A finger vein forgery detection model and method based on fractional weighted fusion

Abstract

The application belongs to the technical field of image recognition, and specifically discloses a finger vein forgery detection model and method based on score weighted fusion, and the training steps of the model include the following steps: after being printed or secondarily collected, the pseudo image will generate additional noise information, therefore, a bilateral filter denoising algorithm and a mean filter denoising algorithm are adopted to extract noise features of the pseudo image and the real image, then two types of noise features and original texture features extracted based on a local binary pattern (LBP) are adopted, finally, a support vector machine (SVM) is used to classify the two types of noise features and the original texture features respectively, the classification results are subjected to score weighted fusion, and finally, a classification result model can be obtained, and then the classification result model is used to recognize the image to be recognized.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology and relates to a finger vein forgery detection model and method based on fractional weighted fusion. Background Technology

[0002] Finger vein recognition technology uses infrared light near a specific wavelength to illuminate the finger and then uses the transmitted light to create an infrared image of the veins. As a biometric identification technology, finger vein recognition has advantages such as low cost, high accuracy, and uniqueness in live identification. However, like other biometric technologies, finger vein recognition also faces the threat and risk of spoofing attacks. Spoofing attacks refer to the act of presenting a forged feature sample to a biometric capture device to deceive the system and gain unauthorized access. Currently existing spoofing attacks mainly include printing attacks, using smart devices to display fake fingers, and using simulated fingers made of beeswax.

[0003] To address the aforementioned spoofing attacks, researchers have proposed various Presentation Attack Detection (PAD) methods. The mainstream solutions can be categorized into two types: methods based on liveness detection and methods based on texture analysis. Liveness detection methods primarily capture liveness signals by adding additional equipment. These methods offer high accuracy but require significant hardware support or consume substantial system resources. Texture analysis-based methods are simpler to implement and lower in cost. They mainly achieve this by designing a feature that can distinguish between genuine and fake finger veins, as forged samples generate a lot of noise information during the creation process that is inconsistent with genuine samples. Existing texture analysis-based methods include the Fourier Spectral Energy Ratio (FSER), Fourier Spectral Bandwidth Energy (FSBE), Binary Statistical Image Features (BSIF), Steerable Pyramids (SP), Multi-Scale Histogram of Oriented Gradients (MHOG), and Windowed DMD (W-DMD). However, these existing texture analysis-based methods typically rely on a single texture feature for spoofing detection, resulting in limited detection performance and low versatility. Summary of the Invention

[0004] Unlike other biometric images, finger vein images are imaged using infrared light sources. The veins appear as shadows with low grayscale values, while the non-textured areas of the finger are brighter with higher grayscale values. To address the shortcomings in accuracy and versatility of existing finger vein counterfeit detection technologies, this application utilizes the aforementioned characteristics of finger vein images and integrates multiple texture features to propose a fractionally weighted fusion-based finger vein counterfeit detection method, thereby improving the performance and generalization of counterfeit detection.

[0005] This invention proposes a finger vein forgery detection model and method based on weighted fusion, which weights and fuses the classification results of bilateral filtering noise features, mean filtering noise features, and original texture features. The basic principle of this method is as follows: forged images generate additional noise information after printing or secondary acquisition. Therefore, bilateral filtering and mean filtering denoising algorithms are used to extract noise features from the forged and original images, respectively. Then, two types of noise features and original texture features extracted based on Local Binary Pattern (LBP) are used. LBP is an operator used to describe the local texture features of an image; its function is to extract local texture features of the image. Finally, Support Vector Machines (SVM) are used to classify the extracted two types of noise features and original texture features respectively. The classification results are then weighted and fused to obtain the final classification result model.

[0006] Firstly, this invention provides a finger vein forgery detection model based on score-weighted fusion, such as... Figure 1 As shown, the model is trained using the following steps:

[0007] Real finger vein images are collected using infrared acquisition devices as the real dataset. Fake finger vein models are created, for example, using laser film. Then, fake finger vein images are collected again using acquisition devices to form a fake dataset. The real dataset and the fake dataset are merged to form the true and false datasets. Finally, the true and false datasets are divided into training set and validation set, and labeled according to their authenticity.

