TWI792017B - Biometric identification system and identification method thereof - Google Patents

Abstract

A biometric identification system includes a sensor for sensing a first biometric feature to generate an image; a first feature extractor coupled to the sensor for generating a first information according to the image, wherein the first information represents uniqueness of the first biometric feature; a second feature extractor coupled to the sensor for generating a second information according to the image, wherein the second information represents true and false features of the image; an identification unit coupled to the first feature extractor and the second feature extractor for generating an identification result according to the first information and/or the second information.

Description

Biometric identification system and identification method

The invention relates to an identification system and an identification method of biological characteristics.

More and more electronic devices or systems use biometrics to identify users. Common biometric recognition includes fingerprint recognition, face recognition, iris recognition, voiceprint recognition and palmprint recognition. However, the current biometric identification still has defects, for example, fingerprint identification cannot effectively distinguish the authenticity of the finger. Therefore, there are still doubts that the biometric system may be spoofed.

The purpose of the present invention is to propose a biological feature identification system and its identification method, which can prevent false biological features from being identified.

According to the present invention, a biometric identification system includes a sensor, a first feature extraction unit, a second feature extraction unit, and a recognition unit. The sensor senses a first biological feature to generate an image. The first feature extraction unit is coupled to the sensor, and generates a first information according to the image, wherein the first information describes the true and false characteristics of the first biological feature. The second feature extraction unit is coupled to the sensor, and generates a second information according to the image, wherein the second information describes the true and false characteristics of the first biological feature. The recognition unit is coupled to the first feature extraction unit and the second feature extraction unit, and generates a recognition result according to the second information or the first information and the second information.

According to the present invention, a biometric identification method includes: sensing a first biometric characteristics to generate an image; generate a first information based on the image, wherein the first information describes the uniqueness of the first biological characteristic; obtain a second information from the image, wherein the second information describes the first The true and false characteristics of a biometric feature; and generating a recognition result according to the second information or the first information and the second information.

The identification system and method of the present invention can effectively improve the security of the biological feature identification system and prevent fake biological features from passing identity authentication.

10: Identification system

11: Sensor

12: The first feature extraction unit

13: The second feature extraction unit

131: Convolutional Neural Networks

1311: Feature extraction part

1312: classification part

14: Memory

15: Identification unit

151: Classifier

FIG. 1 is a block diagram of an embodiment of the biometric identification system of the present invention.

FIG. 2 shows a flow chart of the biometric identification method of the present invention.

Figure 3 shows a block diagram of an embodiment of a CNN and a classifier.

FIG. 1 shows an embodiment of the biometric identification system of the present invention. The identification system 10 includes a sensor 11 , a first feature extraction unit 12 , a second feature extraction unit 13 , a memory 14 , and a recognition unit 15 . The sensor 11 may be an image sensor for sensing a first biometric feature to generate an image (image) A. The first biometric feature may be, for example, a fingerprint, a human face, a palm print or an iris. The first feature extraction unit 12 , the second feature extraction unit 13 and the identification unit 15 can be realized by software or hardware. The first feature extraction unit 12 is coupled to the sensor 11, and generates a first information according to the image A provided by the sensor 11, and the first information describes the uniqueness of the first biological feature. The first information includes multiple sets of feature vectors. In one embodiment, the first feature extraction unit 12 may use a computer vision method to extract features of the image A to generate the first information. The computer vision method can be, for example, Features From Accelerated Segment Test (FAST), Adaptive and generic corner detection based on the accelerated segment test (AGAST), scale-invariant feature transformation ( Scale-invariant feature transform, SIFT), accelerated robustness Algorithms of Speeded Up Robust Features (SURF), KAZE or AKAZE, etc. In another embodiment, the first feature extraction unit 12 is a trained deep learning model, which may be implemented with a model architecture based on, for example, a convolutional neural network (Convolutional Neural Network, CNN) 131 . The method used by the first feature extraction unit 12 may be, but not limited to, Local Binary Patterns (LBP), Local Phase Quantization (LPQ) or Histogram of Oriented Gradient (HOG) And so on algorithm.

