Few-shot finger vein recognition method based on meta-learning

By using technologies such as hollow convolution and average pooling in the meta-learning method to improve the feature extraction network, the problem of insufficient feature extraction capabilities in the existing technology is solved, and the accuracy and efficiency of finger vein recognition are significantly improved.

CN115019356BActive Publication Date: 2024-12-03HANGZHOU LION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210558629.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-20
Publication Date
2024-12-03
Estimated Expiration
2042-05-20

AI Technical Summary

Technical Problem

When existing meta-learning methods process complex background information, their feature extraction capabilities are insufficient, resulting in low efficiency in venous recognition and inability to effectively extract useful information.

Method used

The feature extraction network is improved by using hollow convolution and average pooling, adding receptive fields and adding additional convolutional layers to enhance feature extraction capabilities, combined with batch normalization and ReLU activation functions, and optimize model parameters using cross entropy loss function and stochastic gradient descent .

Benefits of technology

It improves the accuracy of image feature extraction and recognition accuracy, can process complex input images more effectively, and improves the accuracy of finger venous classification results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115019356B_ABST
    Figure CN115019356B_ABST
Patent Text Reader

Abstract

A few-shot finger vein recognition method based on meta-learning, comprising: dividing a finger vein dataset into a meta-training set and a meta-test set; randomly initializing a meta-learning model; for the i-th task T i , extract multiple T i , forming a batch, where T i ~p(T); if there are m tasks, then T={T1,T2,T3,……,T m}; After extracting features through the constructed convolutional neural network, the support set loss value is calculated; for each task T i , calculate the loss of the query set #imgabs0# use stochastic gradient descent with a step size of β to optimize the meta-parameters; task T i The data set is composed of D = {x j ,y j}, where the observed value is x j , the target value is y j , using cross entropy as the loss function for the classification problem; after training, using the initialization parameters learned on the finger vein element training set to fine-tune on the finger vein element test set, outputting the finger vein classification result. The present invention fully guarantees information extraction while enhancing its recognition accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of few-sample finger vein recognition, and in particular relates to a few-sample finger vein recognition method. Background Art

[0002] Meta-learning refers to the use of previously accumulated experience to quickly learn new tasks. The main goal is to understand the connection between the learning strategy and the specific environment in which the strategy is applicable, use meta-knowledge in the learning process to obtain a good model, and be able to adjust its own behavior in the current learning environment to better perform the learning process. Meta-learning focuses on improving the overall learning ability of the model, rather than the ability to solve a specific problem. It neither increases the number of model parameters nor sets any restrictions on the structure of the model.

[0003] The feature extraction network used in meta-learning has a total of four convolution blocks, each of which includes a 3×3 convolution kernel for feature extraction; batch normalization (BN) is used after convolution to prevent gradient vanishing, ReLU is used as the nonlinear activation function, and then a 2×2 maximum pooling layer is connected. When the background information of the input image is more complex, the background information cannot be extracted well, and the four-layer convolution block has a weak ability to extract features of the image, resulting in the failure to effectively extract all useful information, which ultimately leads to low recognition efficiency. Summary of the invention

[0004] In order to solve the above problems in the prior art, the present invention provides a few-sample finger vein recognition method based on meta-learning.

[0005] A few-sample finger vein recognition method based on meta-learning of the present invention comprises the following steps:

[0006] S1, divide the finger vein dataset into a meta-training set and a meta-test set;

[0007] The finger vein element training set randomly extracts multiple groups of N categories to form multiple groups of tasks T i , each task T i It is divided into support set and query set for training meta-learning model. The meta-test set randomly extracts multiple groups of N categories to form multiple groups of tasks T i , each task T i Divide into support set and query set for testing meta-learning models;

[0008] S2, randomly initialize the meta-learning model f θ Parameter θ;

[0009] S3. For the i-th task T i , extract multiple Ti , forming a batch, where T i ~p(T);

[0010] S4. If there are m tasks, then T = {T1, T2, T3, ..., T m}. From each T i In the process, K finger vein images are extracted, and features are extracted from the convolutional neural network model constructed by inputting K images, and the network parameters are designed; specifically, the following are included:

[0011] S41. Change the traditional convolution of the first layer to a dilated convolution with a dilation rate of 2 to increase the receptive field and change the convolution step size to 1. Use batch normalization (BN) after the convolution layer, use ReLU as the nonlinear activation function, and then connect the average pooling with a pooling step size of 2.

