A real human finger vein detection method

Abstract

The present invention discloses a real human finger vein detection method. The present invention specifically includes the following steps: 1. Produce a data set, which is divided into a training set and a test set. In order to increase the diversity of the data set, normal clear images, blurred images, dirty and damaged images are added to the training set and the test set to simulate common image types in life; 2. Construct a real human finger vein detection model based on weighted differential convolution and gradient direction residual unit. The present invention uses local neighborhood relationships to mine detailed texture features in finger vein images that are robust to environmental changes, increase feature stability, and improve the generalization performance of the algorithm. At the same time, the present invention analyzes the gradient direction maps of low-quality real and false human finger vein images, and proposes a gradient direction residual unit structure based on the characteristics of differences in the detailed texture of the true and false sample materials, obtains gradient direction features through horizontal gradients and vertical gradients, and enhances the network gradient direction feature expression.

Description

Technical Field

[0001] The present invention mainly aims at solving the problems that low-quality images of human finger veins are greatly affected by environmental changes and the difference between true and false samples is insufficient. A real human finger vein detection method is proposed, which is a real human finger vein detection network based on deep learning combined with weighted differential convolution and gradient direction residual unit. Background Art

[0002] Finger vein detection is a very important part of the vein recognition system, but its performance is seriously affected by the quality of the finger vein image. Considering the user factors (such as peeling fingers, blurred vein fingers, dirty fingers, etc.) and environmental factors (such as dirty acquisition equipment mirror, etc.) in actual applications, the quality of finger vein imaging is low, and the image contains less effective vein information, which makes it very easy for criminals to forge fake human finger veins to deceive the system (for example, in the registration stage, fake finger veins made of special materials are recognized as real human veins and entered into the database, or non-veins are recognized as veins in the recognition stage), which seriously affects the security of the recognition system. This makes the study of real human finger vein detection algorithms for low-quality images a very critical task.

[0003] At present, the detection algorithm for human finger veins is still mainly based on traditional algorithms, but traditional real human finger vein detection algorithms generally have high design requirements for sensors and other equipment, and have high cost losses. There are also defects such as cumbersome feature extraction process, small number and single type. The detection algorithm based on deep learning convolutional neural network can continuously learn to extract features during the network learning iteration process, and the extracted finger vein image feature information is more objective and rich. Therefore, subsequent researchers mainly use deep learning algorithms for detection.

[0004] With the rapid development of deep learning, convolutional neural networks have been widely used in the detection field. For example, Nogueira et al. proposed a method of using CNN-Alexnet network for detection, which achieved good results in various data sets. However, this network structure is too shallow and cannot fully extract classification features when targeting low-quality finger vein images. Long Min et al. explored the inception structure and further optimized the AlexNet network, achieving good results in real human fingerprint detection. However, the above algorithms all use the traditional stacked convolutional layer structure and focus on extracting high-level semantic features. Although they contain more semantic information, they are weak in describing the detailed features between real human finger veins and fake finger veins. However, the vein information of low-quality finger vein images is scarce, and the semantic information difference between real and fake finger veins is insufficient, which makes the existing algorithms lack the feature difference between real and fake finger veins when targeting low-quality images. In addition, low-quality images are easily affected by environmental factors such as light intensity and direction. If the extracted detailed features are not robust, the generalization of the algorithm will be further reduced.

[0005] From the above, it can be seen that for the real human finger vein detection algorithm, it is necessary to make the proposed features environmentally robust and ensure the generalization ability of the algorithm, and also ensure that the proposed features can fully express the material differences between real human finger veins and fake finger veins. To this end, the present invention proposes a network combining weighted differential convolution and gradient direction residual unit. Summary of the invention

[0006] The purpose of the present invention is to provide a real human finger vein detection method in view of the deficiencies in the prior art.

