A finger vein recognition method and system based on VC-Gabor convolution and multi-branch structure

Abstract

The present invention belongs to the field of automatic control, and particularly relates to a finger vein recognition method based on variable curvature Gabor convolution and a multi-branch structure, and a corresponding finger vein recognition system and finger vein verification device. The solution of the present invention first improves the traditional Gabor filter by adding variable curvature so as to extract information on different bending degrees in the vein curve. Secondly, the variable curvature Gabor function is constructed into a learnable VC-Gabor convolution layer, and the parameters of the VC-Gabor convolution layer are updated by using the backpropagation of the neural network, so that the directions, scales and curvatures of the learned VC-Gabor filters are more abundant, and complex manual parameter adjustment is avoided. Finally, a network model including a VC-Gabor convolution layer and an adaptive multi-branch structure is designed in combination with the convolution layer in the classical CNN network, and then the network model is used for rapid recognition of finger vein images. The present invention solves the defects of insufficient accuracy and poor robustness of the existing finger vein recognition methods.

Description

Technical Field

[0001] The present invention belongs to the field of automatic control, and particularly relates to a finger vein recognition method based on VC-Gabor convolution and a multi-branch structure, as well as a corresponding finger vein recognition system and a finger vein verification device. Background Art

[0002] Identity authentication is to confirm the identity of a user through certain technical means. With the continuous improvement of the requirements for privacy and security in the big data era, traditional identity authentication methods, such as marker authentication using keys, certificates, and passwords, have gradually been replaced by biometric recognition technologies. Commonly used biometric recognition technologies include fingerprint, face, iris, and voice, etc., and these technologies have been widely applied in many electronic devices and Internet of Things devices. In addition, finger vein features are also an emerging biometric recognition. Finger vein features are difficult to be forged due to the need for live detection and their distribution under the skin and other unique advantages, greatly reducing the possibility of being stolen, and having broad application potential.

[0003] The primary problem to be solved in finger vein biometric recognition technology is how to extract finger vein features. Existing methods include methods based on vein patterns, methods based on local binary coding, methods based on feature dimensionality reduction, and methods based on deep learning, etc.

[0004] The method based on vein patterns uses the shape and position information of finger vein blood vessels as the basis for recognition to perform vein recognition. This method can improve image quality and recognition performance. However, in a blurred area affected by irregular shadows and noises, it is difficult for the vein pattern method to effectively extract vein patterns.

[0005] The method based on local binary coding first performs binary coding on the local regions of the image in blocks, and then combines these binary codings to obtain a feature vector of the vein pattern. In the process of feature extraction, the method based on local binary coding needs to define local neighborhoods and radius sizes. Different neighborhood and radius selections may lead to different texture features being extracted, which requires adjustment and optimization according to specific problems. This may require certain experience and professional knowledge, increasing the subjectivity and complexity of the method.

[0006] The method based on dimensionality reduction refers to a method of processing high-dimensional data by reducing its dimensionality and mapping the data into a low-dimensional space. This method can effectively reduce the time and space complexity of data processing and improve the performance of the classifier at the same time. Typical methods include principal component analysis (PCA), 2DPCA, and linear discriminant analysis (LDA). Due to the relatively low feature dimensionality, this method has the advantage of fast recognition speed, but the global features are greatly affected by factors such as posture, occlusion, deformation, and illumination, and have great limitations when used in finger vein image feature extraction.

[0007] In recent years, convolutional neural networks have developed rapidly in the field of image classification. Some researchers have applied convolutional neural networks to finger vein recognition and achieved good results. This method has a high response speed in the recognition stage. However, the manually crafted features can be considered as surface features. This solution is sensitive to noise, difficult to modulate parameters, and not very robust to the transformation of the dataset. Therefore, it is prone to overfitting and the recognition accuracy in the actual application process is still not good. Summary of the Invention

[0008] To solve the deficiencies such as insufficient accuracy and poor robustness existing in various existing finger vein recognition methods, the present invention provides a finger vein recognition method based on VC-Gabor convolution and a multi-branch structure, as well as a corresponding finger vein recognition system and a finger vein verification device.

[0009] The present invention is implemented by the following technical solutions:

[0010] A finger vein recognition method based on VC-Gabor convolution and a multi-branch structure, which includes the following steps:

[0011] S1: Design a Gabor filter with variable curvature. The definition of the Gabor filter with variable curvature is as follows:

[0012]

[0013] In the above formula, x and y represent the pixel coordinate positions; θ represents the angle between the direction of the sine function and the x-axis, that is, the direction of the kernel function of the filter; λ represents the wavelength of the sine function; f represents the curvature of the filter; σ is the standard deviation of the Gaussian function; i represents the imaginary unit.

[0014] S2: Construct a parameter-adjustable VC-Gabor convolution according to the Gabor filter with variable curvature. The construction process of the VC-Gabor convolution is as follows:

[0015] S21: Decompose the Gabor filter with variable curvature into a real part G re and an imaginary part G im .

[0016] S22: Use the decomposed real part to construct the VC-Gabor convolution, and define θ, λ, f, and σ as learnable optimization parameters.

[0017] S23: Calculate the real part G re of the Gabor filter with variable curvature, and respectively take partial derivatives with respect to each parameter θ,

[0018] λ, f, and σ to be optimized.

[0019] S24: Update g using the gradient descent rule of the backpropagation algorithm re for its parameters and establish the update formulas for each optimization parameter.

[0020] S3: Improve the classical CNN network to construct the required finger vein recognition model. The input of the finger vein recognition model is a finger vein image, and the output is the category of the finger vein image. The model construction process is as follows:

[0021] S31: Obtain the classical CNN network, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a Softmax classification layer; the convolutional layer includes a downsampling convolutional layer and a non-downsampling convolutional layer.

[0022] S32: Use the newly designed VC-Gabor convolution as the first layer of the CNN network.

[0023] S33: Use the double-branch convolution module in all downsampling convolutional layers in the convolutional layer.

[0024] S34: Replace all non-downsampling convolutional layers in the convolutional layer with a three-branch convolution module.

[0025] S4: Obtain the finger vein images of a large number of users as sample data to form a data set, divide the data set into a training set, a validation set, and a test set, and add labels to the sample data.

[0026] S5: Train the finger vein recognition model using the training set and validate the model using the validation set.

[0027] S6: Use the trained finger vein recognition model to identify finger vein images on the test set and output the category information of the current user.

[0028] As a further improvement of the present invention, in step 21, the real part g re split out can effectively extract the edge and texture information of the finger vein image. The function is as follows:

[0029]

[0030] The imaginary part g im split out, the function is as follows:

[0031]

[0032] As a further improvement of the present invention, in step S23, the partial derivative functions of each optimization parameter are as follows:

[0033]

[0034] where x' = xcosθ + ysinθ; y' = -xsinθ + ycosθ.

