A finger vein counterfeiting detection method based on multi-feature fusion

Abstract

A finger vein counterfeit detection method based on multi-feature fusion uses the difference in imaging quality between genuine and fake vein images to fuse the noise features and fuzzy features of the image to detect counterfeit attacks. This method is simple to implement. The original finger vein image can be used for subsequent identity recognition while realizing authenticity identification, reducing the consumption of additional computer resources; compared with the finger vein counterfeit detection method that only uses a single image texture feature, this method fuses the noise features and fuzzy features of the vein image, which can increase the distinction between genuine and fake vein images and help improve the accuracy of finger vein counterfeit detection; compared with the traditional direct series or parallel multi-feature fusion method, this method uses weighted feature fusion, takes into account the importance of each feature, reduces the mutual influence between features, and improves the generalization ability of the classifier.

Description

Technical Field

[0001] The present invention belongs to the technical field of finger vein recognition, and in particular relates to a finger vein counterfeit detection method based on multi-feature fusion, which is used to improve the performance of counterfeit attack detection of a finger vein recognition system. Background Art

[0002] Finger vein recognition is a technology that uses the vein structure inside the human finger for identity recognition. Similar to other biometric recognition technologies, finger vein recognition technology is also threatened by spoofing attacks. The so-called spoofing attack means that the attacker uses forged biometric features to deceive the sensor and thus achieve the purpose of invading the system. Although the finger vein itself is an internal feature of the human body and has a certain potential to resist spoofing attacks, the database and related processing systems that store finger vein information still have the possibility of information leakage. Therefore, it is particularly necessary to improve the performance of the recognition system in resisting spoofing attacks.

[0003] At present, counterfeit attacks on finger vein recognition systems are mainly focused on printing attacks. Printing attacks refer to printing finger vein images on A4 white paper, snow-coated paper and other materials to forge a vein model, and then using this model to implement a deceptive attack on the recognition system. There are two main types of solutions to this type of printing attack threat: one is hardware-based and the other is software-based. The hardware-based method detects the activity or life signal of the finger to detect the true and false veins. Compared with the forged finger vein model, the real finger has more living features that can be used for identification, such as thermal infrared radiation features, blood flow features, and blood oxygen saturation features. This type of method has high reliability and recognition accuracy, but the disadvantages are: (1) it requires additional equipment and high hardware costs; (2) it requires continuous acquisition, storage and processing of a series of images to extract relevant information, which increases the system's computing resources, storage space and running time. The software-based method is relatively simple to implement. It only needs to analyze a single image and mainly considers the difference in texture features between true and false images. Such methods mainly include Fourier spectrum energy ratio method, Fourier spectrum bandwidth energy method, binary statistical image feature method and controllable pyramid method based on image texture analysis, etc. However, these methods generally extract a single feature from the image for counterfeit detection, and their performance is limited. Summary of the invention

[0004] In view of the above problems, the present invention fuses the noise features and blur features of the image, and proposes a finger vein counterfeit detection method based on multi-feature fusion. The main principle is that when the finger vein image is collected again after printing, additional noise information will be added to the image. In addition, due to the low printing resolution and the problem of camera defocus during secondary collection, the fuzziness of the image will also increase. Therefore, the difference between the true and false vein images is reflected not only in the noise distribution, but also in the degree of blur. By fusing these two features, the discrimination will be stronger and the recognition rate of counterfeit detection will be higher.