[0008] For each original image I in the real and fake datasets, feature extraction is performed three times. It should be noted that the original image I here includes both real finger vein images and fake finger vein images in the real and fake datasets. The first feature extraction is to perform LBP feature extraction on the original image I to obtain the original texture feature F1 of the image. The second feature extraction is to use the bilateral filtering algorithm to denoise the original image I to obtain the first denoised image I1, then subtract I from I1 to obtain the first noisy image N1, and perform LBP feature extraction on N1 to obtain the first noise feature F2. The third feature extraction is to use the mean filtering algorithm to denoise the original image I to obtain the second denoised image I2, then subtract I from I2 to obtain the second noisy image N2, and perform LBP feature extraction on N2 to obtain the second noise feature F3.

[0009] The original texture features F1, the first noise feature F2, and the second noise feature F3 obtained after three feature extractions of each original image I in the training set are input into a support vector machine for training, resulting in three pre-trained SVM models. Then, the original texture features F1, the first noise feature F2, and the second noise feature F3 obtained after three feature extractions of each original image I in the validation set are input into the corresponding pre-trained SVM models, resulting in three classification results R1, R2, and R3. By weighting the three classification results, the optimal classification weights are obtained through an exhaustive algorithm, resulting in a score-weighted fusion model, i.e., a finger vein spoofing detection model based on score-weighted fusion.

[0010] Secondly, the present invention provides a finger vein spoofing detection method based on fractional weighted fusion, that is, applying the fractional weighted fusion model to identify genuine and fake finger veins. The detection method includes the following steps: performing feature extraction three times on the finger vein image to be identified according to the above feature extraction method, extracting the original texture feature F1', the first noise feature F2', and the second noise feature F3' of the finger vein image to be identified, respectively, and inputting them into the above three pre-trained SVM models to obtain three classification results R1', R2', and R3', and then inputting the three classification results R1', R2', and R3' into the above-trained fractional weighted fusion model for weighted decision, and finally outputting the obtained finger vein spoofing detection result.

[0011] Beneficial effects:

[0012] (1) The discrimination method is simple. After the model is trained, no additional hardware equipment is needed. The authenticity can be determined by directly inputting a single finger vein image. The hardware cost and computing cost are low.

[0013] (2) The present invention performs high-intensity denoising on genuine and fake images, and then extracts the noise features of genuine and fake finger veins by subtracting them from the original image. The noise features extracted in this way have high distinguishability and can improve the accuracy of counterfeit detection.

[0014] (3) Compared with existing finger vein spoofing detection methods that mostly use a single feature for detection, this invention performs a weighted fusion of the classification results of the original texture features and the classification results of two noise features. For low-frequency information such as skin texture and joint texture present in real finger vein infrared images, a bilateral filtering algorithm is used to remove these low-frequency information while retaining high-frequency information such as finger vein patterns and finger edges. By subtracting the original image from the denoised image, the low-frequency information in the real and fake finger vein images can be extracted. The low-frequency information of the fake image consists of unevenly distributed noise and patches. The pixel differences between the real and fake finger vein infrared images are different, and the mean filtering algorithm can make the overall image more blurred and smooth, and the pixel values ​​more uniform. The noise features obtained by subtracting the original image from the denoised image obtained by the mean filtering algorithm can amplify the difference in the fine granularity of the imaging between the real and fake images. Therefore, the weighted fusion of the classification results of the original texture features and the classification results of the two noise features can improve the performance and generalization of spoofing detection. Attached Figure Description

[0015] Figure 1 The flowchart shows the finger vein forgery detection method based on score-weighted fusion.

[0016] Figure 2 This is a schematic diagram of the LBP operator described in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram illustrating the process of obtaining a denoised image by bilateral filtering of the original image and obtaining a noisy image by subtraction in this invention.

