CN109934114A - A finger vein template generation and update algorithm and system - Google Patents

Abstract

The invention relates to the technical field of biometric identification, and discloses a finger vein template selection and update algorithm and system; first, the finger vein images are collected; The image is segmented and enhanced, and then binarized and refined. Then, three template selection methods based on the least squares method, based on the intra-class weighted least squares method and based on the inter-class weighted least squares method are compared respectively. The optimal template generation method is selected; finally, a template database is established based on the optimal template generation method and updated.

Description

A finger vein template generation and update algorithm and system

technical field

The invention relates to the technical field of biometric identification, in particular to a finger vein template generation and update algorithm and system.

Background technique

With the rapid development of Internet technology, the issue of information security has become more and more important. How to effectively identify the identity to protect personal and property safety has become an urgent problem to be solved. Compared with traditional authentication methods such as keys and passwords, biological and behavior-based biometrics are difficult to steal, copy and lose. Therefore, biometric authentication technology has been widely studied and successfully applied to personal identity authentication. At present, the biometrics used for identity authentication are mainly divided into two types:

(1) External features: face, fingerprint and iris, etc.

(2) Internal features: finger veins, palm veins and dorsal hand veins. Compared to external features, intrinsic biometrics are located under the skin making them harder to steal and counterfeit, so they have a higher security feature.

Although the finger vein biometric technology has achieved great results, there are also some shortcomings, such as when the same finger is sampled multiple times at the same time or at different times, due to the offset, rotation, bending and illumination of the finger, etc. The change of conditions makes the obtained images often have large differences, which in turn reduces the recognition performance of the system. To minimize intra-class differences, many current works attempt to capture multiple images as registration templates for authentication. However, this solution requires more storage space and increases the authentication time. In order to solve this problem, other works propose to generate super-registration templates, and then update the super-templates during the authentication process to reduce the intra-class differences in the collected images. The scheme achieves good recognition performance. However, the current template generation and update methods are mainly aimed at biometric recognition systems such as fingerprints, faces and irises, so they cannot be effectively used in finger vein recognition systems. For the finger vein recognition system, there is no relevant template generation and update scheme at present. Therefore, in order to solve this problem, the present invention proposes a method for generating a finger vein registration template to generate a robust registration template. Then, in the authentication stage, we propose a template update method to update the registration template and the finger static template online, so as to save the multiple changes of the same finger during the acquisition process as much as possible, reduce intra-class differences, and improve the system. certification performance.

SUMMARY OF THE INVENTION

technical problems solved

Aiming at the deficiencies of the prior art, the present invention provides an algorithm and system for generating and updating a finger vein template, which generates an optimal template through a weighted least squares method to save storage space and time. At the same time, real-time online update is realized in combination with the newly collected template library of the finger vein recognition system.

Technical solutions

To achieve the above purpose, the present invention is achieved through the following technical solutions:

A method for generating a finger vein template, comprising the following operation steps:

Image acquisition: Obtain the corresponding finger vein image through an infrared camera; perform multiple acquisitions for each individual to the registration image;

Image segmentation: Since our method is to generate and update binary vein templates, it is necessary to perform feature extraction on the collected images to obtain binary images. This paper uses the current deep learning method to extract finger vein features. First, the image is divided into background area, blur area and target area by threshold method, and a training set is established by using background pixels and vein pixels. Then, build a convolutional neural network model and train it with the training data. Finally, the vein features are enhanced using the post-training model, and the enhanced images are binarized;

Generating templates: In the process of finger vein template generation, we propose a template generation method based on the weighted least squares method. In order to reduce intra-class variation, we fuse multiple finger templates of the same class into an optimal template. The main purpose of template generation is to reduce intra-class variation, so we generate an optimal template by minimizing the intra-class distance. First, the definition of the optimal template of finger vein is given. Then, the template generation problem is transformed into an optimization problem based on this definition. Second, in order to improve the performance of the algorithm, the intra-class and inter-class matching scores are calculated to obtain weights and injected into the objective function. Finally, by solving the optimization problem, a robust registration template is obtained.

