenarios such as attendance and access control due to its security and convenience. Existing technologies mostly adopt a teacher-student model distillation training mode, attempting to transfer the recognition capabilities of complex models to lightweight models. However, as the size of the registration database and the number of IDs increase, the problem of insufficient feature alignment accuracy becomes prominent, leading to an increase in false positive and false negative rates, which restricts the reliability of applications in large-scale scenarios.

[0003] Related technologies introduce Centered Kernel Alignment (CKA) loss to optimize feature similarity measurement. However, using only CKA to measure the similarity between features is not robust enough to scale transformation and local noise, and it is also less sensitive to prediction errors, limiting its generalization ability and making it difficult to further reduce the error rate of palm feature recognition.

[0004] How to train a palm recognition model to further reduce the error rate of palm feature recognition and improve the recognition accuracy is an issue that needs attention. Summary of the Invention

[0005] In view of the above problems, this application provides a training method and apparatus for a palm recognition model, and an access control recognition method and apparatus, so as to further reduce the error rate of palm feature recognition by the palm recognition model and improve the recognition accuracy of the palm recognition model.

[0006] To achieve the above objectives, the following specific solutions are proposed:

[0007] A method for training a palm recognition model, comprising:

[0008] Using a pre-trained palm recognition basic model as a student model, student features are extracted from a training batch of samples containing multiple palm images using the student model, and teacher features are extracted from the training batch of samples using a teacher model used to guide the training of the student model.

[0009] By performing multidimensional feature analysis between the student features and the teacher features, the center kernel alignment multidimensional fusion loss value between the student features and the teacher features is calculated;

[0010] Based on the total loss function that incorporates the center kernel aligned multidimensional fusion loss value, the total loss value of the training batch samples under a single round of training is calculated.

[0011] The model parameters of the student model are updated using the total loss value until the student model converges, at which point the student model is used as the hand recognition model.

[0012] Optionally, the step of performing multidimensional feature analysis on the student features and the teacher features to calculate the center kernel alignment multidimensional fusion loss value between the student features and the teacher features includes:

[0013] By performing a relative independence analysis on the student features and the teacher features, a center kernel alignment metric between the teacher features and the student features is calculated.

[0014] Based on the center kernel alignment metric, a center kernel alignment similarity fusion value is calculated by performing feature similarity analysis on the student features and the teacher features.

[0015] Based on the center kernel alignment metric, the center kernel alignment regularization value is calculated by performing a prediction error sensitivity analysis on the student features and the teacher features.

[0016] The center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features.

[0017] Optionally, the step of calculating the center kernel alignment metric between the teacher features and the student features by performing a relative independence analysis on the student features and the teacher features includes:

[0018] Calculate the first kernel feature of the teacher characteristics and the second kernel feature of the student characteristics;

[0019] Calculate the first centered kernel feature of the first kernel feature and the second centered kernel feature of the second kernel feature;

[0020] Based on the first centralized kernel feature and the second centralized kernel feature, calculate the independence index between the teacher feature and the student feature;

[0021] The independence index is normalized to obtain the center kernel alignment metric between the teacher feature and the student feature.

[0022] Optionally, based on the center kernel alignment metric, a center kernel alignment similarity fusion value is calculated by performing feature similarity analysis on the student features and the teacher features, including:

[0023] Calculate the similarity measure between the teacher characteristics and the student characteristics;

[0024] The center kernel alignment similarity fusion value is obtained by weighted summation of the center kernel alignment metric and the similarity metric.

[0025] Optionally, based on the center kernel alignment metric, a center kernel alignment regularization value is calculated by performing a prediction error sensitivity analysis on the student features and the teacher features, including:

[0026] Calculate the mean square error between the teacher characteristics and the student characteristics;

[0027] The mean square error is used to regularize the center kernel alignment metric to obtain the center kernel alignment regularization value.

[0028] Optionally, the center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features, including:

[0029] Using the first balancing hyperparameter as the weighting coefficient of the center kernel alignment similarity fusion value and the second balancing hyperparameter as the weighting coefficient of the center kernel alignment regularization value, the center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features, wherein the sum of the first balancing hyperparameter and the second balancing hyperparameter is 1.

[0030] Optionally, the learning rate of the student model during training is 10. -4 .