[0018] Figure 4 This is a schematic diagram illustrating the process of obtaining a denoised image by mean filtering of the original image and obtaining a noisy image by subtraction in this invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] The following is Example 1, specifically using print attack spoofing identification as an example to illustrate the finger vein spoofing attack detection method. The specific steps are as follows:

[0021] Step S1: Data acquisition and preprocessing.

[0022] A finger vein acquisition device was built using a 2-megapixel high-definition camera and an 850nm infrared light source. Real finger vein images were acquired from 40 volunteers, with six images collected from each of their left and right index, middle, and ring fingers, resulting in a real dataset. Each image in this dataset was labeled with a real label. For each real image, it was printed onto two laser films. The veins on the two films were aligned and then superimposed on a piece of white paper and fixed to create a fake finger vein model. This fake model was then placed on the same acquisition device and subjected to the same parameter conditions for finger vein image acquisition, resulting in a fake dataset. Each image in this dataset was labeled with a fake label. Real images and their corresponding fake images were combined one-to-one; each pair of real and fake images should remain either part of the training set or the validation set in subsequent partitioning processes. Finally, the real and fake datasets were randomly divided into training and validation sets at an 8:2 ratio.

[0023] Step S2: Extraction of original texture features based on LBP.

[0024] For each original image I in the real and fake datasets, LBP feature extraction is performed to obtain the original texture features F1 for each sample. Figure 2 For example, a 3×3 sliding window is defined to traverse the entire image. The center pixel value of the window is selected as the threshold. In a 3×3 neighborhood window centered on the current pixel, if the pixel value of its neighbors is greater than or equal to the current pixel value, the pixel value at that neighborhood position is set to 1; otherwise, it is set to 0. Finally, taking the top-left pixel of the neighborhood window as the first value, the values ​​of each neighboring pixel are saved clockwise as the LBP code of the center pixel. Figure 2 In the image, the LBP code of the center pixel is "01111010". The LBP code of each pixel in the image is calculated, and the LBP codes of all pixels are used as their respective LBP features.

[0025] Step S3: Noise feature extraction based on bilateral filtering.

[0026] A bilateral filtering algorithm is used to denoise the original images of both genuine and fake datasets, resulting in a denoised image. The difference between the original image and the denoised image is then calculated to extract the noise image. Finally, following step S2, LBP feature extraction is performed on the noise image to obtain the bilateral filtered noise features. The bilateral filtering algorithm can remove low-frequency information from the image, such as skin texture and noise, while preserving high-frequency information, such as finger vein patterns and finger edges. By subtracting the original image from the denoised image, low-frequency information in genuine and fake finger vein images can be extracted. The low-frequency information in genuine images mainly consists of skin texture and joint texture, while the low-frequency information in fake images consists of unevenly distributed noise and patches. Therefore, this noise image makes it easier to distinguish between the two types of samples.

[0027] The specific process of applying the bilateral filtering algorithm to denoise the original image to obtain the denoised image I1 is as follows: Define a convolution window H, where p is the center pixel of the convolution window H, and its position is (p x ,p y ), where q is another pixel in H, and its position is (q x ,q y ), O p O q Let p and q be the pixel values, respectively. O can be calculated using formula (1). p The new pixel value O' after the bilateral filtering algorithm p ,

[0028]

[0029] Where σ s It is the time-domain parameter of the spatial kernel, σ r It is the scale kernel parameter, when σ s =7、σ r The noise features extracted when W = 0.2 have high discriminative power. p It is a normalization factor and W p The specific calculation method is shown in formula (2):

[0030]

[0031] After performing bilateral filtering on each pixel of the original image to obtain a denoised image I1, the difference between the original image I and the denoised image I1 is used to obtain the noisy image N1, i.e., N1 = I - I1. A specific example of the original image I, the denoised image I1, and the noisy image N1 is shown below. Figure 3 As shown.

[0032] Finally, the same LBP feature extraction method as in step S2 is used to extract the noise feature F2 from the noisy image N1.