[0035] As a further improvement of the present invention, in step S24, the update formula for each optimization parameter is:

[0036]

[0037] In the above formula, θ t , λ t , f t and σ t are the optimization parameters before update; θ t+1 , λ t+1 , f t+1 and σ t+1 are the optimization parameters after update; α is a preset learning rate; L is a preset objective function;.

[0038] As a further improvement of the present invention, in step S3, the input of the constructed finger vein recognition model is a grayscale image with a channel number c = 1. The channel number of VC-Gabor convolution is 40, and the parameters θ, λ, f, and σ to be optimized are updated during the learning process.

[0039] In the convolutional layer, the size of the feature map obtained by each downsampling convolutional layer is halved and the channel number is doubled; the size and channel number of the feature map obtained by each non-downsampling convolutional layer remain unchanged. The pooling layer uses a global average pooling layer. The number of neurons in the fully connected layer is the total number of categories in the dataset. The Softmax classification layer is used to convert the original output into a probability distribution for multi-class classification tasks.

[0040] As a further improvement of the present invention, in the dual-branch convolution module of step S33, the input feature map enters two branches respectively, and 3×3 convolution and batch normalization are completed in the first branch, and 1×1 convolution and batch normalization are completed in the second branch; then the adaptive weights of each branch are superimposed and fused, and the fused feature is output after passing through the ReLU activation layer.

[0041] The expression of the dual-branch convolution module is as follows:

[0042] Y = ReLU(w1·BN(Conv1(X)) + w2·BN(Conv2(X)))

[0043] In the above formula, X represents the input feature map; Y represents the output feature map; Conv1 represents 3×3 convolution; Conv2 represents 1×1 convolution; BN represents batch normalization; ReLU is the activation function; w1 and w2 are the adaptive weights of the first branch and the second branch respectively.

[0044] As a further improvement of the present invention, in the three-branch convolution module of step S34, the input feature map enters three branches respectively. 3×3 convolution and batch normalization are completed in the first branch, 1×1 convolution and batch normalization are completed in the second branch, and only batch normalization is performed in the third branch. Then, after superimposing the adaptive weights of each branch, they are fused, and the fused features are output after being processed by the ReLU activation layer.

[0045] The expression of the three-branch convolution module is as follows:

[0046] Y = ReLU(w1·BN(Conv1(X)) + w2·BN(Conv2(X)) + w3·BN(X))

[0047] In the above formula, X represents the input feature map; Y represents the output feature map; Conv1 represents 3×3 convolution; Conv2 represents 1×1 convolution; BN represents batch normalization; ReLU is the activation function; w1, w2, and w3 are the adaptive weights of the first branch, the second branch, and the third branch respectively.

[0048] As a further improvement of the present invention, in step S4, the obtained data set is randomly augmented with data augmentation means including rotation, translation, flipping, and simulated illumination changes, and the augmented samples have the same labels as the original samples.

[0049] In step S5, the batch size of the finger vein recognition model in the training stage is set to 32, and it is trained using the Adam optimizer with an initial learning rate of 0.001, and the learning rate is adjusted according to the cosine annealing learning rate adjustment strategy with a period of 20. Among them, the objective function is set as follows:

[0050]

[0051] In the above formula, C is the number of samples, y i represents the one-hot representation of the sample label: when the sample belongs to the i-th class, y i = 1, otherwise y i = 0; p i represents the probability that the sample belongs to the i-th class.

[0052] The present invention also includes a finger vein recognition system, which is used to identify the category of the input finger vein image by using the finger vein recognition method based on VC-Gabor convolution and multi-branch structure as described above.

[0053] The finger vein recognition system includes: an image acquisition device, an image preprocessing module, a template library, and a finger vein recognition model.

[0054] The image acquisition device is used to acquire the original finger vein image of the current user to be recognized. The image preprocessing module is used to preprocess the original finger vein image of the user, and then convert it into a grayscale image that meets the resolution requirements. The template library stores the finger vein feature vectors and corresponding labels of all registered users.

[0055] The input of the finger vein recognition model is the preprocessed grayscale image, and the output is the matching relationship between the current user and the registered users. The finger vein recognition model is a network model designed and trained by using the finger vein recognition method based on VC-Gabor convolution and multi-branch structure as described above; the network model includes a multi-channel filter, a feature extraction module, a fully connected layer, and a classification layer. The multi-channel filter is automatically generated by the VC-Gabor convolution module, and is used to extract the direction, scale, and curvature features of the finger vein image texture. The feature extraction module uses a convolutional layer including a double-branch convolution module and a triple-branch convolution module; the feature extraction module is used to extract multi-scale features of the surface layer and deep layer of the finger vein image; the number of neurons in the fully connected layer is the total number of categories of the data set. The fully connected layer connects the input nodes of the global average pooling layer to each output node of the current layer, so as to perform feature extraction and transformation. The classification layer uses the Softmax function to convert the original output into a probability distribution, so as to match the identity of the current user with the labels of all registered users in the template library.

[0056] The present invention further includes a finger vein verification device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it creates parts of the image preprocessing module, the template library, and the finger vein recognition model in the finger vein recognition system as described above. Then, by using the steps in the finger vein recognition method based on VC-Gabor convolution and multi-branch structure as described above, it realizes the recognition of the finger vein image of the current user collected, and outputs the matching relationship between the current user and the registered users.

[0057] The technical solution provided by the present invention has the following beneficial effects:

[0058] The present invention combines curvature information and direction scale information to construct a Gabor filter with variable curvature, and integrates the variable curvature Gabor filter into the convolutional neural network. By calculating the loss function and using the backpropagation method to optimize the parameters of the VC-Gabor convolutional layer, there is no need to manually modulate the parameters of the filter, making the scale, direction, and curvature information of the VC-Gaobr filter variable and the feature extraction ability stronger.

[0059] Based on the classic CNN network, the present invention introduces a VC-Gabor convolutional layer constructed by variable-curvature Gaobr filters and multi-branch structure convolutions into the network model, thereby enhancing the extraction of finger vein features by the network model, improving the recognition accuracy of the network model, and overcoming the defect of poor generalization of traditional convolutional neural networks in dealing with finger vein image recognition problems. Description of the Drawings

[0060] Figure 1 It is a flowchart of the steps of a finger vein recognition method based on VC-Gabor convolution and multi-branch structure provided in Embodiment 1 of the present invention.

[0061] Figure 2 It is a three-dimensional space image of the real part and the imaginary part of the variable-curvature Gabor function designed in Embodiment 1 of the present invention.

[0062] Figure 3 It is a schematic diagram of the parameters of a common convolution kernel and the VC-Gabor convolution kernel designed in this embodiment.

[0063] Figure 4 It is an architecture diagram of the finger vein recognition model designed in Embodiment 1 of the present invention.

[0064] Figure 5 It is a schematic diagram of the structure of the double-branch convolution module adopted in Embodiment 1 of the present invention.