The second feature extraction unit 13 is coupled to the sensor 11 and generates a second information according to the image A, the second information describes the authenticity of the first biological feature. In one embodiment, the second feature extraction unit 13 may use a computer vision method to extract features of the image A to generate the second information. The method of the computer vision may be using, for example, accelerated segment test features (Features From Accelerated Segment Test, FAST), adaptive general accelerated segmentation detection (Adaptive and generic corner detection based on the accelerated segment test, AGAST), scale invariant feature conversion (Scale-invariant feature transform, SIFT), accelerated robust features (Speeded Up Robust Features, SURF), KAZE or AKAZE and so on. In another embodiment, the second feature extraction unit 13 is a pre-trained deep learning model, which can be implemented, for example, based on a convolutional neural network (Convolutional Neural Network, CNN) 131 or its improved model architecture, Or it can be AlexNet, MobileNet and other deep network models. The method used by the second feature extraction unit 13 may be, but not limited to, Local Binary Patterns (LBP), Local Phase Quantization (LPQ) or Histogram of Oriented Gradient (HOG) And so on algorithm. In one embodiment, the second information includes an embedding feature of the image, and the embedding feature is the vector data generated when the CNN 131 performs data conversion according to the image A, and is used to describe the first biological feature. A true-false property, which is not just a value representing true (1) or false (0).

Although both the first feature extraction unit 12 and the second feature extraction unit 13 are used for feature of extraction. But the purposes of the two are not the same, and the coefficients used to extract features are also different. The first feature extraction unit 12 and the second feature extraction unit 13 can be understood as looking at the image A from different angles. The first feature extraction unit 12 is used to describe the uniqueness of the first biological feature shown in the image A, such as the distribution relationship of some specific feature points. This uniqueness can be used to distinguish different individuals. The second feature extraction unit 13 is used to describe the true and false characteristics of the first biological feature shown in the image A, and is usually used to judge whether the first biological feature comes from a living body.

The memory 14 is respectively coupled to the first feature extraction unit 12 , the second feature extraction unit 13 and the identification unit 15 for storing a first template information and a second template information. The first template information and the second template information are respectively generated by the first feature extraction unit 12 and the second feature extraction unit 13 in the user's enrollment process. In the registration process, the sensor 11 senses a second biological feature of a user to be registered to generate an image B (not shown in the figure), wherein the first biological feature and the second biological feature are the same Kind of biometrics. The first feature extraction unit 12 generates first template information according to the image B, and the first template information describes the uniqueness of the second biological feature. The second feature extraction unit 13 generates the second template information according to the image B, and the second template information describes the true and false characteristics of the second biological feature.

The recognition unit 15 is coupled to the first feature extraction unit 12 , the second feature extraction unit 13 and the memory 14 . The identification unit 15 generates an identification result according to the first information, the second information or the first information and the second information, and the identification result represents pass or fail of identity authentication. In one embodiment, the recognition unit 15 includes a classifier 151 . The classifier 151 is based on a model trained by Machine Learning, which can be implemented by a Support Vector Machine (Support Vector Machine, SVM) or a neural network (Neural Network, NN). The classifier 151 can be a software or a hardware circuit.

The following uses fingerprint identification as an example to illustrate the operation of the identification system 10 of the present invention, but the present invention is not limited to fingerprint identification, and the present invention can also be applied to other biometric identification, such as face recognition recognition, iris recognition and palmprint recognition, etc. When the identification system 10 performs the registration process, the registrant puts his finger on the sensor 11. In this embodiment, the sensor 11 is a fingerprint sensor, which can be an optical fingerprint sensor or a fingerprint sensor. Capacitive fingerprint sensor. The sensor 11 senses the fingerprint of the finger (equivalent to the aforementioned second biometric feature) to generate a fingerprint image Fi1. The first feature extraction unit 12 generates first template information En1 according to the fingerprint image Fi1. The first template information En1 can be understood as describing the fingerprint characteristics of the fingerprint image Fi1. The texture of each person's fingerprint is unique and different from others. The first template information En1 is describing the uniqueness of the fingerprint. The second feature extraction unit 13 generates second template information En2 according to the fingerprint image Fil. The second template information En2 describes the authenticity of the fingerprint image Fi1. After the first template information En1 and the second template information En2 are stored in the memory 14, the registration procedure is completed.

