A finger vein recognition method, model, device and medium

Abstract

This application relates to the field of biometric recognition and image processing technology, and discloses a method, model, device, and medium for finger vein recognition based on a lightweight attention model with few samples. The method includes: acquiring raw finger vein images and extracting regions of interest (ROIs); preprocessing the ROI images; performing data augmentation on the training set images; constructing an LA-VGG network based on a VGG-19 network, introducing CA attention, PReLU, and cross-layer fusion to enhance subtle textures; introducing a loss function with additive angular margins during the training phase; training the LA-VGG network to obtain a finger vein recognition model; inputting the image to be recognized into the model, and outputting the identity recognition result. This invention can effectively alleviate the overfitting problem under small sample conditions, reduce the number of model parameters and computational complexity, improve the extraction capability of key finger vein texture features, and is suitable for deployment and application in resource-constrained devices and complex environments.

Description

Technical Field

[0001] This invention relates to the fields of biometric recognition and artificial intelligence image processing technology, and in particular to a finger vein recognition method, model, electronic device, and computer-readable storage medium based on a few-sample lightweight attention model. Background Technology

[0002] Finger vein recognition is a liveness detection technology that uses near-infrared imaging to acquire information about the vein patterns inside the finger and perform identity authentication. Because human vein patterns are located under the skin, they are difficult to copy and forge, and possess strong individual uniqueness. Therefore, finger vein recognition has high application value in areas such as access control, financial identity authentication, smart terminal verification, and security management systems.

[0003] However, existing finger vein recognition technology still faces at least the following two problems in practical applications.

[0004] First, finger vein data often involves personal privacy, and the collection process is limited by factors such as the scenario, equipment, and user cooperation, making it difficult to obtain large-scale, balanced training samples. In actual deployments, the number of samples available for collection from each user is often small, resulting in a limited training set size. This makes deep learning models prone to overfitting and insufficient generalization ability during training, thus affecting the system's recognition stability and accuracy in real-world environments.

[0005] Secondly, to improve recognition accuracy, many existing solutions typically employ deep convolutional neural networks such as VGG and ResNet for feature extraction. While these networks possess strong representational capabilities, they generally suffer from problems such as large parameter count, high computational complexity, slow inference speed, and high requirements for storage resources and computing power, making them unsuitable for deployment in embedded terminals, portable recognition devices, or edge computing platforms. In particular, the original VGG-19 network has a large number of parameters, with fully connected layers accounting for the majority of the parameter overhead, making it difficult to meet the requirements of lightweight design and real-time performance.

[0006] Furthermore, finger vein images themselves suffer from low contrast, high imaging noise, significant background interference, uneven local brightness, and deviations in acquisition posture. Especially in low-temperature environments, slowed blood flow and vasoconstriction further reduce vein contrast, blur edges, and lower signal-to-noise ratio, leading to further deterioration in recognition performance. Existing solutions typically focus on improvements in a single area, such as image enhancement or network lightweighting, and rarely offer a comprehensive system design for finger vein recognition systems that simultaneously address the challenges of small sample data, complex acquisition environments, and lightweight deployment requirements.

[0007] Therefore, there is an urgent need for a finger vein recognition method that can adapt to small sample training scenarios, balance recognition accuracy and model lightweight requirements, and enhance the model's ability to represent key vein pattern features, in order to solve the problems of insufficient samples, high model complexity, and insufficient environmental adaptability in existing technologies. Summary of the Invention

[0008] To address the aforementioned shortcomings in existing technologies, the present invention aims to provide a finger vein recognition method, model, electronic device, and computer-readable storage medium. The method is based on a lightweight attention model with few samples. This method constructs an image preprocessing and data augmentation workflow tailored to few-sample scenarios and performs lightweighting and attention enhancement modifications on the classic VGG-19 network. While maintaining recognition accuracy, it significantly reduces the number of model parameters and computational complexity, improves the model's deployability on resource-constrained devices, and enhances the model's robustness under complex conditions such as low temperature, acquisition offset, brightness variations, and noise disturbances.