[0033] Step S4: Noise feature extraction based on mean filtering.

[0034] The mean filtering algorithm is used to denoise the original images in the real and fake datasets, resulting in a denoised image. The difference between the original image and the denoised image is then calculated to extract the noise image. Finally, following step S2, LBP feature extraction is performed on the noise image to obtain the mean-filtered noise features. The mean filtering algorithm makes the overall image smoother and more uniform in pixel values. The noise features obtained by subtracting the original image from the denoised image effectively represent the fine-grained imaging of real and fake images. Because adjacent pixel values ​​in a real image are relatively consistent, the mean filtering has a very subtle effect on the pixel value of each individual pixel, resulting in a high degree of fine-grained noise image. Conversely, adjacent pixel values ​​in a fake image are relatively chaotic, and the mean filtering has a relatively large effect on the pixel value of each individual pixel. Therefore, the noise image extracted through mean filtering is easier to distinguish between two samples. The idea behind the mean filtering algorithm is to replace the current pixel value with the mean of the values ​​of its neighboring pixels.

[0035] The specific process of using the mean filtering algorithm to denoise the original images of the real and fake datasets is as follows: Define a convolution window J, where k is the center pixel of the convolution window J, and its position is (k x ,k y Let l be another pixel in J, and its position is (l x ,l y ), O k O l Let k and l be the pixel values, respectively. Then, O can be obtained using formula (3). k The new value O' after mean filtering k .

[0036]

[0037] Where M represents the total number of pixels within the convolution window J. After applying the mean filtering denoising algorithm to each pixel of the original image, a denoised image I2 is obtained. The difference between the original image and the denoised image is used to obtain the noisy image N2, i.e., N2 = I - I2.

[0038] Finally, the noisy image N2 is subjected to the same LBP feature extraction method as in step S2 to obtain the noise feature F3. Specific examples of the original image I, the denoised image I2, and the noisy image N2 are as follows: Figure 4 As shown.

[0039] Step S5: Use SVM for classification prediction.

[0040] The F1, F2, and F3 values ​​of the training set images are input into an SVM with a radial basis function kernel, resulting in three trained SVM models. Then, the F1, F2, and F3 values ​​of the validation set images are input into the trained SVM models to obtain the original texture feature prediction result R1, the bilateral filter noise feature prediction result R2, and the mean filter noise feature prediction result R3.

[0041] SVM is an efficient linear classifier that performs well in handling small sample classification problems. Its principle is to find an optimal linear classifying surface in a high-dimensional space to separate samples and maximize the margin between different samples. Assuming the classification decision function of the vector group F to be classified is f(F), then we have formula (4).

[0042] f(F)=v·F+g (4)

[0043] Where v is the hyperplane used to separate the two classes of samples, and g is the offset constant. By solving the maximum splitting hyperplane problem on the feature F, the parameters v and g in formula (4) can be determined so that the samples have the maximum margin on the hyperplane, thus obtaining the classification decision function f(F).

[0044] In this invention, F1, F2, and F3 of the training set images are input. Taking the original texture feature F1 as an example, after adding the corresponding labels, the SVM training dataset {(e1,s1),...,(e...} is obtained. i ,s i ),…,(e N ,s N )}, where e i It is the original texture feature vector of the i-th finger vein in the training set, s i The true / false label representing the vein of the i-th finger, s i ∈{-1,1}, where 1 represents a real finger vein and -1 represents a fake finger vein, i = 1, 2, ..., N, and N is the total number of samples. The classification decision function f1(F) of the original texture features is obtained by solving the maximum segmentation hyperplane problem on the SVM training dataset.

[0045] Similarly, SVM is trained using F2 and F3 of the labeled training set images to obtain the classification decision function f2(F) for bilateral filtered noise features and the classification decision function f3(F) for mean filtered noise features.

[0046] Finally, the F1, F2, and F3 of the validation set images are used to obtain the classification results R1, R2, and R3 through the decision functions f1(F), f2(F), and f3(F), i.e., R1 = f1(F1), R2 = f2(F2), and R3 = f3(F3).