[0007] The technical solution adopted by the present invention to solve the technical problem comprises the following steps:

[0008] Step 1: Create a data set. This data set is divided into two parts: a training set and a test set. In order to increase the diversity of the data set, normal clear images, blurred images, and dirty and damaged images are added to the training set and the test set to simulate common image types in life;

[0009] 1-1: The fake human finger vein image dataset uses laboratory personnel’s fingers to simulate various experimental types;

[0010] 1-2: The fake human finger vein image dataset uses prostheses of different materials to manually outline the vein image for each experimental type;

[0011] Step 2: Construct a real human finger vein detection model based on weighted differential convolution and gradient direction residual unit.

[0012] Furthermore, the vein real human body detection model is specifically implemented as follows:

[0013] First, the input finger vein image is normalized to 1×224×224. After passing through the 7×7 input convolution layer, the BN layer, and the ReLu layer, the network outputs a feature map with a size of 64×112×112. Then it passes through the gradient direction residual unit in sequence to enhance the finger vein gradient direction features, making the subsequent classification of the network more accurate. Finally, the classification is completed through the fully connected layer and the classification result is output, and back propagated after the standard cross entropy loss calculation.

[0014] Furthermore, all convolutional layers are weighted differential convolutions, and the gradient direction residual unit structure is used to enhance the extraction of the gradient direction features of the finger vein image by the network residual unit, thereby increasing the material difference between real and fake human finger veins.

[0015] Furthermore, the weighted differential convolution is implemented as follows:

[0016] First, the pixel value to be differentiated is obtained by performing weighted calculation on the original image, as shown in formula (1); finally, the original image and the pixel value to be differentiated p are calculated. wlg The difference result performs the normal convolution output feature Figure Ⅰ ;

[0017]

[0018] The original image is the normalized input finger vein image; in formula (1), p wlg represents the pixel value to be differentiated, that is, the result after weighted operation on the original image; p0 represents the center pixel point, p n Represents the neighborhood pixel value, and α is the weight coefficient of the center value.

[0019] Furthermore, the weighted differential convolution is further implemented as follows:

[0020] The specific usage formula (2) of weighted differential convolution is as follows:

[0021] y=θ·y wdc +(1-θ)·y contional (2)

[0022] In formula (2), the y output feature map represents the features of the weighted differential convolution output Figure Ⅰ ;y wdc represents the output result of weighted differential convolution, as shown in formula (3); contional represents the output result of traditional convolution, as shown in formula (4); θ is a hyperparameter used to adjust the proportion of traditional convolution and weighted differential convolution;

[0023]

[0024]

[0025] In formula (3), R represents all pixels in the current neighborhood, w(p i ) means that at pixel p i The weight coefficient at x(p i ) represents the input image at pixel p i The value at x(p wlg ) represents the expanded matrix of pixel values ​​to be differentiated, which is used to perform differential operations with the original image.

[0026] Furthermore, by substituting formula (3) and formula (4) into formula (2), we can obtain:

[0027]

[0028] Among them, the y output feature map represents the features of the weighted difference convolution output Figure Ⅰ .

[0029] Furthermore, the gradient direction residual unit is specifically implemented as follows:

[0030] The gradient direction residual unit consists of three branches: convolution branch, short-circuit connection branch and gradient direction branch; the convolution branch is mainly responsible for the input features Figure Ⅰ Perform convolution operations to obtain effective information; the short-circuit connection branch is mainly responsible for converting the input features Figure Ⅰ The identity is mapped to the output; the gradient direction branch is mainly responsible for extracting the input features Figure Ⅰ The gradient direction information in the output is mapped to the output features Figure II .