When the identification system 10 performs the verification procedure, after the authenticator puts his finger on the sensor 11, the sensor 11 will sense the fingerprint of the finger (equivalent to the aforementioned first biometric feature) to generate a fingerprint Like Fi2, step S10 in FIG. 2 . The first feature extraction unit 12 generates the first information Ve1 according to the fingerprint image Fi1, as shown in step S12 of FIG. 2 . The second feature extraction unit 13 generates second information Ve2 according to the fingerprint image Fi2, as shown in step S14 of FIG. 2 . The second information Ve2 describes the authenticity of the fingerprint image Fi2. Finally, the identification unit 15 generates an identification result Vre according to the first information Ve1, the second information Ve2 or the first information Ve1 and the second information Ve2 to indicate the authentication success or failure, as shown in step S16 of FIG. 2 .

In one embodiment, the recognition unit 15 generates a recognition result Vre according to the difference D1 between the first information Ve1 and the first template information En1 and the difference D2 between the second information Ve2 and the second template information En2. The recognition unit 15 can use but not limited to one of Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, normalized Euclidean distance, Mahalanobis distance and Hamming distance to judge the first information Ve1 and the first template The difference D1 of the information En1 and the difference D2 between the second information Ve2 and the second template information En2 are judged, and then the difference D1 and D2 are multiplied by the weights W1 and W2 respectively, and then added to generate a sum, wherein the weights W1 and W2 are non-zero value. If the sum is less than a preset value TH1, the generated identification result Vre is a value "1", indicating that the authentication is successful. On the contrary, if the sum is greater than the preset value TH1, then If the generated recognition result Vre is a value of "0", it means that the authentication is failed.

In another embodiment, the identification unit 15 includes a classifier 151, which can be implemented by hardware or software, and uses machine learning methods for classification. The classifier 151 can perform a recognition step to generate a recognition result Vre. The recognition step includes but not limited to judging the similarity between the combination of the first information Ve1 and the second information Ve2 and the combination of the first template information En1 and the second template information En2 to generate a recognition result Vre. If the similarity is greater than a preset value TH2, the generated recognition result Vre is a value "1", which means the authentication is successful; otherwise, the generated recognition result Vre is a value "0", which means the authentication has failed.

In one embodiment, the identification unit 15 also includes a classifier 151. The identification unit 15 first uses but is not limited to Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, normalized Euclidean distance, Mahalanobis distance and Hamming One of the distances is used to determine whether the first information Ve1 is similar to the first template information En1. If not, the recognition unit 15 generates a recognition result Vre of "0", indicating that the authentication fails. If yes, the recognition unit 15 then uses the classifier 151 to perform the aforementioned recognition steps to generate a recognition result Vre. The advantage of doing this is that resources can be saved. If it is determined that the first information Ve1 is not similar to the first template information EN1, it is enough to determine that the person to be verified is different from the registered user. It is therefore not necessary to continue this identification step.

In another embodiment, the recognition unit 15 also includes a classifier 151. The recognition unit 15 first uses but is not limited to Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, normalized Euclidean distance, Mahalanobis distance and Han One of the clear distances is used to determine whether the second information Ve2 is similar to the second template information En2. If not, the recognition unit 15 generates a recognition result Vre of "0", indicating that the authentication fails. If yes, the recognition unit 15 then uses the classifier 151 to perform the aforementioned recognition steps to generate a recognition result Vre.

If the first feature extraction unit 12 uses a trained deep learning model, it needs to be trained in advance to obtain the ability of feature extraction. The following uses fingerprint recognition as an example. During the training process, fingerprint images of many different people are provided to a training program T1. The training program T1 has the same model structure as that of the first feature extraction unit 12 . The training program T1 is based on the fingerprint images of these different people and Its corresponding owner obtains a set of coefficients for extracting the features of the fingerprint image. This process is actually to tell the training program T1 which fingerprint image represents whom, and let the training program T1 learn how to classify. The first feature extraction unit 12 uses the set of coefficients to operate to extract the features of the fingerprint image to generate the first information. This first information describes the uniqueness of this fingerprint, and it can also be understood as describing what this fingerprint looks like in a mathematical way. If the first feature extraction unit 12 uses a computer vision method to extract features, it uses defined features to describe what the biological feature looks like, that is, its uniqueness.