[0031] An access control identification method includes:

[0032] Obtain the image of the user's palm to be identified, input the image of the palm to be identified into the palm recognition model as described above, and output the user's palm feature vector;

[0033] The similarity between the palm feature vector and the standard feature vectors of each user stored in the access control database is calculated to obtain the feature comparison similarity.

[0034] If the similarity of the features exceeds the similarity comparison threshold, the access control device will be activated.

[0035] A training device for a palm recognition model, comprising:

[0036] The feature extraction unit is used to extract student features from a training batch of samples containing multiple palm images using a pre-trained palm recognition basic model as a student model, and to extract teacher features from the training batch of samples using a teacher model used to guide the training of the student model.

[0037] The center kernel aligned multidimensional fusion loss value calculation unit is used to calculate the center kernel aligned multidimensional fusion loss value between the student features and the teacher features by performing multidimensional analysis between the features;

[0038] The total loss calculation unit is used to calculate the total loss value of the training batch samples under a single round of training based on the total loss function that incorporates the center kernel aligned multidimensional fusion loss value;

[0039] The model parameter update unit is used to update the model parameters of the student model using the total loss value until the student model converges, and then uses the student model as a palm recognition model.

[0040] Optionally, the central kernel aligned multidimensional fusion loss value calculation unit includes:

[0041] The relative independence analysis unit is used to calculate the center kernel alignment metric between the teacher features and the student features by performing relative independence analysis on the student features and the teacher features;

[0042] The feature similarity analysis unit is used to calculate the center kernel alignment similarity fusion value by performing feature similarity analysis on the student features and the teacher features based on the center kernel alignment metric value.

[0043] The prediction error sensitivity analysis unit is used to calculate the center kernel alignment regularization value by performing prediction error sensitivity analysis on the student features and the teacher features based on the center kernel alignment metric value.

[0044] The multi-factor weighted fusion unit is used to weight and fuse the center kernel alignment similarity fusion value and the center kernel alignment regularization value to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features.

[0045] Optionally, the relative independence analysis unit includes:

[0046] A kernel feature calculation unit is used to calculate the first kernel feature of the teacher's features and the second kernel feature of the student's features;

[0047] A centralized kernel feature calculation unit is used to calculate the first centralized kernel feature of the first kernel feature and the second centralized kernel feature of the second kernel feature;

[0048] An independence index calculation unit is used to calculate the independence index between the teacher feature and the student feature based on the first centralized kernel feature and the second centralized kernel feature;

[0049] The normalization processing unit is used to normalize the independence index to obtain the center kernel alignment metric between the teacher feature and the student feature.

[0050] Optionally, the feature similarity analysis unit includes:

[0051] The first feature similarity analysis subunit is used to calculate the similarity measure between the teacher features and the student features;

[0052] The second feature similarity analysis subunit is used to obtain a center kernel alignment similarity fusion value by weighted summation of the center kernel alignment metric and the similarity metric.

[0053] Optionally, the prediction error sensitivity analysis unit includes:

[0054] The first prediction error sensitivity analysis subunit is used to calculate the mean square error between the teacher characteristics and the student characteristics;

[0055] The second prediction error sensitivity analysis subunit is used to regularize the center kernel alignment metric using the mean square error to obtain a center kernel alignment regularization value.

[0056] Optionally, the multi-factor weighted fusion unit includes:

[0057] A multi-factor weighted fusion subunit is used to use a first balancing hyperparameter as the weighting coefficient of the center kernel alignment similarity fusion value and a second balancing hyperparameter as the weighting coefficient of the center kernel alignment regularization value to weightedly fuse the center kernel alignment similarity fusion value and the center kernel alignment regularization value to obtain the center kernel alignment multidimensional fusion loss value between the student feature and the teacher feature, wherein the sum of the first balancing hyperparameter and the second balancing hyperparameter is 1.

[0058] An access control identification device includes:

[0059] The user palm acquisition unit is used to acquire a user's palm image to be identified, input the palm image to be identified into the palm recognition model as described above, and output the user's palm feature vector.

[0060] The feature comparison similarity calculation unit is used to calculate the similarity between the palm feature vector and the standard feature vectors of each user stored in the access control database to obtain the feature comparison similarity.

[0061] The access control opening unit is used to open the access control device if the similarity of the features exceeds a similarity comparison threshold.