[0065] Figure 6 It is a schematic diagram of the structure of the triple-branch convolution module adopted in Embodiment 1 of the present invention.

[0066] Figure 7 It is a schematic diagram of the framework of the finger vein recognition system provided in Embodiment 2 of the present invention.

[0067] Figure 8 It is a visualization image of the VC-Gabor filter in the performance test experiment.

[0068] Figure 9 It is a two-dimensional visualization image of the traditional Gabor filter with manually adjusted parameters in the performance test experiment.

[0069] Figure 10 It is a visualization image of the traditional Gabor filter with parameters adjusted by the classic CNN network in the performance test experiment.

[0070] Figure 11 It is a case image in the FV-USM dataset.

[0071] Figure 12 It is a case image in the SDUMLA dataset.

[0072] Figure 13Comparison of the feature maps learned by the ordinary convolution kernel and the VC-Gabor convolution kernel in the performance test stage.

[0073] Figure 14 Curves of experimental groups 2, 4, and 5 on the FV-USM dataset in the ablation experiment.

[0074] Figure 15 Curves of experimental groups 2, 4, and 5 on the SDUMLA dataset in the ablation experiment.

[0075] Figure 16 Curves of experimental groups 1, 2, 3, and 5 on the FV-USM dataset in the ablation experiment.

[0076] Figure 17 Curves of experimental groups 1, 2, 3, and 5 on the SDUMLA dataset in the ablation experiment. Detailed implementation manners

[0077] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0078] Embodiment 1

[0079] A finger vein recognition method based on VC-Gabor convolution and multi-branch structure provided in this embodiment. First, the traditional Gabor filter is improved by adding variable curvature to extract different bending degree information in the vein curve. Secondly, the variable curvature Gabor function is constructed into a learnable convolution layer, and the parameters of the VC-Gabor convolution layer are updated by using the backpropagation of the neural network, so that the learned VC-Gabor filter directions, scales and curvatures are more abundant, and complex manual parameter tuning is avoided. Finally, an adaptive multi-branch structure is designed in combination with the convolution layer in the classical CNN network for feature extraction of finger vein images, thereby enhancing the feature extraction ability of the model and effectively avoiding overfitting.

[0080] Specifically, as Figure 1 shown, the technical solution provided in this embodiment includes the following steps:

[0081] S1: Design a Gabor filter with variable curvature.

[0082] The Gabor filter is an effective tool for extracting spatially local texture features. In essence, it is a sinusoidal plane wave modulated by a Gaussian envelope, which can perform local and directional frequency analysis on two-dimensional information, thereby realizing the extraction of texture information. A specially tuned Gabor filter can extract texture features in a specific direction and at a specific scale. Combining the method of multi-channel filtering, Gabor filters with different parameters can extract local texture features on the finger vein image from different directions and resolutions. Moreover, the Gabor filter is insensitive to illumination changes and can well adapt to the complex illumination changes during the acquisition of finger vein images. Therefore, the Gabor filter can be widely applied to scenarios with rich texture information such as finger vein feature extraction.

[0083] By analyzing the finger vein image data, it can be observed that the image contains rich vein pattern information. These vein pattern information is very important for distinguishing different categories of images. Therefore, when extracting the features of the vein image, the information in the vein area should be focused on, while ignoring the background information. The traditional Gabor filter cannot effectively extract the vein linear features with a large number of different curvatures in the finger vein image. In this embodiment, a variable curvature Gabor filter is designed based on the traditional Gabor filter. This filter can construct the finger vein image features at different scales, considering both the direction of the finger vein features and the degree of curvature of the vein curve.

[0084] The definition of the variable curvature Gabor filter designed in this embodiment is as follows:

[0085]

[0086] In the above formula, x and y represent the pixel coordinate positions; θ represents the angle between the direction of the sine function and the x-axis, that is, the direction of the kernel function of the filter; λ represents the wavelength of the sine function; f represents the curvature of the filter; σ is the standard deviation of the Gaussian function; i represents the imaginary unit.

[0087] Among them, adjusting the λ parameter of the filter can change the wavelength of the plane wave of the variable curvature Gabor filter, thereby realizing multi-scale detection; adjusting the θ parameter can change the direction of the sine wave of the filter, thereby realizing multi-direction detection; adjusting the f parameter can change the curvature of the filter and realize the feature extraction of different bending degrees of the vein curve. The setting of each parameter in the variable curvature Gabor filter will have an important impact on its feature extraction ability. In order to obtain a more comprehensive feature representation of the image, usually multiple VC-Gabor filters with different parameters are combined into a filter bank to extract texture features of different scales, different directions and different bending degrees. In this embodiment, a 40-channel VC-Gabor filter with different parameters will be used in the subsequent process to extract finger vein features.

[0088] S2: Construct an adjustable parameter VC-Gabor convolution according to the Gabor filter with variable curvature. The construction process of the VC-Gabor convolution is as follows:

[0089] S21: Decompose the Gabor filter with variable curvature into the real part G re and the imaginary part G im . The Gabor filter with variable curvature is a complex function. After decomposing it, the real part G re extracted from it can effectively extract the edge and texture information of the finger vein image. The function is as follows:

[0090]

[0091] The extracted imaginary part G im The function is as follows:

[0092]

[0093] Among them, the three-dimensional spatial images of the real part and the imaginary part of the Gabor function with variable curvature are as Figure 2 shown.

[0094] S22: Use the decomposed real part to construct the VC-Gabor convolution, and define θ, λ, f, and σ as learnable optimization parameters.

[0095] In the traditional solution, the parameter setting of the Gabor filter is mostly based on experience, which can neither guarantee the optimality of the parameters nor adjust the parameters when the dataset changes, resulting in poor robustness of the algorithm. The advantage of CNN is that it can obtain efficient feature extraction capabilities through a large amount of data training. In this embodiment, the scheme selects to use the VC-Gabor filter to construct the convolutional layer, and then uses the CNN network to assist in optimizing the parameters of the VC-Gabor filter.

[0096] Specifically, in the subsequent solution content, in order to make full use of the characteristics of the Gabor filter with variable curvature and the convolutional neural network, this embodiment uses the real part of the Gabor function with variable curvature to construct the VC-Gabor convolutional layer to replace the traditional convolutional layer, and uses the convolutional neural network to learn the parameters of the VC-Gabor filter. Finally, the learned VC-Gabor convolution kernel is used to extract the direction, scale, and curvature features of the finger vein image texture. Since the VC-Gabor does not directly use learnable weights in the pixel space, the parameters learned in the VC-Gaobr filter are different from the parameters of the traditional convolution kernel. Each pixel of the traditional convolution kernel is a learnable parameter. For example, for a convolution kernel with a channel size of K×K, the number of learned parameters is K 2 , asFigure 3 as shown in (a) of Figure 3 For the VC-Gabor kernel in this embodiment, as shown in part (b) of

[0097] S23: Calculate the real part G of the Gabor filter with variable curvature re ; Take the partial derivatives of the parameters θ, λ, f, and σ that need to be optimized respectively.