[0047] Step S6: Perform score weighted fusion on the classification results and find the optimal weight.

[0048] Weights w1, w2, and w3 are assigned to the prediction results R1, R2, and R3, and the final fused prediction result R can be expressed as formula (5).

[0049]

[0050] Where w1+w2+w3=1, 0≤w1,w2,w3≤1, and sign is the sign function.

[0051] Finally, the predicted result R is compared with the labels of the validation set, and the proportion of the number of predicted results that do not match the dataset labels is calculated out of the total number of samples, denoted as the sample error rate Δ(w1,w2,w3). Then Δ(w1,w2,w3) is a function of the weights w1, w2, and w3, which can be obtained by solving formula (6).

[0052] minΔ(w1,w2,w3)stw1+w2+w3=1,0≤w1,w2,w3≤1 (6)

[0053] The optimal weights can be obtained by solving the above equation using an exhaustive algorithm. With the optimal weights, they can be substituted into (5) to obtain the score-weighted fusion model.

[0054] Step S7: Apply the model to identify genuine and fake finger veins.

[0055] The system inputs the image of the finger vein to be identified, first obtains three feature vectors of the image through steps S2 to S4, then uses the pre-trained SVM model and the score-weighted fusion model obtained in steps S5 to S6 to make a weighted decision, and finally outputs the finger vein impersonation detection result.

[0056] To verify that the proposed method in this embodiment has better recognition performance in finger vein recognition, a comparative experiment of different recognition methods was conducted on the IDIAP FVD dataset (Tome P, Vanoni M, Marcel S. On the vulnerability of finger vein recognition to spoofing[A]. Proceedings of International Conference of the Biometrics Special Interest Group (BIOSIG)[C], 2014:1-10.). The IDIAP FVD dataset includes two types of images: Full and Cropped.

[0057] In the comparative experiment, except for omitting the data acquisition step in step S1 of Example 1, the specific settings of the remaining steps and parameters are the same as those in Example 1 above.

[0058] The performance evaluation metrics used to compare the experimental results are as follows:

[0059] (1) APCER (Attack Presentation Classification Error Rate): The percentage of fake biometrics that are incorrectly identified as legitimate users.

[0060] (2) BPCER (Bona Fide Presentation Classification Error Rate): The proportion of legitimate users who are incorrectly identified as having fake biometrics.

[0061] (3) ACER (Average Classification Error Rate): The mean of APCER and BPCER, and its calculation method is shown in Equation (7).

[0062]

[0063] For specific steps of the identification method used for comparison, please refer to the following published literature:

[0064] (1)MHOG see open literature: Ashari N, Ong TS, Connie T, et al.Multi-ScaleTexture Analysis for Finger Vein Anti-Spoofing[A].Proceedings of IEEEInternational Conference on Artificial Intelligence in Engineering andTechnology(IICAIET)[C],2021,DOI:10.1109 / IICAIET51634.2021.9574036.

[0065] (2) For HDWT and DDWT, see the published literature: Nguyen DT, ParkYH, Shin KY, et al. Fake finger-vein image detection based on Fourier and wavelet transforms[J]. DigitalSignal Processing, 2013, 23(5): 1401-1413.

[0066] The results are shown in Table 1. As can be seen from Table 1, the APCER, BPCER, and ACER of this invention are all lower than those of other methods, indicating that this method is superior to the comparative methods. This method is truly effective.