[0031] Furthermore, the gradient direction branch is mainly implemented by the Sobel operator, by extracting features Figure Ⅰ The horizontal gradient information G x and the vertical gradient information G y After that, the gradient direction G is calculated by the inverse tangent function dir , the calculation formula is as shown in formula (6):

[0032]

[0033] The beneficial effects of the present invention are as follows:

[0034] The present invention combines the convolution operator with the weighted robust local binary pattern, proposes a weighted differential convolution, and uses the local neighborhood relationship to mine the detailed texture features in the finger vein image that are robust to environmental changes, increase the feature stability, and improve the generalization performance of the algorithm. Secondly, the present invention analyzes the gradient direction map of low-quality real and fake human finger vein images, and proposes a gradient direction residual unit structure based on the characteristics of the difference in the detailed texture of the material of the real and fake samples. The gradient direction features are obtained through the horizontal gradient and the vertical gradient, and the network gradient direction feature expression is enhanced, and the essential material difference information of the real and fake samples is enhanced in the features. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The overall network structure diagram of the present invention;

[0036] Figure 2 Schematic diagram of the weighted differential convolution structure of the present invention;

[0037] Figure 3 Schematic diagram of the structure of the gradient direction residual unit of the present invention;

[0038] Figure 4 The network output feature map of the present invention (comparison between real and fake finger veins);

[0039] Figure 5 The network output feature map of the present invention (comparison between real and fake finger veins); DETAILED DESCRIPTION

[0040] The present invention is further described with reference to the following drawings and examples.

[0041] like Figure 1 As shown, a real human finger vein detection method specifically includes the following steps:

[0042] Step 1: Create a data set. This data set is divided into two parts: a training set and a test set. In order to increase the diversity of the data set, normal clear images, blurred images, and dirty and damaged images are added to the training set and the test set to simulate common image types in life.

[0043] 1-1: The fake human finger vein image dataset uses laboratory personnel’s fingers to realistically simulate various experimental types.

[0044] 1-2: The fake human finger vein image dataset uses prostheses made of different materials to artificially outline the vein images for each experimental type.

[0045] Step 2: Construct a real human finger vein detection model based on weighted differential convolution and gradient direction residual unit.

[0046] 2-1: Finger vein real human body detection model structure.

[0047] The schematic diagram of the finger vein real human body detection model proposed in the present invention is as follows: Figure 1 As shown in the figure, first, the input finger vein image is normalized to 1×224×224. After passing through the 7×7 input convolution layer, BN layer, and ReLu layer, the network outputs a feature map of size 64×112×112. Then, it passes through the gradient direction residual unit in sequence to enhance the finger vein gradient direction features, making the subsequent classification of the network more accurate. Finally, the classification is completed through the fully connected layer and the classification result is output. After the standard cross entropy loss is calculated, the back propagation is performed. The convolution layers are all weighted differential convolutions, which are used to enhance the network's robust expression of detailed texture features of finger vein images, enhance the environmental robustness of features, and increase the generalization performance of the network. In addition, the gradient direction residual unit structure is used to enhance the extraction of the gradient direction features of finger vein images by the network residual unit, and increase the material difference between real and fake human finger veins.

[0048] 2-2: Weighted difference convolution.

[0049] In finger vein recognition, low-quality images are usually affected by conditions such as lighting direction, intensity, and noise. If the extracted detail features are not robust, the generalization performance of the algorithm will be reduced. Therefore, for the detection task of low-quality real human finger vein images, it is crucial to learn the robust detail features of the image.

[0050] In the field of image processing, local binary patterns express the texture features of a certain location through the local neighborhood relationship of image pixels, and are relatively robust to environmental changes such as lighting. Therefore, in order to enhance the network's expression of spatial texture details of finger vein images, this paper proposes weighted differential convolution, which further enhances the robustness of texture features in the network feature map by improving the convolution operator. Weighted differential convolution can obtain more pixel-based differential information during the convolution process. At the same time, through the weighted combination of the center pixel value and the neighborhood pixel value, it can avoid abnormal results caused by fluctuations in the center single pixel value, so that the differential convolution can have the ability to resist noise interference, so as to adapt to the situation where noise is common in low-quality finger vein images, so that the acquired features are more robust. The schematic diagram of weighted differential convolution is as follows Figure 2 shown.