[0009] This invention provides the following technical solution: a finger vein recognition method based on a few-shot lightweight attention model, applied in the finger vein recognition process, the method comprising the following steps:

[0010] S1. Obtain the original finger vein image and extract the region of interest from the original finger vein image to obtain a finger vein region of interest image that only contains the effective area of ​​the finger;

[0011] S2. The image of the region of interest of the finger vein is preprocessed to enhance the vein pattern features and suppress background interference and noise, so as to obtain the preprocessed finger vein image.

[0012] S3. Perform data augmentation on the preprocessed finger vein images in the training set to increase the number of training samples and sample diversity. The data augmentation includes at least one of geometric transformation and pixel perturbation, and includes degradation simulation augmentation for imaging characteristics in low temperature environments.

[0013] S4. Construct a lightweight attention-enhanced finger vein recognition network (LA-VGG network or LA-VGG network model). The lightweight attention-enhanced finger vein recognition network is an improvement on the VGG-19 network, and the improvements include:

[0014] The VGG-19 network is structurally pruned to reduce the number of parameters and computational cost of convolutional and fully connected layers.

[0015] Replace the ReLU activation function with the Parametric ReLU activation function;

[0016] A coordinate attention module (CA module) is embedded after the convolutional block, which embeds spatial location information into the channel weights by performing global pooling in the horizontal and vertical directions;

[0017] A multi-scale feature fusion path is established to spatially concatenate the edge features extracted by shallow convolutional blocks with deep semantic features to achieve feature compensation.

[0018] S5. The lightweight attention-enhanced finger vein recognition network is trained using the amplified training set to obtain the trained finger vein recognition model.

[0019] An additive angular margin loss function is introduced during training to enhance the model's discriminative performance under small sample conditions by increasing the angular distance between different categories of features on the hypersphere.

[0020] S6. Input the finger vein image to be identified into the trained finger vein recognition model and output the corresponding identity recognition result.

[0021] Further, step S1 includes:

[0022] Adaptive threshold segmentation or edge detection is performed on the original finger vein image to obtain finger contour information;

[0023] The finger circumscribed region is determined based on the detection of the maximum connected component and / or convex hull.

[0024] Based on the finger aspect ratio and position prior, the circumscribed region of the finger is fine-tuned and standardized for cropping to obtain the region of interest image of the finger vein.

[0025] Further, step S2 includes:

[0026] Local contrast enhancement is performed on the region of interest image of the finger vein using contrast-limited adaptive histogram equalization.

[0027] Gaussian filtering is used to smooth the enhanced image in order to suppress image noise;

[0028] A multi-directional, multi-scale Gabor filter bank is used to enhance the vein patterns in the smoothed image, and the filtering response results of each direction are fused to obtain the preprocessed finger vein image.

[0029] Furthermore, the Gaussian filter kernel size is 5×5 or 7×7, and the standard deviation σ is 1.0 to 1.5;

[0030] The Gabor filter bank includes at least two of the four directions: 0°, 45°, 90°, and 135°, and uses one to two frequency scales to filter the image.

[0031] Furthermore, the data amplification in step S3 includes at least two of the following methods:

[0032] Geometric transformations, including one or more of random rotations, random translations, and flips;

[0033] Pixel transformation, including one or more of random brightness adjustment and random contrast adjustment;

[0034] Noise injection includes adding one or both of Gaussian noise and salt-and-pepper noise.

[0035] Furthermore, the range of the random rotation angle is ±15°;

[0036] The range of the random translation is ±10% of the image height and / or width;

[0037] The random brightness adjustment uses a random scale factor of 0.8 to 1.2;

[0038] The flipping is a horizontal flipping performed with a preset probability;

[0039] The degradation simulation amplification for imaging characteristics in low-temperature environments involves applying Gaussian blur with a standard deviation σ of not less than 2.0 to the preprocessed finger vein image and simultaneously reducing the image contrast to simulate the blurred vein patterns and weakened edges caused by slowed blood flow and vasoconstriction in low-temperature environments.