[0098] The partial derivative functions of each optimization parameter are as follows:

[0099]

[0100] where x' = xcosθ + ysinθ; y' = -xsinθ + ycosθ.

[0101] S24: Update the parameters of G using the gradient descent rule of the backpropagation algorithm re and establish the update formulas for each optimization parameter.

[0102] The update formulas for each optimization parameter are:

[0103]

[0104] In the above formula, θ t , λ t , f t and σ t are the optimization parameters before update; θ t+1 , λ t+1 , f t+1 and σ t+1 are the optimization parameters after update; α is the preset learning rate; L is the preset objective function.

[0105] In this embodiment, the objective function is set as follows:

[0106]

[0107] In the above formula, C is the number of samples, and y i represents the one-hot representation of the sample label: when the sample belongs to the i-th class, y i = 1, otherwise y i = 0; p i represents the probability that the sample belongs to the i-th class.

[0108] S3: Improve the classic CNN network to construct the required finger vein recognition model. The input of the finger vein recognition model is the finger vein image, and the output is the category of the finger vein image.

[0109] In this embodiment, the overall architecture of the constructed finger vein recognition model is as shown in Figure 4 the following. The construction process of this network model generally includes the following steps:

[0110] S31: Obtain a classic CNN network, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a Softmax classification layer; the convolutional layer includes a downsampling convolutional layer and a non-downsampling convolutional layer.

[0111] In the convolutional layer of the CNN network selected in this embodiment, there are three consecutive downsampling convolutional layers and ordinary convolutional layers respectively. Among them, the number of channels of the first downsampling convolutional layer is 128, and the feature map size is half of the previous layer, that is, w = w / 2, h = h / 2. Immediately following, the number of channels and the feature map size of the first convolutional layer remain the same as those of the first downsampling convolutional layer. Then in the second downsampling convolutional layer, the number of channels doubles to 256, and the feature map size is halved again. Immediately following, the number of channels and the feature map size of the second convolutional layer remain the same as those of the second downsampling convolutional layer. Finally, in the third downsampling convolutional layer, the number of channels doubles to 512, and the feature map size continues to be halved. Immediately following, the number of channels and the feature map size of the third convolutional layer remain the same as those of the third downsampling convolutional layer. The pooling layer after the convolutional layer selects global average pooling GAP. The number of neurons in the subsequent fully connected layer is the total number of categories in the dataset. Finally, the Softmax function is selected to convert the output of the neural network into probability values of each category.

[0112] S32: Use the newly designed VC-Gabor convolution as the first layer of the CNN network.

[0113] Specifically, in this embodiment, the convolution kernel in the traditional input layer of the CNN network is replaced with the newly designed VC-Gabor convolution in this time. The parameters of the convolution kernel of the VC-Gabor convolution designed in this embodiment are fixed and are initialized according to the input image. The number of channels is set to 40, and the parameters θ, λ, f, and σ to be optimized are updated during the learning process.

[0114] The input of the finger vein recognition model constructed in this embodiment is a grayscale image with the number of channels c = 1. The first layer of the network model uses a 40-channel learnable VC-Gabor convolution kernel to extract line, direction, and texture information. At the same time, it can also filter out redundant interference information. For example, given an input grayscale image X of a finger vein: X ∈ R 80×128×1 , then the output feature map after passing through the 40-channel VC-Gabor convolutional layer is Y: Y ∈ R 80×128×40It can be seen that the size of the VC-Gabor convolutional layer remains unchanged, but different types of feature information are extracted through different channels of the filter bank. The number of channels directly changes from 1 to 40. The directions, scales, and curvatures of the VC-Gabor filters in this embodiment are variable, so the constructed VC-Gabor convolutional layer can extract richer features.

[0115] The network model structure composed of the single-path structure adopted by the convolutional layer of the traditional CNN network is simple and easy to train, but its performance is not good. Especially when the number of network layers is relatively deep, the problem of gradient disappearance is likely to occur. In order to obtain better feature extraction effects in this embodiment, a multi-branch structure is used in each convolutional block after the VC-Gabor convolutional layer. At the same time, for the feature extraction methods of different convolutional layers, this embodiment also selects two different multi-branch structures, specifically the dual-branch convolutional module and the triple-branch convolutional module. This multi-branch structure has good model representation ability and significantly improves the accuracy of vein recognition.

[0116] S33: Replace all downsampling convolutional layers in the convolutional layer with dual-branch convolutional modules.

[0117] As Figure 5 shown, in the dual-branch convolutional module, the input feature map enters two branches respectively, and 3×3 convolution and batch normalization are completed in the first branch, and 1×1 convolution and batch normalization are completed in the second branch; then after superimposing the adaptive weights of each branch, they are fused, and the fused feature is output after being processed by the ReLU activation layer.

[0118] The expression of the dual-branch convolutional module is as follows:

[0119] Y = ReLU(w1·BN(Conv1(X)) + w2·BN(Conv2(X)))

[0120] In the above formula, X represents the input feature map; Y represents the output feature map; Conv1 represents 3×3 convolution; Conv2 represents 1×1 convolution; Bn represents the batch normalization layer; ReLU is a non-linear activation function; w1 and w2 are the adaptive weights of the first branch and the second branch respectively.

[0121] Taking Figure 4 the first downsampling conv-block in 80×128×40 as an example, assuming that the input feature map from the VC-Gabor convolutional layer is X: X ∈ R Figure X 1 ∈ R 40×64×128; The second branch performs a 1x1 convolution on the input feature map, with a stride of 2 and a padding of 0, and then passes through a batch normalization layer to obtain features Figure X 2∈R 40×64×128 。

[0122] Meanwhile, through network training, a multi-branch adaptive weight w = [w1, w2] is obtained. The features extracted by the two branches are multiplied by the adaptive weights respectively and then added together. Finally, the added feature data passes through a ReLU activation function layer to obtain the output feature map Y∈R of the downsampling conv-block 40×64×128 。

[0123] S34: Replace all convolutional layers in the convolutional layer with a three-branch convolutional module

[0124] As Figure 6 shown, in the three-branch convolutional module, the input feature map enters three branches respectively. In the first branch, a 3×3 convolution and batch normalization are completed. In the second branch, a 1×1 convolution and batch normalization are completed. In the third branch, only batch normalization is performed; then, after superimposing the adaptive weights of each branch, they are fused, and the fused features are output after passing through a ReLU activation layer

[0125] The expression of the three-branch convolutional module is as follows

[0126] Y = ReLU(w1·BN(Conv1(X)) + w2·BN(Conv2(X)) + w3·BN(X))

[0127] In the above formula, X represents the input feature map; Y represents the output feature map; Conv1 represents a 3×3 convolution; Conv2 represents a 1×1 convolution; Bn represents batch normalization; ReLU is a non-linear activation function; w1, w2, and w3 are the adaptive weights of the first branch, the second branch, and the third branch respectively

[0128] Take Figure 4 the first conv-block in Figure X : X∈R 40×64×128 as the input; then the first branch performs a 3x3 convolution on the input feature Figure X with a stride of 1 and a padding of 1, and then passes through a batch normalization layer to obtain the feature X1∈R 112×112×64 ; The second branch performs a 1x1 convolution on the input feature with a stride of 1 and a padding of 0, and then passes through a batch normalization layer to obtain the feature X2∈R 40 ×64×128; The third branch is the shortcut branch, which only performs batch normalization on the input features to obtain X3 ∈ R 40×64×128 , as the structure of this shortcut branch deepens with the model, it can not only obtain more robust feature representations in the deeper layers of the network, but also better handle the problem of gradient disappearance in the deeper layers of the network.