[0062] By employing the aforementioned technical solution, this application uses a pre-trained palm recognition basic model as the student model. The student model extracts student features from a training batch of samples containing multiple palm images. A teacher model, used to guide the student model's training, extracts teacher features from the training batch. Through multidimensional feature analysis of the student and teacher features, a center-kernel alignment multidimensional fusion loss value is calculated between them. Based on the total loss function incorporating the center-kernel alignment multidimensional fusion loss value, the total loss value for the training batch of samples in a single training round is calculated. The total loss value is used to update the student model's parameters until the student model converges, at which point it is used as the palm recognition model. Therefore, performing multidimensional feature analysis, including center-kernel alignment similarity analysis, on the student and teacher features ensures that the total loss value takes into account multidimensional factors including center-kernel alignment, further improving the model's generalization ability, reducing the error rate of palm feature recognition, and enhancing the recognition accuracy of the palm recognition model. Attached Figure Description

[0063] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0064] Figure 1 This is a schematic diagram of a process for training a palm recognition model provided in an embodiment of this application;

[0065] Figure 2 This is a flowchart illustrating a multidimensional analysis of student and teacher characteristics provided in an embodiment of this application.

[0066] Figure 3 This is a schematic diagram of a process for implementing access control identification provided in an embodiment of this application;

[0067] Figure 4 A schematic diagram of a device structure for training a palm recognition model provided in an embodiment of this application;

[0068] Figure 5 This is a schematic diagram of a device structure for implementing access control identification, provided in an embodiment of this application. Detailed Implementation

[0069] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0070] The proposed solution can be implemented based on a terminal with data processing capabilities, such as a computer, cloud, or server.

[0071] Next, combined Figure 1 The training method for the palm recognition model of this application may include the following steps:

[0072] Step S110: Using the pre-trained palm recognition basic model as the student model, extract student features from the training batch samples containing multiple palm images using the student model, and extract teacher features from the training batch samples using the teacher model used to guide the training of the student model.

[0073] It should be noted that the hand images in this embodiment are legally authorized or obtained with the consent of the parties involved. Multiple hand images from the training batch samples can also be obtained from existing publicly available hand datasets for training.

[0074] Specifically, during training, the teacher model's parameters are not updated; only the student model's parameters are updated. Student features can be represented as... Given a shape [B, D1], the teacher's characteristics can be represented as follows: The shape is [B, D2]. Here, B is the number of samples in the training batch, i.e., the batch size, D1 represents the dimension of student features, and D2 represents the dimension of teacher features. D1 can be equal to or not equal to D2.

[0075] Step S120: By performing multidimensional analysis between student features and teacher features, calculate the center kernel alignment multidimensional fusion loss value between student features and teacher features.

[0076] Specifically, relative independence analysis, feature similarity analysis, and prediction error sensitivity analysis can be performed on student and teacher features. It is understandable that by conducting these analyses, the center kernel aligned multidimensional fusion loss value can fully integrate the results of these analyses.

[0077] Step S130: Calculate the total loss value of the training batch samples under a single round of training based on the total loss function of the fused multidimensional fusion loss value with center kernel alignment.

[0078] Specifically, the total loss function can be expressed as:

[0079]

[0080] in, This represents the total loss value. This is the category loss or Arcface loss, specifically the loss calculated based on student features and labels. For the center kernel aligned multidimensional fusion loss value, For the loss fusion weights of class loss or Arcface loss, The loss fusion weight is used to align the multidimensional fusion loss value with the center kernel.

[0081] Step S140: Update the model parameters of the student model using the total loss value until the student model converges, and then use the student model as the hand recognition model.

[0082] Specifically, the learning rate of the student model during training can be set to 10. -4 The training rounds are less than 5, thus slowing down the forgetting of knowledge learned from old tasks by the pre-trained student model.

[0083] Understandably, during training, the improved center kernel alignment loss of two features in a batch is calculated to guide the training of the student model and update the network parameters of the student model. The training rounds are only a few, which takes less time and can quickly bring the parameters of the student model to the optimal level.

[0084] When using only CKA to measure the similarity between features, the palm recognition model has an equal error rate (EER) of 0.2162% and a false acceptance rate (FAR) of around 10. -3 The false rejection rate (FRR) was 0.2333% at FRR=10. -4 The FRR was 0.4038% at FAR=10; -5 The FRR was 0.6420% at FAR=10; -6 The FRR was 0.8704% when FAR < 10. -6 The FRR is 1.5104%.