[0012] S42, the improvements of the second, third and fourth layers are the same as the first layer structure;

[0013] S43, add another convolutional layer after the fourth convolutional layer, whose structure is the same as the first layer to enhance its feature extraction capability;

[0014] Table 1 shows the corresponding parameters of feature maps of each layer.

[0015] Table 1 Feature extraction network settings

[0016]

[0017] S5, extract features through the constructed convolutional neural network and calculate the support set loss value. For each specific task T i , the loss of the support set is The task-specific parameter is φ, which is updated using the first-order gradient of the support set loss or a higher-order gradient, such as a first-order gradient with a step size of α;

[0018]

[0019] S6. For each task T i , calculate the loss of the query set The update of model parameters is based on all tasks, so the goal of meta-learning optimization is;

[0020]

[0021] S7, using stochastic gradient descent with a step size of β to optimize the meta-parameters, the model parameters are optimized as follows;

[0022]

[0023] S8, Task T iThe data set is composed of D = {x j ,y j}, where the observed value is x j , the target value is y j . Use cross entropy as the loss function for classification problems;

[0024]

[0025] S9. After the training is completed, the initialization parameters learned on the finger vein element training set are used to perform fine-tuning on the finger vein element test set, and the finger vein classification results are output; specifically, the following are included:

[0026] Repeat steps S2 to S8. When the model after step S6 converges on the support set of the finger vein element test set, apply it to the query set of the finger vein element test set and output the classification result of the finger vein image.

[0027] The present invention proposes a few-sample finger vein recognition method based on meta-learning. For its feature extraction network, each convolution layer is improved by using dilated convolution and average pooling, and a convolution layer is added to enhance its feature extraction capability. The present invention improves the feature extraction network of meta-learning, enhances its feature extraction capability to cope with more complex input images, and further improves its recognition accuracy.

[0028] By improving its feature extraction network, the present invention can extract more effective image information than the original network, so that the image classification result obtained will be more accurate.

[0029] The advantages of the present invention are: fully guaranteeing information extraction while enhancing its recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a feature extraction network diagram of a few-sample finger vein recognition method based on meta-learning proposed in an example of the present invention;

[0031] Figure 2 The meta-learning flowchart proposed for the example of the present invention;

[0032] Figure 3 The structural diagram of the finger vein original test set and meta-training set proposed for the example of the present invention. DETAILED DESCRIPTION

[0033] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0034] As an application example of a few-sample finger vein recognition method based on meta-learning of the present invention, this embodiment provides a finger vein feature extraction method, including the following steps:

[0035] S1. Divide the finger vein dataset into a meta-training set and a meta-test set; specifically:

[0036] The finger vein element training set randomly extracts multiple groups of N categories to form multiple groups of tasks T i , each task T i It is divided into support set and query set for training meta-learning model. The meta-test set randomly extracts multiple groups of N categories to form multiple groups of tasks T i , each task T i Divide into support set and query set for testing meta-learning models;

[0037] S2, randomly initialize the meta-learning model f θ Parameter θ;

[0038] S3. For the i-th task T i , extract multiple T i , forming a batch, where T i ~p(T);

[0039] S4. If we have m tasks, then T = {T1, T2, T3, ..., T m}. From each T i In the process, K finger vein images are extracted, and features are extracted from the convolutional neural network model constructed by inputting K images, and the network parameters are designed; specifically, the following are included:

[0040] S41. Change the traditional convolution of the first layer to a dilated convolution with a dilation rate of 2 to increase the receptive field and change the convolution step size to 1. Use batch normalization (BN) after the convolution layer, use ReLU as the nonlinear activation function, and then connect the average pooling with a pooling step size of 2.