[0129] Meanwhile, through network training, the multi-branch adaptive weights w = [w1, w2, w3] are obtained. The features extracted by the three branches are respectively multiplied by the adaptive weights and added together. Finally, the added feature data passes through the ReLU non-linear activation function layer to obtain the output feature map Y ∈ R of the conv-block 40×64×128 .

[0130] S4: Obtain the finger vein images of a large number of users as sample data to form a data set, divide the data set into a training set, a validation set and a test set, and add labels to the sample data.

[0131] Among them, in order to overcome the problem of insufficient sample data volume of the original vein images, this embodiment can also choose to use the means of image data augmentation to expand the samples. For example, the original image can be rotated, translated, horizontally flipped or vertically flipped at any angle. And simulate the illumination change to process the original image, such as adjusting the brightness of the image, or overlaying a layer for adjusting the illumination situation, etc.

[0132] S5: Use the training set to train the finger vein recognition model, and use the validation set to verify the model.

[0133] S6: Use the trained finger vein recognition model to identify finger vein images on the test set, and output the category information of the current user.

[0134] The finger vein recognition method based on VC-Gabor convolution and multi-branch structure proposed in this embodiment improves the Gabor filter, adds curvature information to extract different bending degree information in the vein curve; and uses the VC-Gabor function to construct a convolution layer with automatically updated parameters, and then combines the VC-Gabor convolution layer and the multi-branch structure to improve the traditional CNN network, enhancing the feature extraction ability of the network model and effectively avoiding model overfitting.

[0135] The method provided in this embodiment can overcome the problems that the traditional Gabor filter has relatively single direction and scale, cannot effectively extract the curvature information in the vein curve, and is difficult to adjust parameters, resulting in poor finger vein recognition effect. The finger vein recognition accuracy of the network model constructed in this embodiment is also better than the current mainstream solutions.

[0136] Embodiment 2

[0137] This embodiment provides a finger vein recognition system, which uses the finger vein recognition method based on VC-Gabor convolution and multi-branch structure in Embodiment 1 to identify the category of the input finger vein image. As Figure 7 shown, the finger vein recognition system includes: an image acquisition device, an image preprocessing module, a template library, and a finger vein recognition model.

[0138] The image acquisition device is used to acquire the original finger vein image of the current user to be recognized. The image preprocessing module is used to preprocess the original finger vein image of the user, and then convert it into a grayscale image that meets the resolution requirements. The template library stores the finger vein feature vectors and corresponding labels of all registered users.

[0139] The input of the finger vein recognition model is the preprocessed grayscale image, and the output is the matching relationship between the current user and the registered users. The finger vein recognition model is a network model designed and trained using the aforementioned finger vein recognition method based on VC-Gabor convolution and multi-branch structure; the network model includes a multi-channel filter, a feature extraction module, a fully connected layer, and a classification layer. The multi-channel filter is automatically generated by the VC-Gabor convolution module, and is used to extract the direction, scale, and curvature features of the finger vein image texture. The feature extraction module uses a convolutional layer that includes a dual-branch convolution module and a triple-branch convolution module; the feature extraction module is used to extract multi-scale features of the surface and deep layers of the finger vein image; the number of neurons in the fully connected layer is the total number of categories in the dataset, and the fully connected layer connects the input nodes of the global average pooling layer to each output node of the current layer, so as to perform feature extraction and transformation. The classification layer uses the Softmax function to convert the original output into a probability distribution, so as to match the identity of the current user with the labels of all registered users in the template library.

[0140] Embodiment 3

[0141] This embodiment also provides a finger vein verification device, which is a product for actual identity recognition and verification developed using the solutions in Embodiments 1 and 2. The finger vein verification device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it creates parts of the image preprocessing module, the template library, and the finger vein recognition model in the aforementioned finger vein recognition system. Furthermore, it uses the steps in the aforementioned finger vein recognition method based on VC-Gabor convolution and multi-branch structure to implement the recognition of the finger vein image of the current user collected, and outputs the matching relationship between the current user and the registered users.

[0142] The finger vein verification device provided in this embodiment is essentially a computer device, which can be a smart terminal capable of executing programs, a tablet computer, a notebook computer, a desktop computer, a rack-mounted server, a blade server, a tower server, or a cabinet server (including an independent server or a server cluster composed of multiple servers), etc. The computer device in this embodiment at least includes, but is not limited to, a memory and a processor that can communicate with each other through a system bus.

[0143] In this embodiment, the memory (i.e., the readable storage medium) includes flash memory, a hard disk, a multimedia card, a card-type memory (such as an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Of course, the memory can also include both the internal storage unit and the external storage device of the computer device. In this embodiment, the memory is generally used to store the operating system installed on the computer device and various application software. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0144] In some embodiments, the processor can be a central processing unit (CPU), a graphics processing unit (GPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data.

[0145] Performance Test

[0146] In order to verify the effectiveness of the finger vein recognition method based on VC-Gabor convolution and multi-branch structure provided in this embodiment, the inventor formulated a corresponding performance test plan, simulated the solution provided in this embodiment, and set up multiple control groups to compare the technical effects of this case with other solutions.

[0147] I. Test of VC-Gabor filter

[0148] In the experiment, first, the VC-Gabor filter designed in this embodiment, the traditional Gabor filter with manual parameter tuning, and the traditional Gabor filter with parameter tuning based on the classical CNN network were compared. The two-dimensional visualization images of the filters are respectively as Figure 8 、 Figure 9 and Figure 10 .

[0149] Comparing Figures 8 - 10 it can be seen that the Gabor filter with manual parameter tuning has a single direction and scale, so it is difficult to extract rich vein features. The traditional Gabor filter with parameter tuning based on the classical CNN network has diverse directions and scales and can extract vein features in different directions and scales. However, this filter cannot effectively extract the line features of different bending degrees of the vein curve, and these bending degree features are crucial for distinguishing different types of finger vein images. The VC-Gabor filter provided in this case overcomes this problem and can obtain feature information with rich details, obvious texture, and curvature.