[0051] The specific process is as follows: first, the original image (normalized input finger vein image) is weighted to obtain the pixel value to be differentiated, as shown in formula (1). Finally, the original image and the pixel value to be differentiated p are calculated. wlg The difference result performs the normal convolution output feature Figure Ⅰ .

[0052]

[0053] In formula (1), pwlg represents the pixel value to be differentiated, that is, the result after weighted operation on the original image. p0 represents the center pixel, p n Represents the neighborhood pixel value, α is the weight coefficient of the center value, which is used to adjust the importance of the center pixel value. In this experiment, it is taken as 1.5.

[0054] The specific usage formula (2) of the weighted differential convolution of the present invention is as follows:

[0055] y=θ·y wdc +(1-θ)·y contional (2)

[0056] In formula (2), the y output feature map represents the features of the weighted differential convolution output Figure Ⅰ ;y wdc Indicates the use of Figure 2 The output result of the convolution is shown in formula (3). contional represents the output result of traditional convolution, as shown in formula (4). θ is a hyperparameter used to adjust the proportion of traditional convolution and weighted differential convolution. In this experiment, it is set to 0.6.

[0057]

[0058]

[0059] In formula (3), R represents all pixels in the current neighborhood, w(p i ) means that at pixel p i The weight coefficient at x(p i ) represents the input image at pixel p i The value at x(p wlg ) represents the expanded matrix of the pixel values ​​to be differentiated, which is used to perform the differential operation with the original input.

[0060] Substituting formula (3) and formula (4) into formula (2) yields:

[0061]

[0062] 2-3: Gradient-directed residual unit.

[0063] Since the real human finger vein image contains human skin texture details, it has many edges and the same texture has smooth connections, while the fake human finger vein is limited by the manufacturing material and cannot completely simulate the human skin texture details. Therefore, this information can be used to distinguish between real and fake human finger veins. In order to enhance the feature difference between real and fake human finger veins, the present invention analyzes the gradient direction features of real and fake human finger veins and proposes a gradient direction residual unit, so that the feature map can better maintain the gradient direction information when propagating in the network.

[0064] The difference between the real and fake human finger vein gradient directions is that the gradient direction image of the real finger vein has a more consistent gradient direction change, and contains more material details of the human skin texture, while the gradient direction image of the fake human finger vein has a more random gradient direction change. There is an obvious difference in material texture between the gradient direction images of the real and fake human finger veins, which is crucial for distinguishing the real from the fake in real human finger vein detection. In order to enhance the feature difference between the real and fake human finger veins, the present invention proposes a gradient direction residual unit, the structure of which is as follows: Figure 3 shown.

[0065] The gradient direction residual unit is mainly divided into three branches: convolution branch, short-circuit connection branch and gradient direction branch. The convolution branch is mainly responsible for the input features. Figure Ⅰ Perform convolution operations to obtain effective information and enhance the network learning ability; the short-circuit connection branch is mainly responsible for converting the input features Figure Ⅰ The identity is mapped to the output to avoid overfitting and network degradation problems caused by the deepening of the network layers; the gradient direction branch is mainly responsible for extracting the input features Figure Ⅰ The gradient direction information in the output is mapped to the output features Figure II , to avoid the loss of gradient direction information in the feature map as the network deepens.

[0066] The gradient direction branch is mainly realized by the Sobel operator. In the field of image processing, the Sobel operator is mainly used to calculate the grayscale approximation of the image gradient. Applying this operator to a certain pixel in the finger vein image will obtain the gradient information of the corresponding direction.

[0067] By extracting features Figure Ⅰ The horizontal gradient information G x and the vertical gradient information G y After that, the gradient direction G is calculated by the inverse tangent function dir , the calculation formula is as shown in formula (6):

[0068]

[0069] Example:

[0070] The deep learning framework used in the experiment of the present invention is Pytorch. The single batch size in the training phase is set to 16, the global learning rate is set to 0.0001, and the server GPU used for training is NVIDIA-TITAN-RTX.