[0085] After performing inter-feature multidimensional analysis, including center-kernel alignment similarity analysis, on teacher and student features, the resulting palm recognition model achieved an EER of 0.1172%; at FAR=10 -3 The FRR was 0.1250% at FAR=10; -4 The FRR was 0.2375% at FAR=10. -5 The FRR was 0.3549% at FAR=10; -6 The FRR was 0.5135% when FAR < 10. -6 The FRR is 0.7524%.

[0086] This shows that after the palm feature model was improved, the FRR under different FAR conditions was effectively reduced, indicating that the improved palm feature model can improve the accuracy of palm feature recognition.

[0087] The hand recognition model training method provided in this embodiment uses a pre-trained hand recognition basic model as a student model. The student model extracts student features from a training batch of samples containing multiple hand images. A teacher model, used to guide the student model's training, extracts teacher features from the training batch. Through multidimensional feature analysis of the student and teacher features, a center-kernel alignment multidimensional fusion loss value is calculated. Based on the total loss function incorporating the center-kernel alignment multidimensional fusion loss value, the total loss value for the training batch of samples in a single training round is calculated. The total loss value is used to update the student model's parameters until the student model converges, at which point it is used as the hand recognition model. Therefore, performing multidimensional feature analysis, including center-kernel alignment similarity analysis, on the student and teacher features ensures that the total loss value takes into account multidimensional factors including center-kernel alignment, further improving the model's generalization ability, reducing the error rate of hand feature recognition, and increasing the recognition accuracy of the hand recognition model.

[0088] In some embodiments of this application, the process of calculating the center kernel alignment multidimensional fusion loss value between student features and teacher features by performing multidimensional analysis between student features and teacher features is described, such as... Figure 2 As shown, the process may include:

[0089] Step S210: Calculate the center alignment metric between teacher and student features by performing a relative independence analysis on student and teacher features.

[0090] Specifically, step S210 can first calculate the first kernel feature of the teacher features and the second kernel feature of the student features.

[0091] The kernel features can be calculated using either a linear kernel or an RBF kernel. The first kernel feature can be represented as... The second kernel feature can be represented as The shapes of both the first and second nuclear features can be [B, B].

[0092] Furthermore, the first centered kernel feature of the first kernel feature and the second centered kernel feature of the second kernel feature can be calculated.

[0093] Specifically, the first-centered kernel feature and the second-centered kernel feature can be calculated using the following formula:

[0094]

[0095] in, It is the identity matrix. Given a square matrix with all elements equal to 1, the shape is [B, B]. The normalized matrix has the shape [B, B]. The first centralized kernel feature, This is a second-centered nuclear feature.

[0096] Furthermore, the independence index between teacher features and student features can be calculated based on the first and second centralized kernel features.

[0097] Specifically, the independence criterion can be the Hilbert-Schmidt independence criterion, used to measure the independence of student and teacher characteristics. The specific independence criterion is as follows:

[0098]

[0099] in, Represents the trace of a matrix.

[0100] Finally, the independence index can be normalized to obtain the center kernel alignment metric between teacher and student characteristics.

[0101] Specifically, the normalization process for the independence index is as follows:

[0102] .

[0103] Step S220: Based on the center kernel alignment metric, calculate the center kernel alignment similarity fusion value by performing feature similarity analysis on student features and teacher features.

[0104] Specifically, step S220 can first calculate the similarity measure between teacher features and student features. The similarity measure can be the cosine similarity between student features and teacher features. Then, the center kernel alignment metric and the similarity metric are weighted and summed to obtain the center kernel alignment similarity fusion value.

[0105] The center kernel alignment similarity fusion value can be expressed as:

[0106]

[0107] in, The center kernel alignment similarity fusion value, The fusion weights for the center kernel alignment metric. The fusion weight is the cosine similarity between student features and teacher features.

[0108] Step S230: Based on the center kernel alignment metric, calculate the center kernel alignment regularization value by performing a prediction error sensitivity analysis on student features and teacher features.

[0109] Specifically, step S230 can first calculate the mean squared error between student characteristics and teacher characteristics. Then, the center kernel alignment metric is regularized using the mean squared error to obtain a center kernel alignment regularization value, which enhances the sensitivity to prediction errors.