[0005] A finger vein counterfeiting detection method based on multi-feature fusion includes the following steps:

[0006] Step 1: For the input image I, perform differential filtering on the image in four directions: 0°, 45°, 90°, and 135°, and obtain the differential matrix M in the four directions: → , M ↑ and As the noise information of the image;

[0007] Step 2: For the input image I, the guided filtering method is used to extract the blur information of the image;

[0008] Step 3: Divide the result obtained in step 2 into 2*2 equal blocks to obtain four fuzzy information blocks;

[0009] Step 4: Extract LBP texture features from the difference matrices in the four directions obtained in step 1, and concatenate the results as the noise feature F1 of the vein image;

[0010] Step 5: Extract LBP texture features from the four fuzzy information blocks obtained in step 3, and concatenate the results to serve as the fuzzy feature F2 of the vein image;

[0011] Step 6: Linear weighted fusion of noise feature F1 and fuzzy feature F2;

[0012] Step 7: Training the classification model; using the support vector machine to train and classify the results after feature fusion, the particle swarm algorithm is used during the training process to optimize the feature weights to obtain the best weights and classification model;

[0013] Step 8: Apply the classification model; input the finger vein image to be detected into the classification model to obtain the result of counterfeit detection.

[0014] Furthermore, in step 1, the difference matrices M corresponding to the four directions are → , M ↑ and The calculation method is as follows:

[0015] M → (x,y)=I(x,y)-I(x+1,y) (1)

[0016]

[0017] M ↑ (x,y)=I(x,y)-I(x,y+1) (3)

[0018]

[0019] Among them, I(x,y)∈{0,1,2,...,255}, x and y represent the pixel positions of image I respectively.

[0020] Furthermore, step 2 includes the following sub-steps:

[0021] Step 2.1, for the input image I, set the guided image p of the guided filter equal to I, and record the filtered output image as q; assume that p and q are in the window w k There is a local linear relationship in the window, that is, q is p in the window w k The local linear transformation in which the window w k Taking pixel k as the center and r as the radius, the calculation of each pixel in q is shown in formula (5):

[0022]

[0023] Where i and k are pixel indices, α k and b k is the coefficient of the linear function, and in the local window w k is a constant;

[0024] Step 2.2, let the difference between I and q be E, then constrain q through equation (6), and solve the coefficient α after minimizing E k and b k :

[0025]

[0026] Where ε>0, ε is to prevent α from being too large k The introduced regularization parameter, I i represents a pixel i of the input image I, and i∈w k ;

[0027] The linear coefficients can be solved using the least squares method where u k and σ k 2 Respectively represent in window w kThe mean and variance in |w| is the window w k The number of pixels in Represents the input image I in window w k The mean value in ;

[0028] Step 2.3, when calculating the specific output of a certain pixel, take the average of all linear function values ​​including the pixel, and the final filtering output is shown in formula (7):

[0029]

[0030] in, w i are all windows containing pixel k, and pixel i is the center of the window; the result q of the filtering output will be used as the blurred information of the image.

[0031] Furthermore, in step 4 and step 5, the process of LBP feature extraction is to select a pixel point, and in a 3*3 neighborhood window centered on the pixel point, if the pixel value of its neighborhood is greater than or equal to the pixel value, the pixel value of the neighborhood bit is set to 1, otherwise it is set to 0; finally, the pixel in the upper left corner of the neighborhood window is taken as the first bit, and the values ​​of each neighborhood are saved in a clockwise direction as the LBP code of the central pixel point.

[0032] Furthermore, in step 6, the fused feature F is shown in formula (8):

[0033] F=z1F1+z2F2 (8)

[0034] Among them, z1 and z2 are the corresponding weights of F1 and F2, satisfying z1+z2=1, 0≤z1,z2≤1.

[0035] Furthermore, in step 7, the particle swarm algorithm is used to optimize z1 and z2 during the training process so that the detection performance of the model is optimal; the particle swarm algorithm is used to optimize the weights, and the process is as follows: after randomly initializing the particle swarm distribution, the counterfeit detection performance corresponding to each particle is evaluated within the number of iterations; then, each particle updates its optimal value according to the current fitness value and the historical optimal fitness value; then the optimal value of the entire population is updated; then, all particles update their own speed and position; finally, the above steps are repeated according to whether the number of iterations is reached; among which, the speed and position update formulas of the particle swarm algorithm are as follows.