If the second feature extraction unit 13 uses a trained deep learning model, it also needs to be trained in advance to obtain the ability of feature extraction. The following uses fingerprint recognition as an example. During the training process, a large number of real fingerprint images and fake fingerprint images need to be prepared and provided to the training program T2, which has the same model structure as the first feature extraction unit 13 . The aforementioned large number of real fingerprint images are obtained by sensing the fingerprints of many different people, all of which are fingerprint images obtained from living bodies. These fingerprint images are formed on materials such as silicon gel and sensed to obtain the aforementioned large number of fake fingerprint images. By telling the training program T2 which are real fingerprint images and which are fake fingerprint images, the training program T2 can obtain a set of coefficients for identifying real fingerprints or fake fingerprints after learning. This process is actually allowing the training program T2 to learn how to identify true and false fingerprints. The second feature extraction unit 13 uses the set of coefficients to extract features of the fingerprint image to generate second information. It is also possible to directly judge whether it is a real fingerprint or a fake fingerprint with this second information, but the present invention does not do this. The second feature extraction unit 13 of the present invention mainly obtains a set of feature values generated after the feature extraction and before judging the true and false fingerprints as the second information. Therefore, the second information describes the authenticity of the fingerprint, and it can also be understood as describing the authenticity of the fingerprint in a mathematical way. If the first feature extraction unit 12 uses a computer vision method to extract features, it uses already defined features to describe the true and false characteristics of the biological feature.

By providing a large number of images in advance to the first feature extraction unit 12 and the second feature extraction unit 12 The fetching unit 13 can obtain a large amount of first information and second information for the machine learning model of the classifier 151 to learn and classify. So that the classifier 151 has the ability to judge the success or failure of identity verification. Taking fingerprint recognition as an example, it is necessary to prepare a training program T3 having the same model structure as that of the classifier 151 . Next, a large number of combinations of the first information and the second information are provided to the training program T3, and the training program T3 is told which combinations are verified successfully and which combinations are verified failed. For example, tell the training program T3 to use the first information generated by the real fingerprint image and the second information to represent the verification success (pass), and to use the first information generated by the real fingerprint image and the second information generated by the fake fingerprint to represent the verification failure (fail). ). From this process, the training program T3 can learn how to judge whether the combination of the first information and the second information represents the verification success or failure.

As can be understood from the above training and identification descriptions, the present invention uses the first information generated by the first feature extraction unit 11 to judge whether the currently sensed fingerprint is similar to the registered fingerprint and through the second feature extraction unit 12 The generated second information is used to judge whether the authenticity of the currently sensed fingerprint is close to the authenticity of the registered fingerprint to determine whether the identity verification is passed.

FIG. 2 provides an embodiment to illustrate the architecture of the CNN 131 , which mainly includes a feature extraction part 1311 and a classification part 1312 . The image A is processed by the feature extraction part 1311 and the classification part 1312, and the embedded feature information is generated in the classifier 1312. The embedded feature information represents the true and false characteristics of the biological feature (such as a fingerprint) shown in the image A . The embedded feature information may be a set of values, such as "1101000". Please note that the second feature extraction unit 13 does not use the classification part 1312 to classify to generate a true or false recognition result, but obtains the embedded feature information generated by the classifier 1312 . If a feature extraction unit is used to directly determine whether a biometric feature to be verified is true or false, there may be a problem of accuracy. Taking fingerprint recognition as an example, if fake fingerprints of 10 materials are provided to train the feature extraction unit, for fake fingerprints other than these 10 materials, the feature extraction unit cannot accurately judge whether it is true or false. The second feature extraction unit 13 of the present invention does not directly judge whether the biological feature is true or false, but obtains the embedded feature information describing the true and false characteristics of the biological feature to follow the registered second template information in the memory for comparison. Therefore, for the new material that has not been learned by the second extraction feature unit 13, for the identity of the present invention The impact on the accuracy of validation is low.

The above descriptions of the preferred embodiments of the present invention are for the purpose of illustration, and are not intended to limit the present invention to the disclosed form. It is possible to modify or change based on the above teachings or learning from the embodiments of the present invention. The embodiment is selected and described in order to explain the principle of the present invention and to allow those familiar with the art to use the present invention in various embodiments for practical application. Decide.