[0040] Furthermore, in step S4:

[0041] The structural pruning includes retaining the first three convolutional blocks of the VGG-19 network and removing the conv4_3 and conv4_4 layers from the fourth convolutional block;

[0042] At the same time, the second fully connected layer in the VGG-19 network is removed to reduce the number of model parameters and computational complexity;

[0043] The activation function employs Parametric ReLU, whose negative half-axis contains a learnable slope parameter α. The slope parameter α is automatically updated iteratively during model training to adaptively adjust the response intensity of different feature channels to negative information.

[0044] The attention module is a coordinate attention module (CA module), which is embedded after the last convolutional layer and before the activation function in each of the retained convolutional blocks;

[0045] The convolutional block attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule generates channel attention weights based on global average pooling and global max pooling, and the spatial attention submodule generates spatial attention weights based on average pooling and max pooling along the channel dimension.

[0046] This application also provides a finger vein recognition model, which is trained using the finger vein recognition method according to any one of claims 1-7. The finger vein recognition model is a lightweight attention enhancement network used to extract features from the input finger vein image, classify the categories, and output the identity recognition result.

[0047] This application also provides an electronic device, including:

[0048] One or more processors;

[0049] Memory;

[0050] And computer programs stored in the memory and capable of running on the one or more processors;

[0051] When the one or more processors execute the computer program, they implement the finger vein recognition method as described above, and / or call the finger vein recognition model as described above to perform identity recognition on the finger vein image to be recognized.

[0052] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the finger vein recognition method as described above, and / or invokes the finger vein recognition model as described above to perform identity recognition on the finger vein image to be recognized.

[0053] The beneficial effects of this invention are as follows:

[0054] 1. This invention constructs a hierarchical image preprocessing and augmentation mechanism suitable for small sample finger vein scenarios. Through ROI extraction, image enhancement, multi-type data augmentation, and low-temperature degradation simulation, it effectively improves the quantity, quality, and distribution diversity of training samples, fundamentally alleviating the overfitting problem in the small sample training process.

[0055] 2. This invention makes targeted lightweight modifications to the classic VGG-19 network. By removing some convolutional layers and fully connected layers, the model parameter size and computational complexity are greatly reduced. While maintaining strong feature extraction capabilities, the feasibility of deploying the model on embedded devices and resource-constrained terminals is significantly improved.

[0056] 3. This invention uses Parametric ReLU instead of traditional ReLU, which effectively alleviates the problem of neuron death caused by the gradient being zero in the negative range, and enables the network to maintain stable gradient propagation and good training dynamics during the training process after lightweight pruning.

[0057] 4. This invention embeds a CA module to capture the spatial location information of vein patterns, replaces ReLU with Parametric ReLU, and establishes a cross-layer feature fusion path to preserve subtle textures. During the training phase, a loss function with additive angular margins is introduced to optimize the feature space distribution. This allows the model to focus more on the key feature channels and key spatial regions of vein patterns, reducing interference from background noise and irrelevant skin textures, thereby further improving recognition accuracy and robustness. This invention introduces a CA coordinate attention mechanism into the improved network. By decomposing channel attention into one-dimensional feature encoding along both horizontal and vertical directions, the model can capture cross-channel orientation perception and position-sensitive information. This mechanism enables the network to accurately locate the key topological coordinates of finger vein patterns, effectively suppressing background noise and interference from finger edges, thus significantly improving recognition accuracy and robustness with extremely low parameter increments.

[0058] 5. This invention addresses the degradation of venous imaging caused by slowed blood flow and vasoconstriction in low-temperature environments by constructing simulated data, thereby enhancing the model's adaptability to complex environments and special usage scenarios. Attached Figure Description

[0059] The invention will now be described in more detail with reference to embodiments and the accompanying drawings.