[0071] The test set uses a self-collected and produced real human finger vein detection dataset for research. The images in this database are collected from fake fingers of different materials and artificially outlined vein patterns. The database can be roughly divided into the following three categories based on the common confusion of true and false samples in actual applications of finger vein detection:

[0072] (1) Normal and clear finger: The real sample is a real normal and clear finger, and the false sample is an artificially made finger with normal veins;

[0073] (2) Low-quality fuzzy fingers: The real sample is a real fuzzy finger, and the fake sample is an artificially created fuzzy vein finger;

[0074] (3) Low-quality dirty and damaged fingers: The real samples are real dirty and damaged fingers, and the fake samples are artificially made dirty and damaged venous fingers.

[0075] The dataset is divided into training set and test set, with a total of 25,000 images in the training set and 5,000 images in the test set. The real and fake human finger vein image datasets each account for half of the training set and test set. In the training set, there are three experimental types, normal and clear, blurred, and dirty and damaged, each accounting for more than 8,000 images, more than 4,000 of which are real samples and the other half are fake samples. The test set also includes three experimental types, each accounting for more than 1,600 images, half of which are real samples and the other half are fake samples. While ensuring the diversity of the training set and test set, avoid the imbalance and long tail of samples.

[0076] In order to prove that network extraction can effectively increase the environmental robustness and material difference of real and fake sample features, this experiment outputs the feature map of the middle layer of the network. For the convenience of observation, this experiment selects the output of the second-level residual unit of the network of the present invention and the F-net network for comparison, as shown in Figure 2. Figure 3 and Figure 4 As shown, the environmental robustness and material diversity of the features proposed by the real human finger vein detection network of the weighted differential convolution and gradient residual unit of the present invention are respectively demonstrated.

[0077] (1) Environmental robustness

[0078] In order to prove that the network feature extraction of the present invention is environmentally robust, the experiment selects false sample images collected in different environments for feature extraction, and the vein structures of the false sample images are the same. Figure 4As shown. Wherein a refers to the vein false sample image 1 (for the convenience of description, hereinafter referred to as prosthesis 1), b is the characteristic image (prosthesis 1) obtained by using the network of the present invention, c is the characteristic image (prosthesis 1) obtained by using the algorithm F-net network, d refers to the vein false sample image 2 (for the convenience of description, hereinafter referred to as prosthesis 2), e is the characteristic image (prosthesis 2) obtained by using the network of the present invention, and f is the characteristic image (prosthesis 2) obtained by using the algorithm F-net network.

[0079] It can be seen that: when the original image environment changes, the features extracted by the F-net network do not express the robust features of the original image, and the final features are significantly affected by environmental changes (such as lighting direction and intensity), which will reduce the generalization performance of the algorithm; and when the original image environment changes, the features extracted by the network of the present invention change very slightly, proving that the real human finger vein detection network based on weighted differential convolution and gradient residual unit of the present invention mines texture and gradient direction features through the local neighborhood relationship of pixel points, and the output features have higher environmental robustness.

[0080] (2) Material Difference

[0081] In order to prove that the network extraction features of the present invention can reflect the material differences between real and fake human finger veins, the experiment selected real sample images and fake sample images for feature extraction. Figure 5 As shown. Among them, a is a real human finger vein image (live), b is a real sample feature image (live) obtained by using the algorithm F-net network, c is a real sample feature image (live) obtained by using the network of the present invention, d is a fake human finger vein image (prosthesis), e is a fake sample feature image (prosthesis) obtained by using the algorithm F-net network, and f is a fake sample feature image (prosthesis) obtained by using the network of the present invention.