10: Identification system

11: Sensor

12: The first feature extraction unit

13: The second feature extraction unit

131: Convolutional Neural Networks

14: Memory

15: Identification unit

151: Classifier

Claims (20)

A biometric identification system, comprising: a sensor for sensing a first biometric to generate an image; a first feature extraction unit coupled to the sensor to generate a first a piece of information, the first piece of information describes the uniqueness of the first biological feature; a second feature extraction unit, coupled to the sensor, the second feature extraction unit generates a second piece of information according to the image, the second The information describes the true and false characteristics of the first biological feature; an identification unit, coupled to the first feature extraction unit and the second feature extraction unit, generates a recognition result; and a memory, coupled to the first feature extraction unit, the second feature extraction unit and the recognition unit, for storing a first template information and a second template information; wherein, the recognition unit includes a A classifier, the classifier is used to perform a recognition step to judge the similarity between the combination of the first information and the second information and the combination of the first template information and the second template information, so as to generate the recognition result. The identification system according to claim 1, wherein the identification result represents passing or failing the identity authentication. The identification system according to claim 1, wherein the sensor is a fingerprint sensor, and the first biometric feature is a fingerprint. The recognition system according to claim 1, wherein the classifier includes a support vector machine or a neural network. The recognition system of claim item 1, wherein the recognition unit further includes first judging whether the first information is similar to the first template information, and if so, the recognition unit uses the classifier to perform Perform the identification step. The recognition system according to claim 1, wherein the recognition unit further includes first determining whether the second information is similar to the second template information, and if so, the recognition unit uses the classifier to perform the recognition step. The identification system according to claim 5 or 6, wherein the identification unit further includes a judgment using one of Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, normalized Euclidean distance, Mahalanobis distance and Hamming distance Whether the first information is similar to the first template information or whether the second information is similar to the second template information. The recognition system according to claim 1, wherein the first feature extraction unit or the second feature extraction unit includes a deep learning model. The recognition system according to claim 1, wherein the first feature extraction unit or the second feature extraction unit includes a convolutional neural network. The identification system according to claim 1, wherein the first feature extraction unit or the second feature extraction unit uses accelerated segment test features, adaptive general accelerated segmentation detection, scale invariant feature conversion, accelerated robust features, KAZE, AKAZE, local Binary mode, local phase quantization or histogram of oriented gradient algorithm to obtain the first information or the second information. A biological feature identification method, comprising the following steps: A. sensing a first biological feature to generate an image; B. generating a first information according to the image, wherein the first information describes the first biological feature Uniqueness; C. Obtain a second information from the image, wherein the second information describes the true and false characteristics of the first biological feature; D. Obtain a first template information and a second template information from a memory; as well as E. Generate a recognition result according to the second information or the first information and the second information; wherein, the step E includes using a classifier to perform a recognition step to determine the combination of the first information and the second information and the similarity between the combination of the first template information and the second template information to generate the recognition result. The identification method as claimed in claim 11 further includes judging whether to pass or fail the identity authentication according to the identification result. The identification method according to claim 11, wherein the first biometric feature is a fingerprint. The identification method of claim 11 further includes implementing the classifier with a support vector machine or a neural network. The identification method according to claim 11, wherein the step E further includes: judging whether the first information is similar to the first template information; and if so, performing the identification step. The identification method according to claim 11, wherein the step E further includes: judging whether the second information is similar to the second template information; and if so, performing the identification step. The identification method of claim item 15 or 16, wherein the step E further includes using one of Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, standardized Euclidean distance, Mahalanobis distance and Hamming distance to judge Whether the first information is similar to the first template information or whether the second information is similar to the second template information. The identification method as claimed in item 11, wherein the step B includes using accelerated segment test features, adaptive general accelerated segmentation detection, scale invariant feature conversion, accelerated robust features, KAZE, AKAZE, local binary mode, local phase quantization or direction Gradient histogram algorithm to obtain the first information. The identification method as claimed in item 11, wherein the step B includes using a deep learning model to obtain the first information or the second information. The identification method according to claim 11, wherein the step B includes using a convolutional neural network to obtain the first information or the second information.

Images