[0060] Figure 1 This is a schematic diagram of the overall process of the lightweight finger vein recognition method based on small samples according to the present invention.

[0061] Figure 2 A schematic flowchart of a finger vein recognition method 100 provided according to an embodiment of the present disclosure is shown.

[0062] Figure 3 A flowchart illustrating a method 200 for extracting the region of interest from the original finger vein image is shown.

[0063] Figure 4 This is a schematic diagram of ROI extraction from finger vein images in this invention.

[0064] Figure 5 A flowchart of a preprocessing method 300 for enhancing vein texture features and suppressing background interference and noise in a region of interest image of a finger vein is shown.

[0065] Figure 6This is a schematic diagram of image enhancement processing in this invention.

[0066] Figure 7 This is a schematic diagram of the data augmentation strategy in this invention.

[0067] Figure 8 This is a schematic diagram of the Gaussian filtering process applied to two finger vein images taken in a low-temperature environment, as described in this invention.

[0068] Figure 9 This is a schematic diagram of the improved VGG-19 network, also known as the LA-VGG network structure, in this invention.

[0069] Figure 10 This is a schematic diagram of the coordinate attention mechanism in this invention.

[0070] Figure 11 This is a schematic diagram of the multi-scale feature fusion path mechanism in this invention.

[0071] Figure 12 This is a classic, unmodified VGG-19 network architecture diagram. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Raw images of finger veins often suffer from low contrast, background interference, uneven lighting, and noise. To improve image quality and highlight vein features, Figure 1 A modular schematic diagram of the method 100 provided by the present invention is shown, such as... Figure 1 As shown, the finger vein recognition method based on a small-sample lightweight attention model provided by the present invention generally includes a preprocessing stage, a feature extraction stage, and a result output stage.

[0074] The preprocessing stage comprises two parts: data augmentation and image enhancement. Data augmentation includes image enhancement operations such as histogram equalization, Gaussian filtering, Gabor filtering, and bilateral filtering. Image enhancement includes training sample augmentation methods such as flipping, rotation, translation, brightness adjustment, and noise addition. The processed image is then input into the feature extraction stage. In this stage, the processed image is fed into an improved VGG-19 network, where key features are enhanced using spatial and channel attention mechanisms, ultimately outputting the recognition result. Figure 2A flowchart illustrating a finger vein recognition method 100 according to an embodiment of the present disclosure is shown. Method 100 is applied during finger vein recognition and includes the following steps:

[0075] S1. Obtain the original finger vein image and extract the region of interest from the original finger vein image to obtain a finger vein region of interest image that only contains the effective area of ​​the finger;

[0076] S2. The image of the region of interest of the finger vein is preprocessed to enhance the vein pattern features and suppress background interference and noise, so as to obtain the preprocessed finger vein image.

[0077] S3. Perform data augmentation on the preprocessed finger vein images in the training set to increase the number of training samples and sample diversity. The data augmentation includes at least one of geometric transformation and pixel perturbation, and includes degradation simulation augmentation for imaging characteristics in low temperature environments.

[0078] S4. Construct a lightweight attention-enhanced finger vein recognition network (LA-VGG network or LA-VGG network model). The lightweight attention-enhanced finger vein recognition network is an improvement on the VGG-19 network, and the improvements include:

[0079] The VGG-19 network is structurally pruned to reduce the number of parameters and computational cost of convolutional and fully connected layers.

[0080] Replace the ReLU activation function with the Parametric ReLU activation function;

[0081] A coordinate attention module is embedded after the convolutional block, which embeds spatial location information into the channel weights by performing global pooling in the horizontal and vertical directions;

[0082] A multi-scale feature fusion path is established to spatially concatenate the edge features extracted by shallow convolutional blocks with deep semantic features to achieve feature compensation.

[0083] S5. The LA-VGG is trained using the amplified training set to obtain the trained finger vein recognition model.