Update template: During the authentication process, due to the influence of various factors, the collected finger vein characteristics change, which makes it difficult to achieve effective matching and authentication between the images obtained by the same finger and the registration template. In order to improve the matching accuracy, we propose a template update method to achieve online update of templates. First, in the authentication phase, the images to be authenticated are matched with the original registration images, and the authentication images whose matching scores are greater than a given threshold are stored to update the original registration template. Then, the inter-class distance and the intra-class distance are calculated by the authentication image and the original registration template, and the two distances are fused to obtain the weight of each image. Similar to template generation, template updating is transformed into an optimization problem through the definition of template quality, and a new registration template is obtained by least squares method, and the original registration template is updated with this template.

A. Validate Template Quality Definitions

The purpose of the existing biometric template generation is to improve the performance of the authentication system, mainly to reduce the verification error rate. Therefore, biometric template generation should focus on the reduction of authentication error rate, rather than based on people's subjective perception of registration templates, such as images with high contrast that people usually think can be used as registration templates. In the existing recognition system, when a user is recognized, the system will obtain the biometric data of a user again, and perform image preprocessing on it, and match the extracted features with the data stored in the data set. The accuracy of verification is closely related to the stability of people's biometric data in different periods. In other words, the validation error rate is mainly caused by intra-class variation. Therefore, we consider a high-quality finger vein registration template to be an image that has the smallest intra-class distance from its similar registration images.

B. Weight calculation

Templates are fused from multiple registration samples. In fact, there are some differences between registered samples, so we assume that different samples have different weights. Intuitively, samples with smaller intra-class similarity and larger inter-class similarity should have a larger weight in template generation. Therefore, we decide the weight it occupies in template generation by calculating the similarity between this sample and other registered samples.

1) Similarity calculation: In order to obtain a robust weight, first, the similarity between two registration templates is defined as in Equation (2). Then, weights are automatically assigned to each sample based on intra-class similarity and inter-class similarity. Suppose E and F are binarized feature images extracted from two registered samples with sizes x and y, respectively. The width and height of E are expanded to 2w+x and 2h+y respectively, and its expanded image is shown below.

The similarity of E and F can be calculated by formula (2):

in

2) Intra-class similarity calculation: For each finger vein image sample, we calculate the distance between this finger and other finger vein samples in the same class according to Equation (2), and calculate its average value as the weight of the sample. Suppose there are N classes and each class has M registered samples. X _m,n refers to the mth sample in the nth category, and the weight X _m,n is calculated as follows:

d(X _i,n ,X _m,n ) refers to the similarity between samples X _i,n and samples X _m,n . w _m,n is actually the average intra-class similarity of X _m,n and other M-1 registered samples in the same class.

3) Inter-class similarity calculation: Similarly, we calculate the inter-class distance for each registered sample to obtain weights. Suppose w' _m,n is the weight of the mth sample of the nth class, which is calculated by the following formula:

d(X _m,j , X _m,n ) refers to the similarity between sample X _m,j and sample X _m,n . Equation (6) calculates the average similarity of sample X _m,n and all samples of other N-1 categories.

4) Similarity fusion: Different samples have different intra-class and inter-class similarities. We fuse their weights by the following formula:

Determines the respective proportions of intra-class similarity and inter-class similarity in template generation.

C. Template generation

For the nth category, we mainly generate a template T _n to minimize the distance between this template and its registered samples, so the template generation problem is transformed to solve the following optimization problem.

In the formula, W _m,n refers to the weight of the mth sample of the nth category. Equation (8) can be solved by the least squares method. Therefore, the template T _n in the nth category can be obtained by formula (9):

By formula (3) and formula (9), we can get the samples with the smallest intra-class distance and the largest inter-class distance in template generation.

Finger Vein Template Update

During validation, finger vein quality changes due to changes in the conditions under which the images are generated, resulting in large intra-class variations. Due to the limited memory and computing power of finger vein devices, it is difficult to obtain multiple finger vein images from the same finger as registration templates. To solve this problem, we propose an online template update method.