[0036] V e (t+1)=wV e (t)+c1R1(p best –S e (t))+c2R2(g best -S e (t)) (9)

[0037] S e (t+1)=S e (t)+V e (t+1) (10)

[0038] Among them, V e (t) is the velocity of the e-th particle at time t, S e (t) is the position of the e-th particle at time t, p best Save the historical optimal position of the particle, g best The optimal position of all particles in the world is saved, w is the inertia weight, c1 and c2 are acceleration factors, R1 and R2 are random numbers in [0,1], particles are searched in one-dimensional space, the number of particles and the maximum number of iterations are set, the position and speed are limited to the interval [0,1] and [-0.2,0.2] respectively, and the fitness function is the recognition rate of the model; after training, the optimal feature fusion weights z1 and z2 are obtained, and the best classification model is saved.

[0039] Beneficial effects of the present invention:

[0040] (1) Simple implementation: the original finger vein image can be used for subsequent identity recognition while realizing authenticity identification, thus reducing the consumption of additional computer resources.

[0041] (2) Compared with the finger vein counterfeit detection method that only uses a single image texture feature, the present invention integrates the noise feature and fuzzy feature of the vein image, which can increase the distinction between true and false vein images and help improve the accuracy of finger vein counterfeit detection.

[0042] (3) Compared with the traditional direct serial or parallel multi-feature fusion method, the present invention uses weighted feature fusion, which takes into account the importance of each feature, reduces the mutual influence between features, and improves the generalization ability of the classifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flow chart of a finger vein counterfeiting detection method based on multi-feature fusion in an embodiment of the present invention.

[0044] Figure 2 Schematic diagram of the LBP operator in an embodiment of the present invention.

[0045] Figure 3 Flow chart of the particle swarm algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0047] The present invention is a finger vein counterfeiting detection method based on multi-feature fusion, and its process is as follows: Figure 1 As shown. Before extracting features, the present invention does not need to do any preprocessing. When extracting image noise information, the present invention introduces a differential filtering method to perform differential filtering on the image in four directions, and uses the obtained differential matrix as the noise information of the image. When extracting image fuzzy information, the present invention introduces a guided filtering method that can restore the real information of the image to a large extent to extract information. Taking into account the characteristics of finger vein images, such as the presence of highlight areas in the proximal and distal joints in the image, in order to better characterize this local structural characteristic, the obtained image fuzzy information is equally divided into 2*2 blocks. Then, for the extracted noise information and fuzzy information blocks, LBP is used to extract texture features respectively to obtain the noise features and fuzzy features of the image, and the linear weighted method is used to fuse these two features. In the training model stage, SVM is used as a classifier, and the weights of the feature weighted fusion are determined by optimizing the particle swarm algorithm. Finally, after obtaining the optimal weights and classification model, the finger vein image can be directly input for finger vein counterfeiting detection.

[0048] The main steps of this method are as follows.

[0049] Step 1: Use differential filtering to extract noise information from finger vein images.

[0050] For the input image I, the image is differentially filtered in four directions of 0°, 45°, 90° and 135°, and the corresponding differential matrices M in the four directions can be obtained. → , M ↑ and The calculation method of each of them is as follows.

[0051] M → (x,y)=I(x,y)-I(x+1,y) (1)

[0052]

[0053] M ↑ (x,y)=I(x,y)-I(x,y+1) (3)

[0054]

[0055] Among them, I(x,y)∈{0,1,2,...,255}, x and y represent the pixel positions of image I respectively.

[0056] Step 2: Use guided filtering to extract fuzzy information from finger vein images.

[0057] For the input image I, set the guided image p of the guided filter equal to I, and record the filtered output image as q. Assume that p and q are in the window w k There is a local linear relationship (w k With pixel k as the center and r as the radius), that is, q is p in window w k The local linear transformation in , then the calculation of each pixel in q is shown in formula (5).

[0058]

[0059] Where i and k are pixel indices, α k and b k is the coefficient of the linear function, and in the local window w k where is a constant. Let the difference between I and q be E. Then constrain q by equation (6). After minimizing E, the coefficient α can be solved. k and b k .