[0084] An additive angular margin loss function is introduced during training to enhance the model's discriminative performance under small sample conditions by increasing the angular distance between different categories of features on the hypersphere.

[0085] S6. Input the finger vein image to be identified into the trained finger vein recognition model and output the corresponding identity recognition result.

[0086] It should be noted that, as Figure 1As shown, the preprocessing module and feature extraction module of the implementation method 100 of this invention are not simply connected in series, but rather jointly serve the model training and testing recognition process under small sample conditions. That is, in the training phase, the input image is processed by ROI extraction, enhancement processing, and augmentation processing before entering the LA-VGG model for training; in the inference phase, the image to be recognized is processed by at least ROI extraction and basic enhancement processing before entering the model to complete the recognition.

[0087] Extracting the region of interest (ROI) from the original finger vein image and eliminating background areas unrelated to the finger allows computational resources to be concentrated on the effective vein information, improving subsequent processing efficiency and model recognition accuracy. In some embodiments, an automatic ROI extraction algorithm based on edge detection and contour analysis is employed. Figure 3 A flowchart illustrating a method 200 for extracting a region of interest from the original finger vein image is shown. Method 200 includes the following steps:

[0088] S11. Perform adaptive threshold segmentation or edge detection on the original finger vein image to obtain finger contour information;

[0089] S12. Determine the finger circumscribing region based on the detection of the maximum connected component and / or convex hull;

[0090] S13. Based on the finger aspect ratio and position prior, the external region of the finger is fine-tuned and standardized for cropping to obtain the region of interest image of the finger vein.

[0091] In one embodiment, ROI extraction can be performed according to the following steps:

[0092] S11. Perform Canny edge detection on the original finger vein image to extract the finger edge contour information. Since finger vein images are usually accompanied by irrelevant information such as background, fixtures, noise shadows, etc. in the acquisition window, contour detection is used first to separate the main finger area from the original image.

[0093] S12. The main finger regions in the image are located through maximum connected component filtering and convex hull detection. For the extracted edge contours, the connected regions with the largest area are retained first, and the approximate circumscribed rectangular region of the finger is determined by combining shape features.

[0094] S13. Based on the geometric prior of the finger, perform fine-tuning and standardized cropping of the position. For example, combine the proportional relationship between the length and width of the finger to modify the boundary of the circumscribed rectangle so that the cropped area maintains a relatively uniform position and scale among different samples.

[0095] Therefore, a Region of Interest (ROI) image containing only the effective area of ​​the finger can be extracted from the original finger vein image. ROI extraction is as follows: Figure 4As shown, Figure 4 The image on the left shows the original finger vein image, with the red dashed box marking the Region of Interest (ROI) to be cropped. The image on the right shows the cropped ROI. After this step, the background area, device edge area, and irrelevant dark areas are effectively removed, facilitating subsequent image enhancement and feature learning. Thus, the ROI extraction method 200 can extract an ROI image containing only the effective area of ​​the finger from the original finger vein image. By extracting the region of interest from the original finger vein image using the above method, interference from invalid information can be reduced, allowing subsequent computational resources to be concentrated on the effective vein information area, thereby improving the accuracy and efficiency of subsequent feature extraction and classification.

[0096] After the ROI extraction is completed, in order to further improve the visibility and distinguishability of vein pattern features, the present invention performs image enhancement processing on the ROI image sequentially. Figure 5 A flowchart of a preprocessing method 300 for enhancing vein texture features and suppressing background interference and noise in a region of interest image of a finger vein is shown. Method 300 includes the following steps:

[0097] S21. Local contrast enhancement is performed on the region of interest image of the finger vein using contrast-limited adaptive histogram equalization.

[0098] S22. Apply Gaussian filtering to smooth the enhanced image in order to suppress image noise;

[0099] S23. A multi-directional, multi-scale Gabor filter bank is used to enhance the vein patterns of the smoothed image, and the filtering response results of each direction are fused to obtain the preprocessed finger vein image.