[0082] Depend on Figure 5 It can be seen that: for non-vein areas, the features obtained by the algorithm F-net network are fuzzy, while the network of the present invention can extract finer-grained material textures, increase the more intrinsic difference information between true and false samples, and can be used in the task of distinguishing true and false samples, thereby improving the accuracy of distinction; similarly, for vein areas, the algorithm F-net network has fuzzy finger vein features, resulting in serious overlap of vein areas in the feature map, which will lead to a decrease in the accuracy of the network in distinguishing true and false samples, while the finger vein texture obtained by the network of the present invention is clearer and more obvious, and there will be no problem of confusion between true and false samples, that is, the real and false human finger vein features proposed by the real human finger vein detection network based on weighted differential convolution and gradient residual unit of the present invention are more different.

[0083] Euclidean distance is the most common indicator for measuring image differences. The smaller the Euclidean distance, the smaller the difference between images. In order to verify that the feature map extracted by the network of the present invention can effectively widen the difference between true and false samples, this experiment uses the Euclidean distance between the feature map of the real sample and the feature map of the false sample to show the difference. The larger the Euclidean distance, the greater the difference and the easier it is to distinguish. The smaller the Euclidean distance, the more difficult it is to distinguish.

[0084] In order to further verify the effectiveness of the real human finger vein detection algorithm of the present invention, this experiment selected the final output feature graphs of the F-Net algorithm and the algorithm in this paper for comparison. The three indicators of minimum Euclidean distance, maximum Euclidean distance and average Euclidean distance were statistically analyzed, and the results are shown in Table 1.

[0085] Network structure Minimum Euclidean distance Maximum Euclidean distance Average Euclidean distance F-Net 8.0493 42.3964 25.8560 This article network 39.6690 58.0494 47.6516

[0086] Table 1 Comparison of Euclidean distance of feature graphs The network proposed by the present invention is trained using the training set to obtain the optimal parameter model and then a performance verification test is conducted. In order to demonstrate the performance of real human finger vein detection, this experiment first selects the AlexNet network using ordinary convolution and the F-Net algorithm using ordinary convolution to test in various types of data sets. In addition, in order to highlight the effectiveness of the weighted differential convolution and gradient direction unit proposed by the present invention, the F-Net using weighted differential convolution, the algorithm of the present invention using weighted differential convolution (and not using gradient direction unit), and the algorithm of the present invention using both weighted differential convolution and gradient direction unit are selected. The experimental results are shown in Table 2.

[0087]

[0088] Table 2 Comparison of different network recognition rate algorithms

[0089] In order to more intuitively compare and verify the real human finger vein detection performance of the present invention, this experiment uses the average classification error rate ACE to evaluate the classification results of the real human finger vein detection algorithm. The lower the generalization ability of the network, the higher its ACE value. The evaluation index is defined as:

[0090]

[0091] In the formula, SFPR represents the ratio of real samples to false samples, and SFNR represents the ratio of false samples to real samples. The comparison results of the network of the present invention and other networks are shown in Table 3.

[0092]

[0093]

[0094] Table 3 ACE comparison results of algorithms on corresponding datasets

[0095] It can be seen from Table 3 that in the normal image data set, the blurred image data set, and the peeling image data set, the weighted differential convolution and the gradient direction unit proposed in the present invention can bring performance improvement, and when the network based on weighted differential convolution and gradient direction unit is adopted, the correct detection rate of real samples is significantly improved compared with the AlexNet network and the F-Net network in each type of data set, which proves the effectiveness of the algorithm based on the weighted differential convolution and gradient direction unit network of the present invention.

[0096] In summary, the present invention proposes a real human finger vein detection network that combines weighted differential convolution and gradient directional residual unit. It aims at the problems that low-quality finger vein images are greatly affected by environmental changes and the difference between true and false samples is insufficient. The present invention first designs a weighted differential convolution to use the local neighborhood relationship to mine the detailed texture features in the finger vein image that are robust to environmental changes, thereby increasing the stability of the features; secondly, a gradient directional residual unit structure is designed to enhance the essential material difference information of the true and false samples in the features. Through experimental analysis and comparison, the present invention can effectively increase the environmental robustness and material difference of the proposed features, and can effectively improve the real human finger vein detection performance for low-quality images.