[0150] I. Training and Performance Testing of the Network Model

[0151] 2.1. Dataset:

[0152] The training and testing of the network model in this experiment were mainly carried out under two public datasets, FV-USM and SDMULA. Among them, the FV-USM database comes from Sunway University Malaysia. It includes the left and right index fingers and middle fingers of 123 subjects, and each finger has 6 images. The images were collected twice, with 2952 images collected each time, for a total of 5904 images. The cropped ROI image resolution is 100x300 pixels. Several typical sample images are as Figure 11 shown. The SDMULA dataset comes from Shandong University and contains finger vein images of 636 fingers of 106 subjects. 6 grayscale images were obtained from the index finger, middle finger, and ring finger of the left and right hands respectively, consisting of a total of 3816 images. The cropped ROI image resolution is 150x96 pixels, and several typical sample images are as Figure 12 shown.

[0153] In this experiment, the region of interest was extracted from the original finger vein image as the input image, and the dataset was divided into a training set, a validation set, and a test set in a ratio of 4:1:1. To improve the generalization ability of the model and prevent overfitting, in this embodiment, the training set was augmented by rotating, translating, flipping, and simulating illumination changes.

[0154] 2.2. Experimental Environment and Parameter Settings:

[0155] In this experiment, the test scheme was implemented using the Pytorch framework, and the experiment was conducted on a Linux server. The main hardware configurations are shown in Table 1.

[0156] Table 1: Hardware configurations adopted in the experiment

[0157]

[0158] During the simulation experiment, the batch size in the network model training stage was set to 32, and the Adam optimizer was used for training. The initial learning rate was 0.001, and the learning rate was adjusted according to the cosine annealing learning rate adjustment strategy with a period of 20.

[0159] 2.3. Evaluation metrics:

[0160] To fairly evaluate the performance of different schemes, this experiment uses two evaluation metrics, accuracy (ACC) and equal error rate (EER), to evaluate the performance of the model, and the DET curve is used to more clearly compare the performance of the model. The accuracy formula is as follows:

[0161]

[0162] EER comprehensively measures and balances the performance of the false rejection rate (FRR) and the false acceptance rate (FAR). By adjusting the threshold, when FAR and FRR are equal, the values of FAR and FRR at this time are the EER. The definitions of FAR and FRR are as follows:

[0163]

[0164]

[0165] Among them, FP refers to the number of examples that are originally negative samples but are judged as positive samples; TN refers to the number of examples that are originally negative samples and are also judged as negative samples; TP refers to the number of examples that are originally positive samples and are also judged as positive samples; FN refers to the number of examples that are originally positive samples but are judged as negative samples.

[0166] The DET curve (Detection Error Tradeoff Curve) is a commonly used classifier evaluation metric and plays an objective role in evaluating the performance of the model. It is a curve with the false positive rate (False Positive Rate, FPR) and the false negative rate (False Negative Rate, FNR) as the coordinate axes, which helps us weigh the trade-off relationship between the false positive rate and the false negative rate of the classifier.

[0167] Under normal circumstances, the DET curve is a curve in the lower left direction of the coordinate axis. In this embodiment, the performance of classifiers can be judged by comparing the area and shape of the DET curves of the classifiers. The larger the area enclosed by the curve and the upper right coordinate axis or the closer it is to the coordinate origin (0, 0), the better the performance of the classifier.

[0168] 2.4. Convolution Kernel Size Setting of VC-Gabor Convolution Module

[0169] The size of the VC-Gabor convolution kernel is an important parameter in the convolution operation. Convolution kernels of different sizes have different response capabilities to images. To select the VC-Gabor convolution kernel size that makes the model performance optimal as much as possible, in this experiment, other conditions were the same, only the size of the VC-Gabor convolution kernel was different. K = 3, K = 9, and K = 15 were taken in turn, and experiments were carried out on the FV-USM dataset and the SDUMLA dataset. The experimental results are shown in Table 2:

[0170] Table 1: Performance Comparison of the Proposed Scheme of the Invention under Different VC-Gabor Convolution Kernels

[0171]

[0172] It can be seen from Table 2 that when the convolution kernel is too small (3×3) or too large (15×15), the feature extraction ability cannot reach the optimal. In this experiment, when the convolution kernel size K = 9, the accuracy and EER of the model are relatively optimal. Therefore, in subsequent experiments, the size of the VC-Gabor convolution kernel is set to K = 9.

[0173] 2.5. Model Visualization

[0174] To prove that the model proposed in this embodiment can better learn the features of the vein region, in this experiment, a conventional CNN convolution scheme was used as the control group, and the feature maps learned by the control group and the VG-Gabor layer in this case were visualized. Figure 13 Shows the convolution results of the first convolution kernels of the trained ordinary convolutional neural network CNN and the VC-Gabor convolutional neural network on the same input image.

[0175] It can be seen from the figure that the feature map learned by the VC-Gabor convolution kernel (right) is richer in vein curves than the feature map learned by the ordinary convolution kernel (left). The VC-Gabor filter makes the curved parts in the vein curve more obvious, which shows that the VC-Gabor convolution filter proposed in this embodiment can effectively extract the vein line features in the finger vein image.

[0176] 2.6. Ablation Experiment

[0177] To evaluate the contribution of the VC-Gabor convolutional layer and the multi-branch feature extraction structure used in the model proposed in this embodiment to the algorithm performance, this experiment designed several control groups for ablation experiments and conducted experiments on the FV-USM and SDUMLA datasets respectively.

[0178] The settings of each experimental group in the ablation experiment are as follows:

[0179] (1) CNN+SBS (single-branch Structure): The first convolutional layer of the network uses ordinary convolutional filters, and all subsequent feature extraction blocks use a single-branch structure.

[0180] (2) CNN+MBS (multi-branch structure): The first convolutional layer of the network uses ordinary convolutional filters, and all subsequent feature extraction blocks use a multi-branch structure.

[0181] (3) VC-Gabor+SBS: The first convolutional layer of the network uses VC-Gabor convolutional filters, and all subsequent feature extraction blocks use a single-branch structure.

[0182] (4) Gabor+MBS: The first convolutional layer of the network uses Gabor convolutional filters, and all subsequent feature extraction blocks use a multi-branch structure.

[0183] (5) VC-Gabor+MBS: The first convolutional layer of the network uses VC-Gabor convolutional filters, and all subsequent feature extraction blocks use a multi-branch structure.