[0100] like Figure 6 As shown, Method 300 follows the standard image processing flow of "first enhancing contrast, then smoothing noise, and finally extracting features".

[0101] In one embodiment, method 300 includes the following steps:

[0102] S21. First, Limit Contrast Adaptive Histogram Equalization (CLAHE) is used to process the ROI image. The image is divided into small blocks and each block is subjected to independent histogram equalization, which effectively solves the problems of background noise amplification and local over-brightness / over-darkness that may be caused by global histogram equalization.

[0103] S22. Use Gaussian filtering to suppress and smooth noise. Set the filter kernel size to (5×5) or (7×7), and the standard deviation σ to between 1.0 and 1.5 depending on the image resolution. This step provides a cleaner input for subsequent Gabor filtering.

[0104] S23. Gabor filtering is used to enhance vein pattern features. A multi-directional, multi-scale Gabor filter bank is designed. Through experiments, for finger vein images, four typical directions (0°, 45°, 90°, and 135°) are selected, and one or two optimal frequency scales are chosen to filter the smoothed image. Finally, the most significant features in the filter response maps of each direction are fused to obtain an output image with extremely enhanced vein patterns.

[0105] To address the overfitting problem caused by small sample training in finger vein recognition, this invention implements a rigorous data augmentation strategy on the training set images after preprocessing. This significantly increases the diversity and quantity of data, expands the sample distribution, and improves the model's adaptability to pose shifts, brightness changes, and noise disturbances. Figure 7 As shown, the present invention can perform one or more combinations of the following amplification operations on the original ROI image.

[0106] 1) Geometric transformation amplification

[0107] Geometric transformations are used to simulate the user's posture changes and positional shifts during actual data acquisition. In one embodiment, the following operations are included:

[0108] Random rotation: The image is randomly rotated within a range of ±15° to simulate the situation where a finger is placed at a slight tilt on the acquisition device.

[0109] Random translation: Translation transformation is performed within ±10% of the image height and width to simulate the positional shift of the finger within the imaging window.

[0110] Horizontal Flip: Performs a horizontal flip operation with a preset probability to increase the geometric diversity of image samples. In some embodiments, the preset probability can be a probability value of 0.35, 0.5, or 0.75.

[0111] Vertical flip: Under certain implementation conditions, images can also be vertically flipped to further expand the training sample space.

[0112] Proportional scaling: In one embodiment, the image may also be scaled proportionally to enhance the model’s tolerance to changes in imaging scale.

[0113] Figure 7 Centered on the original image, the diagram illustrates various geometric transformations such as vertical flipping, horizontal flipping, 90-degree flipping, image movement, and proportional magnification.

[0114] 2) Pixel-level transformation amplification

[0115] Pixel-level transformation is used to simulate changes in device imaging conditions and ambient lighting conditions. In one embodiment, it includes:

[0116] Random brightness adjustment: Image pixels are adjusted using a random scaling factor between 0.8 and 1.2 to simulate different exposure levels and lighting conditions.

[0117] Random contrast adjustment: Changes the image contrast within a set range to adapt the model to fluctuations in image quality.

[0118] Figure 7 The image also illustrates two processing methods: increasing brightness and decreasing brightness.

[0119] 3) Noise injection amplification

[0120] To improve the model's robustness to imaging noise, this invention also adds random noise to the image. In one embodiment, Gaussian noise or salt-and-pepper noise may be added. Figure 7 This illustrates the image effect after adding salt and pepper noise.

[0121] Through the above-mentioned combined data augmentation method, the present invention can expand the limited training samples into training datasets that are tens or even hundreds of times larger, which helps to alleviate the problem of model overfitting under small sample conditions.

[0122] In some application scenarios, such as outdoor environments in northern winters, low-temperature storage environments, and around cold chain equipment, finger vein recognition devices may operate under low-temperature conditions. Low temperatures can cause peripheral blood flow to slow down and blood vessels to constrict, resulting in near-infrared vein images exhibiting characteristics such as decreased contrast, blurred vein patterns, and weakened edges.