[0060]

[0061] Where ε>0, ε is to prevent α from being too large k The introduced regularization parameter, I i represents a pixel i of the input image I, and i∈w k The linear coefficients can be solved using the least squares method. where u k and σ k 2 Respectively represent in window w k The mean and variance in |w| is the window w k The number of pixels in Represents the input image I in window w k The mean value in .

[0062] Since a pixel belongs to multiple different windows, it means that each pixel can be described by multiple linear functions. Therefore, when calculating the specific output of a certain point, the average of all linear function values ​​including the pixel can be taken. The final filtering output is shown in formula (7).

[0063]

[0064] in, w i are all windows containing pixel k, and pixel i is the center of the window. The result q of the filtering output will be used as the blurred information of the image.

[0065] Step 3: Divide the blurred information obtained in step 2 into 2*2 equal blocks to obtain four blurred information blocks of image I.

[0066] Step 4: Perform LBP texture feature extraction on the four difference matrices obtained in step 1, and concatenate the results as the noise feature F1 of image I.

[0067] The calculation example of LBP feature extraction is as follows Figure 2 As shown in the figure. The process is to select a pixel point, and in a 3*3 neighborhood window centered on the pixel point, if the pixel value of its neighborhood is greater than or equal to the pixel value, the pixel value of the neighborhood bit is set to 1, otherwise it is set to 0. Finally, the pixel in the upper left corner of the neighborhood window is the first bit, and the values ​​of each neighborhood are saved in a clockwise direction as the LBP code of the central pixel point. Figure 2 In the example, the LBP code of the central pixel is: “10001110”. After calculating the LBP code for each pixel in the differential matrix, the LBP codes obtained for all pixels are used as the LBP features of the differential matrix.

[0068] Step 5: Perform LBP texture feature extraction on the four fuzzy information blocks obtained in step 3 (the process is the same as step 4), and concatenate the results as the fuzzy feature F2 of image I.

[0069] Step 6: Multi-feature fusion.

[0070] The noise feature F1 and the fuzzy feature F2 are fused using a linear weighted method to obtain the fused feature F, as shown in formula (8).

[0071] F=z1F1+z2F2 (8)

[0072] Among them, z1 and z2 are the corresponding weights of F1 and F2, satisfying z1+z2=1, 0≤z1,z2≤1.

[0073] Step 7: Train the classification model.

[0074] The SVM model is used as a classifier, and the particle swarm algorithm is used during the training process to optimize z1 and z2, so that the detection performance of the model is optimal. The process of using the particle swarm algorithm to optimize the weights is as follows: Figure 3 As shown. After randomly initializing the particle swarm distribution, the counterfeit detection performance of each particle is evaluated within the number of iterations. Then, each particle updates its optimal value based on the current fitness value and the historical optimal fitness value. Then the optimal value of the entire population is updated. After that, all particles update their own speed and position. Finally, the above steps are repeated according to whether the number of iterations is reached. Among them, the speed and position update formulas of the particle swarm algorithm are as follows.

[0075] V e (t+1)=wV e (t)+c1R1(p best –Se (t))+c2R2(g best -S e (t)) (9)

[0076] S e (t+1)=S e (t)+V e (t+1) (10)

[0077] Among them, V e (t) is the velocity of the e-th particle at time t, S e (t) is the position of the e-th particle at time t, p best Save the historical optimal position of the particle, g best The optimal position of all particles in the world is saved, the inertia weight w is set to 0.8, the acceleration factors c1 and c2 are set to 0.5, R1 and R2 are random numbers [0,1], the particles are searched in one-dimensional space, the number of particles is set to 20, the maximum number of iterations is set to 30, the position and speed are limited to the interval [0,1] and [-0.2,0.2] respectively, and the fitness function is the recognition rate of the model. After training, the optimal feature fusion weights z1 and z2 can be obtained, and the best classification model can be saved.