[0184] After the ablation experiment, the experimental results of each experimental group tested on the FV-USM and SDUMLA datasets are shown in Table 3:

[0185] Table 3: Test results of each experimental group in the ablation experiment

[0186]

[0187]

[0188] The DET curves obtained from the experiments of the network models of experimental groups 2, 4, and 5 on the FV-USM and SDUMLA datasets are plotted respectively as Figure 14 and Figure 15 shown. Combining Table 3 and the data in Figure 14 and Figure 15 it can be known that:

[0189] Under the condition of using the multi-branch structure feature extraction, replacing the first layer of the network with Gabor convolution can improve the recognition accuracy of the network model by 0.06% and 0.09% respectively on the USM dataset and the SDU dataset, and the equal error rates are reduced by 0.06% and 0.04% respectively. The experimental results show that adding Gabor convolution to the neural network can effectively extract the direction and scale information of vein patterns, thereby improving the recognition rate and reducing the equal error rate. When the first layer of the network is replaced with VC-Gabor convolution, the recognition accuracy is further improved and the equal error rate is further decreased. Compared with ordinary convolution, the accuracy is increased by 0.13% and 0.33% respectively on the two datasets, and the equal error rates are decreased by 0.10% and 0.14% respectively. This shows that the VC-Gabor filter proposed in this embodiment can more effectively extract the curvature information of vein images, which is beneficial to improving the feature extraction ability of the model.

[0190] Furthermore, the DET curves obtained from the experiments of the network models of experimental groups 1, 2, 3, and 5 on the FV-USM and SDUMLA datasets are plotted respectively as Figure 16 and Figure 17 shown. Combining Table 3 and the data in Figure 16 and Figure 17 , it can be seen that:

[0191] Under the condition of ensuring the same other conditions, using multi-branch structure feature extraction has better results than single-branch structure. The accuracy rates are increased by 0.21% and 0.36% respectively on the two datasets, and the equal error rates are reduced by 0.13% and 0.09% respectively. Similarly, by comparing experimental group 2 and experimental group 5, it can also be found that the experimental results of the multi-branch structure are still better than those of the single-branch structure. The accuracy rates are increased by 0.19% and 0.20% respectively on the two datasets, and the equal error rates are reduced by 0.16% and 0.10% respectively. Therefore, the adaptive multi-branch structure feature proposed in this embodiment is beneficial to improving the feature extraction ability of finger vein images, thereby improving the recognition performance.

[0192] In summary, whether under the multi-branch condition or the single-branch condition, through experiments, it can be proved that the experimental effects of the variable curvature Gabor convolution layer and the multi-branch structure proposed in this embodiment have been improved on the FV-USM dataset and the SDUMLA dataset.

[0193] 2.7. Comparative Experiments

[0194] To verify the effectiveness of the proposed scheme in this paper, this experiment also compares this case with existing finger vein recognition schemes based on traditional methods and deep learning. The scheme of this embodiment as the control group includes:

[0195] “

[12] Yakun Zhang,Weijun Li,Liping Zhang,Xin Ning,Linjun Sun,Yaxuan Lu: Adaptive Learning Gabor Filter for Finger-Vein Recognition. IEEE Access 7: 159821-159830(2019)

[0196]

[32] Yang L,Yang G,Yin Y,et al. Sliding window-based region of interest extraction for finger vein images[J]. Sensors,2013,13(3): 3799-3815.

[0197]

[33] Qiu S,Liu Y,Zhou Y,et al. Finger-vein recognition based on dual-sliding window localization and pseudo-elliptical transformer[J]. Expert Systems with Applications,2016,64: 618-632.

[0198]

[34] Huang B,Dai Y,Li R,et al. Finger-vein authentication based on wide line detector and pattern normalization[C] / / 2010 20th international conference on pattern recognition. IEEE,2010: 1269-1272.

[0199]

[35] H. Hong,M. Lee,and K. Park, “Convolutional neural network-based finger-vein recognition using NIR image sensors,” Sensors,vol.17,no.6,p.1297, Jun. 2017.

[0200]

[36] Zhao D, Ma H, Yang Z, et al. Finger vein recognition based on lightweight CNN combining center loss and dynamic regularization[J]. Infrared Physics & Technology, 2020, 105: 103221

[0201]

[37] Huang J, Tu M, Yang W, et al. Joint Attention Network for Finger Vein Authentication[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1 - 11.

[0202]

[38] Li Y, Lu H, Wang Y, et al. ViT - Cap: A Novel Vision Transformer - Based Capsule Network Model for Finger Vein Recognition[J]. Applied Sciences, 2022, 12(20): 10364.

[0203]

[39] Zhongxia Zhang, Mingwen Wang: Finger vein recognition based on lightweight convolutional attention model. IET Image Process. 17(6): 1864 - 1873(2023)”

[0204] The performance evaluation results of the proposed solution in this embodiment and the control group solution on the FV - USM dataset and SDUMLA dataset are shown in Tables 4 and 5 respectively:

[0205] Table 4: Performance evaluation results of this case and the control group solution on the FV - USM dataset

[0206]

[0207]

[0208] Table 5: Performance evaluation results of this case and the control group solution on the SDUMLA - HMT dataset

[0209]

[0210] Analysis of the data in Table 4 and Table 5 shows that:

[0211] Compared with the traditional finger vein recognition schemes

[32]

[33]

[34] , the scheme proposed in this embodiment has the highest recognition accuracy improved by 4.74% and 3.51% respectively on the FV-USM and SDUMLA datasets, and the equal error rate reduced by 2.56% and 2.26% respectively at most. This shows that the finger vein recognition method based on VC-Gabor convolution and multi-branch structure proposed in this embodiment can extract deeper features and has better performance.

[0212] Compared with the deep learning-based finger vein recognition schemes proposed by Hong et al.

[35] , Zhao et al.

[36] , Huang et al.

[37] , Li et al.

[38] , and Zhang et al.

[39] , the scheme proposed in this embodiment has the highest recognition accuracy improved by 5.31% and 4.73% respectively on the FV-USM and SDUMLA datasets, and the equal error rate reduced by 4.15% and 3.77% respectively at most. This shows that the multi-scale features extracted by the finger vein recognition method based on VC-Gabor convolution and multi-branch structure proposed in this embodiment are more discriminative for distinguishing different categories of finger vein images.

[0213] Compared with the adaptive AGCNN method proposed by Zhang et al.

[12] , the network in this embodiment has the accuracy improved by 7.65% and 4.64% respectively on the two datasets, and the equal error rate reduced by 4.01% and 3.32%. This shows that the improved VC-Gabor convolution layer and multi-branch network structure in this embodiment can significantly improve the recognition performance of finger vein images.