[0097] The examples of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above examples. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention, which should also be regarded as the protection scope of the present invention.

Claims

1. A real human finger vein detection method, characterized in that The steps include: Step 1: Create a data set. This data set is divided into two parts: a training set and a test set. In order to increase the diversity of the data set, normal clear images, blurred images, and dirty and damaged images are added to the training set and the test set to simulate common image types in life; 1-1: The fake human finger vein image dataset uses laboratory personnel’s fingers to simulate various experimental types; 1-2: The fake human finger vein image dataset uses prostheses of different materials to manually outline the vein image for each experimental type; Step 2: Construct a real human finger vein detection model based on weighted differential convolution and gradient direction residual unit; The specific implementation of the vein real human body detection model is as follows: First, the input finger vein image is normalized to 1×224×224. After passing through the 7×7 input convolution layer, the BN layer, and the ReLu layer, the network outputs a feature map with a size of 64×112×112. Then, it passes through the gradient direction residual unit in sequence to enhance the finger vein gradient direction features, making the subsequent classification of the network more accurate. Finally, the classification is completed through the fully connected layer and the classification result is output. After the standard cross entropy loss is calculated, the network is back-propagated. The weighted difference convolution is implemented as follows: First, the pixel value to be differentiated is obtained by performing weighted calculation on the original image, as shown in formula (1); finally, the original image and the pixel value to be differentiated p are calculated. wlg The difference result performs the feature map I of the output of ordinary convolution; The original image is the normalized input finger vein image; in formula (1), p wlg represents the pixel value to be differentiated, that is, the result after weighted operation on the original image; p0 represents the center pixel point, p n represents the neighborhood pixel value, and α is the weight coefficient of the center value; The specific formula (2) used for weighted differential convolution is as follows: y=θ·y wdc +(1-θ)·y contional (2) In formula (2), y output feature map represents the feature map I output by weighted differential convolution; y wdc represents the output result of weighted differential convolution, as shown in formula (3); contional represents the output result of traditional convolution, as shown in formula (4); θ is a hyperparameter used to adjust the proportion of traditional convolution and weighted differential convolution; In formula (3), R represents all pixels in the current neighborhood, w(p i ) indicates that at pixel p i The weight coefficient at x(p i ) represents the input image at pixel p i The value at x(p wlg ) represents the extended matrix of the pixel values ​​to be differentiated, which is used to perform a differential operation with the original image; Substituting formula (3) and formula (4) into formula (2) yields: Among them, y output feature map represents the feature map I output by weighted differential convolution; The specific implementation of the gradient direction residual unit is as follows: The gradient direction residual unit consists of three branches: a convolution branch, a short-circuit connection branch and a gradient direction branch; the convolution branch is mainly responsible for performing convolution operations on the input feature map I to obtain effective information from it; the short-circuit connection branch is mainly responsible for mapping the input feature map I to the output; the gradient direction branch is mainly responsible for extracting the gradient direction information in the input feature map I and mapping it to the output feature map II.

2. A real human finger vein detection method according to claim 1, characterized in that The convolution layers are all weighted differential convolutions, and the gradient direction residual unit structure is used to enhance the extraction of the gradient direction features of the finger vein image by the network residual unit, thereby increasing the material difference between real and fake human finger veins.

3. A real human finger vein detection method according to claim 1, characterized in that The gradient direction branch is mainly realized by the Sobel operator, which extracts the horizontal gradient information G in the feature map I. x and the vertical gradient information G y After that, the gradient direction G is calculated by the inverse tangent function dir , the calculation formula is as shown in formula (6):

Images