[0078] Step 8: Apply the classification model. Input the finger vein image to be detected into the classification model obtained in step 7 to obtain the result of finger vein counterfeit detection.

[0079] The invention is based on the idea of ​​image texture analysis, fully utilizes the difference between noise features and fuzzy features of authentic and fake finger vein images, performs weighted fusion of the two features, and improves the accuracy of finger vein counterfeiting detection.

[0080] The present invention uses the public IDIAP FVD dataset (Tome P, Vanoni M, Marcel S. On the vulnerability of finger vein recognition to spoofing [C] / / 2014 International Conference of the Biometrics Special Interest Group (BIOSIG). IEEE, 2014: 1-10.) to conduct experiments. The dataset includes two types of images: Full and Cropped. The performance evaluation indicators used are as follows:

[0081] (1) APCER (Attack Presentation Classification Error Rate): the proportion of counterfeit biometrics that are mistakenly identified as legitimate users;

[0082] (2) BPCER (Bona Fide Presentation Classification Error Rate): the proportion of legitimate users who are mistakenly identified as counterfeit biometrics;

[0083] (3)ACER (Average Classification Error Rate): The average of APCER and BPCER, which is calculated as shown in formula (11).

[0084]

[0085] The present invention uses a weighted method to fuse multiple features, and uses a particle swarm algorithm to optimize weights. The detection performance of the method of the present invention is compared with that of the method of multi-feature series fusion and multi-feature parallel fusion, and the results are shown in Table 1. As can be seen from the table below, the APCER, BPCER, and ACER of the present invention on both Full and Cropped data sets are all 0, and the results are better than multi-feature series fusion and multi-feature parallel fusion. This shows that the use of weighted fusion makes better use of the classification advantages of each feature itself, and the generalization performance of the classifier is improved.

[0086]

[0087] Then, the method of FSER, DDWT, HDWT, FSER-DWT and other methods proposed by Nguyen et al. (Nguyen DT, Park YH, Shin KY, et al. Fake finger-vein image detection based on fourier and wavelet transforms [J]. Digital Signal Processing, 2013, 23 (5): 1401-1413) are compared, and the results are shown in the following table. As can be seen from Table 2, the results of the present invention are better than these methods. This is because these methods mostly extract a single feature from the image to identify the authenticity of the finger vein, and it is difficult to completely distinguish the authentic finger vein from the fake one.

[0088]

[0089]

[0090] In summary, the detection performance of the present invention has obvious advantages. The fusion of noise features and fuzzy features increases the differentiation between true and false veins, and the counterfeit detection performance is better.

[0091] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A finger vein counterfeiting detection method based on multi-feature fusion, characterized by: The method comprises the following steps: Step 1: For the input image I, perform differential filtering on the image in four directions: 0°, 45°, 90°, and 135°, and obtain the differential matrix M in the four directions: → 、M ↗ 、M ↑ and M ↖ , as the noise information of the image; Step 2: For the input image I, the guided filtering method is used to extract the blur information of the image; Step 2 includes the following steps: Step 2.1, for the input image I, set the guided image p of the guided filter equal to I, and record the filtered output image as q; assume that p and q are in the window w k There is a local linear relationship in the window, that is, q is p in the window w k The local linear transformation in which the window w k Taking pixel k as the center and r as the radius, the calculation of each pixel in q is shown in formula (5): Where i and k are pixel indices, α k and b k is the coefficient, and in the local window w k is a constant; Step 2.2, let the difference between I and q be E, then constrain q through equation (6), and solve the coefficient α after minimizing E k and b k : Where ε>0, ε is to prevent α from being too large k The introduced regularization parameter, I i represents a pixel i of the input image I, and i∈w k ; The linear coefficients can be solved using the least squares method where u k and σ k 2 Respectively represent in window w k The mean and variance in |w| is the window w k The number of pixels in Represents the input image I in window w k The mean value in ; Step 2.3, when calculating the specific output of a certain pixel, take the average of all linear function values ​​including the pixel, and the final filtering output is shown in formula (7): in, w i are all windows containing pixel k, and pixel i is the center of the window; the result q of the filtering output will be used as the blurred information of the image; Step 3: Divide the result obtained in step 2 into 2*2 equal blocks to obtain four fuzzy information blocks; Step 4: Perform LBP texture feature extraction on the difference matrix in the four directions obtained in step 1, and concatenate the results as the noise feature F1 of the image; Step 5: Perform LBP texture feature extraction on the four fuzzy information blocks obtained in step 3, and concatenate the results as the fuzzy feature F2 of the image; Step 6: Linear weighted fusion of noise feature F1 and fuzzy feature F2; Step 7: Training the classification model; using the support vector machine to train and classify the results after feature fusion, the particle swarm algorithm is used during the training process to optimize the feature weights to obtain the best weights and classification model; Step 8: Apply the classification model; input the finger vein image to be detected into the classification model to obtain the result of counterfeit detection.