[0214] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A finger vein recognition method based on VC-Gabor convolution and multi-branch structure, characterized in that: It includes the following steps: S1: Design a Gabor filter with variable curvature. The variable curvature Gabor filter is defined as follows: In the above formula, x and y represent the pixel coordinate position; θ represents the angle between the direction of the sine function and the x-axis, which is the direction of the kernel function of the filter; λ represents the wavelength of the sine function; f represents the curvature of the filter; σ is the standard deviation of the Gaussian function; i represents the imaginary unit; S2: Construct a parameter-adjustable VC-Gabor convolution based on the variable curvature Gabor filter; the construction process of the VC-Gabor convolution is as follows: S21: Decompose the variable curvature Gabor filter into the real part G re and the imaginary part G im ; S22: Use the decomposed real part to construct VC-Gabor convolution, and define θ, λ, f and σ as learnable optimization parameters; S23: Calculate the real part G of the Gabor filter with variable curvature re , perform partial derivatives on each optimization parameter θ, λ, f and σ respectively; S24: Update G using the gradient descent rule of the backpropagation algorithm re Parameters and establish update formulas for each optimization parameter; S3: Improve the classic CNN network to construct the required finger vein recognition model. The input of the finger vein recognition model is the finger vein image, and the output is the category of the finger vein image. The model construction process is as follows: S31: Obtain a classic CNN network, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and a Softmax classification layer; the convolutional layer includes a downsampling convolutional layer and a non-downsampling convolutional layer; S32: Use the newly designed VC-Gabor convolution as the first layer of the CNN network; S33: Replace all downsampling convolution layers in the convolution layer with two-branch convolution modules; S34: Replace the non-subsampled convolutional layer in the convolutional layer with a three-branch convolutional module; S4: Obtain finger vein images of a large number of users as sample data to form a dataset, divide the dataset into training set, validation set and test set, and add labels to the sample data; S5: training the finger vein recognition model using the training set, and verifying the model using the verification set; S6: Using the trained finger vein recognition model to perform finger vein image recognition on a test set, and outputting category information of the current user.

2. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 1, characterized in that: In step 21, the real number part G is split re It can effectively extract the edge and texture information of the finger vein image. The function is as follows: The decomposed imaginary part G im The function is as follows:

3. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 2, characterized in that: In step S23, the partial derivative function of each optimization parameter is as follows: Among them, x'=xcosθ+ysinθ; y'=-xsinθ+ycosθ.

4. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 3, characterized in that: In step S24, the update formula of each optimization parameter is: In the above formula, θ t ,λ t 、f t and σ t is the optimization parameter before updating; θ t+1 ,λ t+1 、f t+1 and σ t+1 is the updated optimization parameter; α is the preset learning rate; L is the preset objective function.

5. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 4, characterized in that: In step S3, the input of the constructed finger vein recognition model is a grayscale image with a channel number c=1; The number of channels of the VC-Gabor convolutional layer is 40, and the parameters θ, λ, f, and σ that need to be optimized are updated during the learning process; In the convolution layer, the size of the feature map obtained by each downsampling convolution module is halved and the number of channels is doubled; the size and number of channels of the feature map obtained by each non-downsampling convolution module remain unchanged; The pooling layer adopts a global average pooling layer; The number of neurons in the fully connected layer is the total number of categories in the dataset; The Softmax classification layer is used to convert the original output into a probability distribution, thereby performing a multi-category classification task.

6. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 5, characterized in that: In the dual-branch convolution module of step S33, The input feature map enters two branches respectively, and completes 3×3 convolution and batch normalization in the first branch, and completes 1×1 convolution and batch normalization in the second branch; then the adaptive weights of each branch are superimposed and fused, and the fused features are processed by the ReLU activation layer and output; The expression of the dual-branch convolution module is as follows: Y=ReLU(w1·BN(Conv1(X))+w2·BN(Conv2(X))) In the above formula, X represents the input feature map; Y represents the output feature map; Conv1 represents 3×3 convolution; Conv2 represents 1×1 convolution; BN represents batch normalization; ReLU is the activation function; w1 and w2 are the adaptive weights of the first branch and the second branch respectively.

7. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 5, characterized in that: In the three-branch convolution module of step S34, The input feature map enters three branches respectively, and completes 3×3 convolution and batch normalization in the first branch, 1×1 convolution and batch normalization in the second branch, and only batch normalization in the third branch. The adaptive weights of each branch are then superimposed and fused, and the fused features are processed by the ReLU activation layer and output. The expression of the three-branch convolution module is as follows: Y=ReLU(w1·BN(Conv1(X))+w2·BN(Conv2(X))+w3·BN(X)) In the above formula, X represents the input feature map; Y represents the output feature map; Conv1 represents 3×3 convolution; Conv2 represents 1×1 convolution; BN represents batch normalization; ReLU is the activation function; w1, w2, and w3 are the adaptive weights of the first branch, second branch, and third branch, respectively.

8. The finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to claim 5, characterized in that: In step S4, the acquired data set is randomly augmented with data samples using data augmentation methods including rotation, translation, flipping, and simulating illumination changes, and the labels of the augmented samples are consistent with those of the original samples; In step S5, the batch size of the finger vein recognition model in the training phase is set to 32, and the Adam optimizer is used for training. The initial learning rate is 0.001, and the learning rate is adjusted according to the cosine annealing learning rate adjustment strategy with a period of 20. The objective function is set as follows: In the above formula, C is the number of samples, y i One-hot representation of sample labels: When the sample belongs to the i-th category, y i =1, otherwise y i =0;p i It represents the probability that the sample belongs to the i-th category.

9. A finger vein recognition system, characterized by: It is used to identify the category of an input finger vein image using the finger vein recognition method based on VC-Gabor convolution and multi-branch structure as described in any one of claims 1 to 8; the finger vein recognition system includes: An image acquisition device, which is used to acquire an original finger vein image of the current user to be identified; An image preprocessing module is used to preprocess the user's original finger vein image and convert it into a grayscale image that meets the resolution requirements; A template library storing finger vein feature vectors and corresponding labels for all registered users; a finger vein recognition model whose input is the preprocessed grayscale image and whose output is a matching relationship between the current user and registered users; the finger vein recognition model is a network model designed and trained using the finger vein recognition method based on VC-Gabor convolution and a multi-branch structure as described in any one of claims 1 to 8; the network model comprises a multi-channel filter, a feature extraction module, a fully connected layer, and a classification layer; the multi-channel filter is automatically generated by the VC-Gabor convolution module and is used to extract the direction, scale, and curvature features of the finger vein image texture; the feature extraction module uses a convolution layer including a two-branch convolution module and a three-branch convolution module; the feature extraction module is used to extract multi-scale features of the surface and deep layers of the finger vein image; the number of neurons in the fully connected layer is equal to the total number of categories in the dataset; the fully connected layer connects the input node of the global average pooling layer to each output node of the current layer to perform feature extraction and conversion; the classification layer uses a Softmax function to convert the raw output into a probability distribution, thereby matching the current user with the labels of all registered users in the template library.

10. A finger vein authentication device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the computer program, it creates an image preprocessing module, a template library, and part of a finger vein recognition model in the finger vein recognition system according to claim 9; and then adopts the steps of the finger vein recognition method based on VC-Gabor convolution and multi-branch structure according to any one of claims 1 to 8 to recognize the collected finger vein image of the current user and output a matching relationship between the current user and the registered users.

Images