A weight calculation

In the testing phase, feature extraction is performed on each input image and the similarity between the feature map and the template is computed for validation. If its similarity is greater than some predefined threshold, the input image (unlabeled image) is stored in the data to update the template. It is assumed that N templates corresponding to N classes are stored in the recognition system, and each class has stored K input images. Since we utilize N templates and K input images to update the templates, we compute their weights separately.

X _k,n refers to the kth input image in the nth category, k=1,2,..,K. The original template is denoted by T _o,n , n=1,2,...N. The weight of the original template is represented by W _0,n .

where w _0,n denotes its intra-class average similarity to k input samples. w _0,n is calculated by Equation (11)

w' _0,n is calculated by equation (12)

Equation (12) actually computes the average inter-class similarity of template T _0,n with other N-1 templates.

After calculating the weight of the template, we calculate the weight of the input image. The weight of the kth input sample x _k,n is calculated by formula (13)

In formula (13), w _k,n is the average similarity between the k-th input image and other K-1 input images and its corresponding template

w' _k,n refers to the average similarity between the kth input image in the nth sample and the other n-1 templates.

B template update

Like formula (8), the improved template T _n (k) in the nth category is obtained by solving the following objective function

Finally, generate the improved template

Note that Tn* in Equation (9) and Tn _* in Equation (17) are two probability maps. To complete the verification, we binarize them, and then store the resulting binarized image for verification of the registered template.

beneficial effect

The present invention proposes a finger vein template selection and update algorithm and system for the first time. Compared with the prior art, the present invention has the following beneficial effects:

1. The invention first proposes to generate and update the registration template in the finger vein recognition system.

2. The present invention proposes a weighted least squares method using inter-class distance and intra-class distance. The template generation of this method can not only obtain the optimal template, but also save a lot of memory storage space and finger vein matching time.

3. The present invention proposes an online finger vein authentication system template update algorithm, which can update the registration template in real time, reduce the differences between categories of finger vein images, and improve the recognition accuracy of the system.

4. The weighted least squares method of the present invention can be applied not only to the finger vein recognition system, but also to the establishment of other biometric templates, and has good scalability.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

FIG. 1 is a flow chart for generating a finger vein template according to the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Example:

A method for generating a finger vein template of the present embodiment includes the following operation steps:

Image acquisition: Obtain the corresponding finger vein image through an infrared camera; perform multiple acquisitions for each individual to the registration image;

Image segmentation: Since our method is to generate and update binary vein templates, it is necessary to perform feature extraction on the collected images to obtain binary images. This paper uses the current deep learning method to extract finger vein features. First, the image is divided into background area, blur area and target area by threshold method, and a training set is established by using background pixels and vein pixels. Then, build a convolutional neural network model and train it with the training data. Finally, the vein features are enhanced using the post-training model, and the enhanced images are binarized;

Generating templates: In the process of finger vein template generation, we propose a template generation method based on the weighted least squares method. In order to reduce intra-class variation, we fuse multiple finger templates of the same class into an optimal template. The main purpose of template generation is to reduce intra-class variation, so we generate an optimal template by minimizing the intra-class distance. First, the definition of the optimal template of finger vein is given. Then, the template generation problem is transformed into an optimization problem based on this definition. Second, in order to improve the performance of the algorithm, the intra-class and inter-class matching scores are calculated to obtain weights and injected into the objective function. Finally, by solving the optimization problem, a robust registration template is obtained.

Update template: During the authentication process, due to the influence of various factors, the collected finger vein characteristics change, which makes it difficult to achieve effective matching and authentication between the images obtained by the same finger and the registration template. In order to improve the matching accuracy, we propose a template update method to achieve online update of templates. First, in the authentication phase, the images to be authenticated are matched with the original registration images, and the authentication images whose matching scores are greater than a given threshold are stored to update the original registration template. Then, the inter-class distance and the intra-class distance are calculated by the authentication image and the original registration template, and the two distances are fused to obtain the weight of each image. Similar to template generation, template updating is transformed into an optimization problem through the definition of template quality, and a new registration template is obtained by least squares method, and the original registration template is updated with this template.