[0110] The process of regularizing the center kernel alignment metric using the mean square error is as follows:

[0111]

[0112] in, express The intensity of contribution.

[0113] Step S240: Weighted fusion of center kernel alignment similarity fusion value and center kernel alignment regularization value to obtain center kernel alignment multidimensional fusion loss value between student features and teacher features.

[0114] Specifically, the first balancing hyperparameter can be used. As a weighting coefficient for the center kernel alignment similarity fusion value, the second balancing hyperparameter As a weighting coefficient for the center kernel alignment regularization value, the center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between student features and teacher features. Among them, the first balancing hyperparameter... With the second equilibrium hyperparameter The sum of is 1.

[0115] The center kernel alignment multidimensional fusion loss value can be:

[0116]

[0117] Among them, the first equilibrium hyperparameter With the second equilibrium hyperparameter Adjustments can be made based on the experiment.

[0118] Therefore, the introduction of similarity metrics (such as cosine similarity) makes the improved center kernel alignment metric more robust to scaling transformations and local noise, especially suitable for non-Euclidean data. The addition of mean squared error enhances sensitivity to large errors and optimizes the model's ability to handle outliers. In knowledge distillation tasks, the improved center kernel alignment metric can more accurately capture feature alignment relationships across models, improving the transfer performance of student models. The improved formula retains the scale invariance and orthogonal transformation invariance of CKA, while providing a more intuitive optimization objective through mean squared error and similarity metrics. Mathematically, the composite loss function still satisfies the convex optimization condition, ensuring convergence to the global optimum.

[0119] Next, this application also provides an access control identification method, such as... Figure 3 As shown, the access control identification method may include the following steps:

[0120] Step S310: Obtain the user's palm image to be recognized, input the palm image to be recognized into the palm recognition model, and output the user's palm feature vector.

[0121] Specifically, the image of the hand to be recognized can be obtained by acquiring the current user's hand using a data acquisition device installed on the access control system. This image is then input into a trained hand recognition model for feature vector extraction. The hand recognition model can be deployed locally on the access control system.

[0122] The palm recognition model was trained using the palm recognition model training method described in the aforementioned method embodiment.

[0123] Step S320: Calculate the similarity between the palm feature vector and the user's standard feature vector stored in the access control database to obtain the feature comparison similarity.

[0124] Specifically, after extracting the palm feature vector, the similarity between the user's palm feature vector and the palm feature vectors of each user stored in the access control database is calculated to obtain the feature similarity of each user.

[0125] Step S330: If the feature comparison similarity exceeds the similarity comparison threshold, then the access control device is activated.

[0126] Specifically, multiple features can be compared to a similarity threshold. If a feature comparison exceeds the similarity threshold, it is considered that a user who has pre-registered and recorded palm information has requested to open the access control device, and the access control device can be opened.

[0127] The access control recognition method provided in this embodiment acquires a user's palm image to be recognized, inputs the palm image into an improved palm recognition model, outputs the user's palm feature vector, and calculates the similarity between the palm feature vector and the user's standard feature vector stored in the access control database. If the feature similarity exceeds a similarity comparison threshold, the access control device is opened. Therefore, because the improved palm recognition model performs multidimensional feature analysis, including center kernel alignment similarity analysis, on student and teacher features during the training phase, the total loss value can take into account multidimensional factors including center kernel alignment, further improving the model's generalization ability. Thus, when the model is applied to access control palm recognition, it can further reduce the error rate of palm feature recognition and improve the recognition accuracy of the palm recognition model.

[0128] The apparatus for training a palm recognition model provided in the embodiments of this application will be described below. The apparatus for training a palm recognition model described below can be referred to in correspondence with the training method for training a palm recognition model described above.

[0129] See Figure 4 , Figure 4 This is a schematic diagram of a device structure for training a palm recognition model, as disclosed in an embodiment of this application.

[0130] like Figure 4 As shown, the device may include:

[0131] The feature extraction unit 11 is used to extract student features from a training batch of samples containing multiple palm images using a pre-trained palm recognition basic model as a student model, and to extract teacher features from the training batch of samples using a teacher model used to guide the training of the student model.

[0132] The center kernel alignment multidimensional fusion loss value calculation unit 12 is used to calculate the center kernel alignment multidimensional fusion loss value between the student features and the teacher features by performing multidimensional analysis between the features;

[0133] The total loss calculation unit 13 is used to calculate the total loss value of the training batch samples under a single round of training based on the total loss function that incorporates the center kernel aligned multidimensional fusion loss value.