[0041] S42, the improvements of the second, third and fourth layers are the same as the first layer structure;

[0042] S43, add another convolutional layer after the fourth convolutional layer, whose structure is the same as the first layer to enhance its feature extraction capability;

[0043] Table 1 shows the corresponding parameters of feature maps of each layer.

[0044] Table 1 Feature extraction network settings

[0045]

[0046] S5, extract features through the constructed convolutional neural network and calculate the support set loss value. For each specific task T i , the loss of the support set is The task-specific parameter is φ, which is updated using the first-order gradient of the support set loss or a higher-order gradient, such as a first-order gradient with a step size of α;

[0047]

[0048] S6. For each task T i , calculate the loss of the query set The update of model parameters is based on all tasks, so the goal of meta-learning optimization is;

[0049]

[0050] S7, using stochastic gradient descent with a step size of β to optimize the meta-parameters, the model parameters are optimized as follows;

[0051]

[0052] S8, Task T i The data set is composed of D = {x j ,y j}, where the observed value is x j , the target value is y j . Use cross entropy as the loss function for classification problems;

[0053]

[0054] S9. After the training is completed, the initialization parameters learned on the finger vein element training set are used to perform fine-tuning on the finger vein element test set, and the finger vein classification results are output; specifically, the following are included:

[0055] Repeat steps S2 to S8. When the model after step S6 converges on the support set of the finger vein element test set, apply it to the query set of the finger vein element test set and output the classification result of the finger vein image.

[0056] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A few-sample finger vein recognition method based on meta-learning, comprising the following steps: S1, divide the finger vein dataset into a meta-training set and a meta-test set; The finger vein element training set randomly extracts multiple groups of N categories to form multiple groups of tasks T i , each task T i It is divided into support set and query set for training meta-learning model. The meta-test set randomly extracts multiple groups of N categories to form multiple groups of tasks T i , each task T i Divide into support set and query set for testing meta-learning models; S2, randomly initialize the meta-learning model f θ Parameter θ; S3. For the i-th task T i , extract multiple T i , forming a batch, where T i ~p(T); S4. If there are m tasks, then T = {T1, T2, T3, ..., T m}; From each T i In the process, K finger vein images are extracted, and features are extracted from the convolutional neural network model constructed by inputting K images, and the network parameters are designed; specifically, the following are included: S41. Change the traditional convolution of the first layer to a dilated convolution with a dilation rate of 2 to increase the receptive field and change the convolution step size to 1. Use batch normalization (BN) after the convolution layer, use ReLU as the nonlinear activation function, and then connect the average pooling with a pooling step size of 2. S42, the improvements of the second, third and fourth layers are the same as the first layer structure; S43, add another convolutional layer after the fourth convolutional layer, whose structure is the same as the first layer to enhance its feature extraction capability; Table 1 shows the corresponding parameters of the feature maps of each layer; Table 1 Feature extraction network settings S5, extract features through the constructed convolutional neural network and calculate the support set loss value. For each specific task T i , the loss of the support set is The task-specific parameter is φ, which is updated using the first-order gradient of the support set loss or higher-order gradients, using the first-order gradient with a step size of α; S6. For each task T i , calculate the loss of the query set The update of model parameters is based on all tasks, so the goal of meta-learning optimization is; S7, using stochastic gradient descent with a step size of β to optimize the meta-parameters, the model parameters are optimized as follows; S8, Task T i The data set is composed of D = {x j ,y j }, where the observed value is x j , the target value is y j ; Use cross entropy as the loss function for classification problems; S9. After the training is completed, the initialization parameters learned on the finger vein element training set are used to perform fine-tuning on the finger vein element test set, and the finger vein classification results are output; specifically, the following are included: Repeat steps S2 to S8. When the model after step S6 converges on the support set of the finger vein element test set, apply it to the query set of the finger vein element test set and output the classification result of the finger vein image.