D. Validation Template Quality Definitions

The main purpose of existing template generation methods is to improve performance by reducing the validation error rate. Therefore, biometric template generation is mainly defined by minimizing distances, rather than based on people's subjective perception of registration templates. In the existing recognition system, when a user is recognized, the system will obtain the biometric data of a user again, and perform image preprocessing on it, and match the extracted features with the data stored in the data set. The accuracy of validation is closely related to the biometric data of people at different time periods. In other words, the validation accuracy is mainly caused by intra-class variation, so we assume a high-quality finger vein with the smallest intra-class distance among all registered templates.

E. Weight calculation

Templates are fused from multiple registration samples. In fact, there are some differences between registered samples, so we assume that different samples have different weights. Intuitively, samples with larger intra-class similarity and smaller inter-class similarity should have a larger weight in template generation. Therefore, we determine the weight it occupies in template generation by calculating the distance between this sample and other registered samples.

5) Similarity calculation: In order to obtain a robust weight, first, the similarity between two registration templates is defined, as in Equation (2). Then, weights are automatically assigned to each sample based on intra-class similarity and inter-class similarity. Suppose E and F are the binarized feature maps extracted from two registered samples with sizes x, y, respectively. The width and height of E are expanded to 2w+x and 2h+y respectively, and its expanded image is shown below.

The similarity of E and F can be calculated by formula (2):

in

6) Intra-class similarity calculation: For each finger vein image, we calculate the distance between this finger and other fingers in the same class according to equation (2), calculate its average and use it as a weight. Now suppose there are N categories, and each category has M registered samples. Xm,n refers to the mth sample in the nth category, and the weight Xm,n is calculated as follows:

d(Xi,n,Xm,n) refers to the similarity between samples Xi,n and samples Xm,n. Wm,n is actually the average intra-class similarity of Xm,n and other M-1 registered samples in the same class.

7) Inter-class similarity calculation: Similarly, we calculate the inter-class distance of each registered sample to obtain weights. The weight W′ _m,n refers to the mth sample of the nth class, and it is calculated by the following formula:

d(Xi,n,Xm,n) refers to the similarity between samples Xi,n and samples Xm,n. Equation (6) calculates the average similarity of sample Xm,n and all samples of other N-1 categories.

8) Similarity fusion: Different samples have different intra-class and inter-class similarities, and we fuse their weights by the following formula:

Determines the respective proportions of intra-class similarity and inter-class similarity in template generation.

F. Template Generation

For the nth class, we mainly aim to generate a template Tn such that there is minimal intra-class variation between this template and its registered samples, so the template generation problem is transformed into an optimization problem.

In the formula, Wm,n refers to the weight of the mth sample of the nth category. Equation (8) can be solved by the least squares method. Therefore, the template Tn in the nth category can be obtained by formula (9):

By formula (3) and formula (9), we can get the samples with the smallest intra-class distance and the largest inter-class distance in template generation.

Finger Vein Template Update

During validation, finger vein quality changes due to changes in the conditions under which the images are generated, resulting in large intra-class variations. Due to the limited memory and computing power of finger vein devices, it is difficult to obtain multiple finger vein images from the same finger. To solve this problem, we propose a method to continuously update the template through online operations.

A weight calculation

In the testing phase, feature processing is performed on each input image and the similarity between the feature map and the template is calculated for verification. If its similarity is greater than some predefined threshold, this input image (unlabeled image) will be stored to update the template during validation. Suppose there are N templates in N categories, and each category stores K input images. Since the update of the template is based on N categories and K input images, we calculate their weights separately.

Xk,n refers to the Kth input image in the Nth category, K=1,2,..,K. The original template is denoted by T _o,n , n=1,2,...N. The weight of the original template is represented by W _0,n .

where w _0,n refers to the intra-class average similarity of the K input samples. w _0,n is calculated by Equation (11)

w' _0,n is calculated by equation (12)

Equation (12) calculates the average inter-class similarity of template T _0,n and other N-1 templates.

The weight of the Kth input sample x _k,n is calculated by Equation (13)

In formula (13), w _k,n is the average similarity between the kth input and the other K-1 inputs

w' _k,n refers to the average similarity between the kth input in the nth sample and the other n-1 templates.