[0134] The model parameter update unit 14 is used to update the model parameters of the student model using the total loss value until the student model converges, and then use the student model as a palm recognition model.

[0135] Optionally, the central kernel aligned multidimensional fusion loss value calculation unit includes:

[0136] The relative independence analysis unit is used to calculate the center kernel alignment metric between the teacher features and the student features by performing relative independence analysis on the student features and the teacher features;

[0137] The feature similarity analysis unit is used to calculate the center kernel alignment similarity fusion value by performing feature similarity analysis on the student features and the teacher features based on the center kernel alignment metric value.

[0138] The prediction error sensitivity analysis unit is used to calculate the center kernel alignment regularization value by performing prediction error sensitivity analysis on the student features and the teacher features based on the center kernel alignment metric value.

[0139] The multi-factor weighted fusion unit is used to weight and fuse the center kernel alignment similarity fusion value and the center kernel alignment regularization value to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features.

[0140] Optionally, the relative independence analysis unit includes:

[0141] A kernel feature calculation unit is used to calculate the first kernel feature of the teacher's features and the second kernel feature of the student's features;

[0142] A centralized kernel feature calculation unit is used to calculate the first centralized kernel feature of the first kernel feature and the second centralized kernel feature of the second kernel feature;

[0143] An independence index calculation unit is used to calculate the independence index between the teacher feature and the student feature based on the first centralized kernel feature and the second centralized kernel feature;

[0144] The normalization processing unit is used to normalize the independence index to obtain the center kernel alignment metric between the teacher feature and the student feature.

[0145] Optionally, the feature similarity analysis unit includes:

[0146] The first feature similarity analysis subunit is used to calculate the similarity measure between the teacher features and the student features;

[0147] The second feature similarity analysis subunit is used to obtain a center kernel alignment similarity fusion value by weighted summation of the center kernel alignment metric and the similarity metric.

[0148] Optionally, the prediction error sensitivity analysis unit includes:

[0149] The first prediction error sensitivity analysis subunit is used to calculate the mean square error between the teacher characteristics and the student characteristics;

[0150] The second prediction error sensitivity analysis subunit is used to regularize the center kernel alignment metric using the mean square error to obtain a center kernel alignment regularization value.

[0151] Optionally, the multi-factor weighted fusion unit includes:

[0152] A multi-factor weighted fusion subunit is used to use a first balancing hyperparameter as the weighting coefficient of the center kernel alignment similarity fusion value and a second balancing hyperparameter as the weighting coefficient of the center kernel alignment regularization value to weightedly fuse the center kernel alignment similarity fusion value and the center kernel alignment regularization value to obtain the center kernel alignment multidimensional fusion loss value between the student feature and the teacher feature, wherein the sum of the first balancing hyperparameter and the second balancing hyperparameter is 1.

[0153] Optionally, the learning rate of the student model during training is 10. -4 .

[0154] The following describes the apparatus for implementing access control identification provided in the embodiments of this application. The apparatus for implementing access control identification described below can be referred to in correspondence with the method for implementing access control identification described above.

[0155] See Figure 5 , Figure 5 This is a schematic diagram of a device structure for implementing access control identification, as disclosed in an embodiment of this application.

[0156] like Figure 5 As shown, the device may include:

[0157] User palm acquisition unit 21 is used to acquire a user's palm image to be identified, input the palm image to be identified into the palm recognition model as described above, and output the user's palm feature vector.

[0158] The feature comparison similarity calculation unit 22 is used to calculate the similarity between the palm feature vector and the standard feature vectors of each user stored in the access control database to obtain the feature comparison similarity.

[0159] Access control opening unit 23 is used to open the access control device if the feature comparison similarity exceeds the similarity comparison threshold.

[0160] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0161] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0162] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A training method for a palm recognition model, characterized in that, include: Using a pre-trained palm recognition basic model as a student model, student features are extracted from a training batch of samples containing multiple palm images using the student model, and teacher features are extracted from the training batch of samples using a teacher model used to guide the training of the student model. By performing multidimensional feature analysis between the student features and the teacher features, the center kernel alignment multidimensional fusion loss value between the student features and the teacher features is calculated; Based on the total loss function that incorporates the center kernel aligned multidimensional fusion loss value, the total loss value of the training batch samples under a single round of training is calculated. The model parameters of the student model are updated using the total loss value until the student model converges, at which point the student model is used as the hand recognition model.