2. According to claim 1, a finger vein counterfeiting detection method based on multi-feature fusion is characterized by: In step 1, the difference matrix M corresponding to the four directions is → 、M ↗ 、M ↑ and M ↖ The calculation method is as follows: M → (x,y)=I(x,y)-I(x+1,y) (1) M ↗ (x,y)=I(x,y)-I(x+1,y+1) (2) M ↑ (x,y)=I(x,y)-I(x,y+1) (3) M ↖ (x,y)=I(x,y)-I(x-1,y+1) (4) Among them, I(x,y)∈{0,1,2,...,255}, x and y represent the pixel positions of image I respectively.

3. According to claim 1, a finger vein counterfeiting detection method based on multi-feature fusion is characterized by: In step 4 and step 5, the process of LBP feature extraction is to select a pixel point, and in a 3*3 neighborhood window centered on the pixel point, if the pixel value of its neighborhood is greater than or equal to the pixel value, the pixel value of the neighborhood is set to 1, otherwise it is set to 0; finally, take the pixel in the upper left corner of the neighborhood window as the first bit, and save the values ​​of each neighborhood in a clockwise direction as the LBP code of the central pixel point.

4. According to claim 1, a finger vein counterfeiting detection method based on multi-feature fusion is characterized by: In step 6, the fused feature F is shown in formula (8): F=z1 F1+z2 F2 (8) Among them, z1 and z2 are the corresponding weights of F1 and F2, satisfying z1+z2=1, 0≤z1,z2≤1.

5. According to claim 1, a finger vein counterfeiting detection method based on multi-feature fusion is characterized in that: In step 7, the particle swarm algorithm is used to optimize z1 and z2 during the training process so that the detection performance of the model is optimal; the particle swarm algorithm is used to optimize the weights, and the process is as follows: after randomly initializing the particle swarm distribution, the counterfeit detection performance corresponding to each particle is evaluated within the number of iterations; then, each particle updates its optimal value according to the current fitness value and the historical optimal fitness value; then the optimal value of the entire population is updated; then, all particles update their own speed and position; finally, the above steps are repeated according to whether the number of iterations is reached; among them, the speed and position update formulas of the particle swarm algorithm are as follows: V e (t+1)=wV e (t)+c1R1(p best -S e (t))+c2R2(g best -S e (t)) (9) S e (t+1)=S e (t)+V e (t+1) (10) Among them, V e (t) is the velocity of the e-th particle at time t, S e (t) is the position of the e-th particle at time t, p best Save the historical optimal position of the particle, g best The optimal position of all particles in the world is saved, w is the inertia weight, c1 and c2 are acceleration factors, R1 and R2 are random numbers in [0,1], particles are searched in one-dimensional space, the number of particles and the maximum number of iterations are set, the position and speed are limited to the interval [0,1] and [-0.2,0.2] respectively, and the fitness function is the recognition rate of the model; after training, the optimal feature fusion weights z1 and z2 are obtained, and the best classification model is saved.

Images