B template improvements

Like formula (8), the improved template T _n (k) in the nth category is obtained by calculating the following objective function

Finally, generate the improved template

Note that Tn* in Equation (9) and Tn _* in Equation (17) are two probability maps. In order to complete the verification, they should be binarized, and then the stored binarized image is used for the verification of the registered template.

It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any such actual relationship or sequence exists. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The recorded technical solutions are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (6)

1. a finger vein template generation method, is characterized in that, comprises the following operation steps: Image acquisition: Obtain the corresponding finger vein image through an infrared camera; perform multiple acquisitions for each individual to the registration image; Image segmentation: Since our method is to generate and update binary vein templates, it is necessary to perform feature extraction on the collected images to obtain binary images. This paper uses the current deep learning method to extract finger vein features. First, the image is divided into background area, blur area and target area by threshold method, and a training set is established by using background pixels and vein pixels. Then, build a convolutional neural network model and train it with the training data. Finally, the vein features are enhanced using the post-training model, and the enhanced images are binarized; Generating templates: In the process of finger vein template generation, we propose a template generation method based on the weighted least squares method. In order to reduce intra-class variation, we fuse multiple finger templates of the same class into an optimal template. The main purpose of template generation is to reduce intra-class variation, so we generate an optimal template by minimizing the intra-class distance. First, the definition of the optimal template of finger vein is given. Then, the template generation problem is transformed into an optimization problem based on this definition. Second, in order to improve the performance of the algorithm, the intra-class and inter-class matching scores are calculated to obtain weights and injected into the objective function. Finally, by solving the optimization problem, a robust registration template is obtained. Update template: During the authentication process, due to the influence of various factors, the collected finger vein characteristics change, which makes it difficult to achieve effective matching and authentication between the images obtained by the same finger and the registration template. In order to improve the matching accuracy, we propose a template update method to achieve online update of templates. First, in the authentication phase, the images to be authenticated are matched with the original registration images, and the authentication images whose matching scores are greater than a given threshold are stored to update the original registration template. Then, the inter-class distance and the intra-class distance are calculated by the authentication image and the original registration template, and the two distances are fused to obtain the weight of each image. Similar to template generation, template updating is transformed into an optimization problem through the definition of template quality, and a new registration template is obtained by least squares method, and the original registration template is updated with this template. 2. a kind of finger vein template generation method according to claim 1, is characterized in that, described in-class similarity is calculated as: d(Xi,n,Xm,n) refers to the similarity between samples Xi,n and samples Xm,n. Wm,n is actually the average intra-class similarity of Xm,n and other M-1 registered samples in the same class. 3. a kind of finger vein template generation method according to claim 1, is characterized in that, described similarity between classes is calculated as: d(Xi,n,Xm,n) refers to the similarity between samples Xi,n and samples Xm,n. Equation (2) calculates the average similarity of sample Xm,n and all samples of other N-1 categories. 4. a kind of finger vein template generation method according to claim 1, is characterized in that, described similarity merges to be: Determines the respective proportions of intra-class similarity and inter-class similarity in template generation. 5. a kind of finger vein template generation method according to claim 1, is characterized in that, described finger vein template is generated as: In the formula, Wm,n refers to the weight of the mth sample of the nth category. 6. a kind of finger vein template generation method according to claim 1, is characterized in that, described finger vein template is updated as: X _k,n refers to the Kth input image in the Nth category, K=1,2,..,K. The original template is denoted by T _o,n , n=1,2,...N. The weight of the original template is represented by W _0,n . where w _0,n refers to the intra-class average similarity of the K input samples. w _0,n is calculated by formula (6) w' _0,n is calculated by formula (7) Equation (7) calculates the average inter-class similarity of template T _0,n and other N-1 templates. The weight of the Kth input sample x _k,n is calculated by Equation (8) In formula (9), w _k,n is the average similarity between the kth input and the other K-1 inputs w' _k,n refers to the average similarity between the kth input in the nth sample and the other n-1 templates. Template update The improved template T _n (k) in the nth category is obtained by computing the following objective function Finally, generate the improved template Note that Tn* in Equation (4) and Tn _* in Equation (12) are two probability maps. In order to complete the verification, they should be binarized, and then the stored binarized image is used for the verification of the registered template.