2. The method according to claim 1, characterized in that, The step of performing multidimensional feature analysis on the student features and teacher features to calculate the center kernel alignment multidimensional fusion loss value between the student features and teacher features includes: By performing a relative independence analysis on the student features and the teacher features, a center kernel alignment metric between the teacher features and the student features is calculated. Based on the center kernel alignment metric, a center kernel alignment similarity fusion value is calculated by performing feature similarity analysis on the student features and the teacher features. Based on the center kernel alignment metric, the center kernel alignment regularization value is calculated by performing a prediction error sensitivity analysis on the student features and the teacher features. The center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features.

3. The method according to claim 2, characterized in that, The step of performing a relative independence analysis on the student features and the teacher features, and calculating the center kernel alignment metric between the teacher features and the student features, includes: Calculate the first kernel feature of the teacher characteristics and the second kernel feature of the student characteristics; Calculate the first centered kernel feature of the first kernel feature and the second centered kernel feature of the second kernel feature; Based on the first centralized kernel feature and the second centralized kernel feature, calculate the independence index between the teacher feature and the student feature; The independence index is normalized to obtain the center kernel alignment metric between the teacher feature and the student feature.

4. The method according to claim 2, characterized in that, Based on the central kernel alignment metric, a central kernel alignment similarity fusion value is calculated by performing feature similarity analysis on the student features and the teacher features, including: Calculate the similarity measure between the teacher characteristics and the student characteristics; The center kernel alignment similarity fusion value is obtained by weighted summation of the center kernel alignment metric and the similarity metric.

5. The method according to claim 2, characterized in that, Based on the center kernel alignment metric, a center kernel alignment regularization value is calculated by performing a prediction error sensitivity analysis on the student features and the teacher features, including: Calculate the mean square error between the teacher characteristics and the student characteristics; The mean square error is used to regularize the center kernel alignment metric to obtain the center kernel alignment regularization value.

6. The method according to claim 2, characterized in that, The center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features, including: Using the first balancing hyperparameter as the weighting coefficient of the center kernel alignment similarity fusion value and the second balancing hyperparameter as the weighting coefficient of the center kernel alignment regularization value, the center kernel alignment similarity fusion value and the center kernel alignment regularization value are weighted and fused to obtain the center kernel alignment multidimensional fusion loss value between the student features and the teacher features, wherein the sum of the first balancing hyperparameter and the second balancing hyperparameter is 1.

7. The method according to any one of claims 1-6, characterized in that, The student model has a learning rate of 10 during training. -4 .

8. An access control identification method, characterized in that, include: Obtain a user's palm image to be identified, input the palm image to be identified into the palm recognition model as described in any one of claims 1-7, and output the user's palm feature vector; The similarity between the palm feature vector and the standard feature vectors of each user stored in the access control database is calculated to obtain the feature comparison similarity. If the similarity of the features exceeds the similarity comparison threshold, the access control device will be activated.

9. A training device for a palm recognition model, characterized in that, include: The feature extraction unit is used to extract student features from a training batch of samples containing multiple palm images using a pre-trained palm recognition basic model as a student model, and to extract teacher features from the training batch of samples using a teacher model used to guide the training of the student model. The center kernel aligned multidimensional fusion loss value calculation unit is used to calculate the center kernel aligned multidimensional fusion loss value between the student features and the teacher features by performing multidimensional analysis between the features; The total loss calculation unit is used to calculate the total loss value of the training batch samples under a single round of training based on the total loss function that incorporates the center kernel aligned multidimensional fusion loss value; The model parameter update unit is used to update the model parameters of the student model using the total loss value until the student model converges, and then uses the student model as a palm recognition model.

10. An access control identification device, characterized in that, include: The user palm acquisition unit is used to acquire a user's palm image to be identified, input the palm image to be identified into the palm recognition model as described in claim 9, and output the user's palm feature vector. The feature comparison similarity calculation unit is used to calculate the similarity between the palm feature vector and the standard feature vectors of each user stored in the access control database to obtain the feature comparison similarity. The access control opening unit is used to open the access control device if the similarity of the features exceeds a similarity comparison threshold.

Images