[0067] Table 1

[0068]

[0069] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A finger vein spoofing detection model based on score-weighted fusion, characterized in that, The detection model is trained through the following steps: Real finger vein images are collected using an infrared acquisition device as the real dataset. A fake finger vein model is created, and then the acquisition device is used to collect fake finger vein images to form a fake dataset. The real dataset and the fake dataset are merged to form the true and false datasets. Finally, the true and false datasets are divided into training set and validation set, and labeled according to their authenticity. For each original image I in the true and false datasets, three feature extractions are performed. The first feature extraction is to perform LBP feature extraction on the original image I to obtain the original texture feature F1. The second feature extraction is to use a bilateral filtering algorithm to denoise the original image I to obtain the first denoised image I1, then subtract I from I1 to obtain the first noisy image N1, and perform LBP feature extraction on N1 to obtain the first noise feature F2. The third feature extraction is to use a mean filtering algorithm to denoise the original image I to obtain the second denoised image I2, then subtract I from I2 to obtain the second noisy image N2, and perform LBP feature extraction on N2 to obtain the second noise feature F3. The original texture features F1, the first noise feature F2, and the second noise feature F3 obtained after three feature extractions of each original image I in the training set are input into a support vector machine for training, resulting in three pre-trained SVM models. Then, the original texture features F1, the first noise feature F2, and the second noise feature F3 obtained after three feature extractions of each original image I in the validation set are input into the corresponding pre-trained SVM models, resulting in three classification results R1, R2, and R3. By weighting the three classification results, the optimal classification weights are obtained through an exhaustive algorithm, resulting in a score-weighted fusion model, i.e., a finger vein spoofing detection model based on score-weighted fusion. The specific process of using the bilateral filtering algorithm to denoise the original image I to obtain the first denoised image I1 is as follows: Define a convolution window H, where p is the center pixel of the convolution window H, and its position is (p x ,p y ), where q is another pixel in H, and its position is (q x ,q y ), O p O q Let p and q be the pixel values ​​respectively; O can be calculated using formula (1). p New pixel value after bilateral filtering algorithm , (1) Where σ s It is the time-domain parameter of the spatial kernel, σ r It is the scale kernel parameter, W p It is a normalization factor and W p The specific calculation method is shown in formula (2): (2) The denoised image I1 is obtained by performing bilateral filtering on each pixel of the original image. The specific process of using the mean filtering algorithm to denoise the original images of the real and fake datasets is as follows: J is a convolution window, k is the center pixel of the convolution window J, and its position is (k x ,k y Let l be another pixel in J, and its position is (l x ,l y ), O k O l Let k and l be the pixel values ​​respectively; then O can be obtained through formula (3). k The new value after mean filtering ; (3) Where M represents the total number of pixels in the convolution window J; after applying the mean filtering denoising algorithm to each pixel of the original image, the denoised image I2 is obtained.

2. The finger vein spoofing detection model based on fractional weighted fusion as described in claim 1, characterized in that, The forged finger vein model was created using laser film.

3. The finger vein spoofing detection model based on fractional weighted fusion as described in claim 2, characterized in that, s s =7、s r =0.2。 4. The finger vein spoofing detection model based on fractional weighted fusion as described in claim 1, characterized in that, The specific steps for weighted fusion of classification results and finding the optimal weights are as follows: assign weights w1, w2, and w3 to the prediction results R1, R2, and R3. The final fused prediction result R can be expressed as formula (5). (5) Where w1+w2+w3=1, 0≤w1, w2, w3≤1, and sign is the sign function; Finally, the prediction result R is compared with the labels of the validation set, and the proportion of the number of prediction results that are not equal to the labels of the dataset is calculated as the sample error rate Δ(w1,w2,w3). Then Δ(w1,w2,w3) is a function of the weights w1, w2, and w3, which can be obtained by solving formula (6). (6) The optimal weights are obtained by solving the above formula through an exhaustive algorithm. The optimal weights are then input into formula (5) to obtain the score-weighted fusion model.

5. A finger vein forgery detection method based on fractional weighted fusion, characterized in that, The finger vein image to be identified is subjected to three feature extractions according to the feature extraction method described in claim 1, to extract the original texture feature F1' and the first noise feature F2' of the finger vein image to be identified, respectively. 、 The second noise feature F3' is input into the three pre-trained SVM models described in claim 1 to obtain three classification results R1', R2' and R3'. Then, the three classification results R1', R2' and R3' are input into the score-weighted fusion model described in claim 1 for weighted decision-making, and finally the finger vein forgery detection result is output.

Images