[0123] To enhance the model's adaptability to such special environments, this invention further designs a low-temperature environment data simulation method.

[0124] like Figure 8 As shown, two different preprocessed, clearer finger vein images were subjected to Gaussian filtering with a large standard deviation, and the overall image contrast was appropriately reduced to simulate the blurred imaging effect caused by the weakening of venous blood flow signals under low temperature conditions.

[0125] In one embodiment, the standard deviation σ of the Gaussian filter can be set to 2.0 or higher, allowing for a wider filtering range and resulting in softer vein edges and reduced detail. Simultaneously, by reducing contrast, the difference between light and dark gray is minimized, more closely resembling images acquired under low-temperature conditions.

[0126] Figure 8 The left and right sides illustrate the gradual blurring effect of Gaussian filtering on two different simulated low-temperature vein images. When these simulated images are added to the training set as amplified samples, the robustness of the model to low-temperature conditions can be effectively improved.

[0127] We need to add a model diagram of the VGG19 before any modifications.

[0128] The existing VGG19 network model has more than 140 million parameters, of which the fully connected layer accounts for more than 90%. The large number of parameters not only requires a lot of storage resources, but also leads to slow model inference speed, which cannot meet the lightweight requirements of embedded devices (such as portable finger vein recognition terminals).

[0129] like Figure 12 As shown, the existing VGG19 network model uses ReLU as the activation function. When the input value is negative, the ReLU function outputs 0, causing the gradient of the corresponding neuron to remain 0 during backpropagation, thus permanently "dying" the neuron and losing its feature extraction ability. In finger vein images, the pixel values ​​of some key features (such as the edges of vein textures) are relatively small and easily affected by the ReLU function, leading to the loss of feature information. The core of finger vein recognition lies in extracting key features such as the topological structure and branch details of vein textures. However, the original VGG19 network uses an "equal" feature extraction method, failing to distinguish the importance of different regions in the image. It may treat non-critical information such as background noise and finger skin texture equally with vein features, reducing the model's utilization of effective features in small sample data and affecting recognition accuracy.

[0130] To address the aforementioned issues, this algorithm is based on the classic VGG-19 network model in existing technologies. Through targeted structural optimization and mechanism improvements, it aims to comprehensively reconstruct the classic VGG-19 network, significantly reducing model complexity and parameter count while maintaining its powerful feature extraction capabilities. An attention mechanism is also introduced to enhance its ability to represent subtle finger vein features. The improvements mainly include three aspects: 1) structured pruning and channel compression; 2) activation function optimization; and 3) embedding of coordinate attention mechanism and optimization of the discriminative loss function. The improved network is called LA-VGG (Lightweight Attention-aware VGG).

[0131] Regarding the improvement of the LA-VGG network over the classic VGG-19 network in this application, in aspect 1): the VGG-19 network is structurally pruned to reduce the number of parameters and computational cost of convolutional and fully connected layers. For example... Figure 9 As shown, this is the lightweight attention-enhanced finger vein recognition network constructed in step S4 of this application, hereinafter referred to as the LA-VGG network. The feature extraction part of the LA-VGG network mainly includes multiple convolutional blocks and pooling layers.

[0132] The first convolutional block corresponds to an input image size of approximately 112×112×3. This first convolutional block includes Conv1_1, Parametric ReLU1, Conv1_2, Parametric ReLU2, and pool1. Specifically, it extracts shallow edge and texture information through two convolutional layers, with Parametric ReLU activation applied after each convolutional layer, followed by a first pooling operation to reduce spatial resolution and increase the receptive field. This part... Figure 9 The dashed box ① indicates that the ReLU activation function at the corresponding position in the original network has been replaced with the Parametric ReLU activation function. Specifically, this is reflected in the activation mapping process between the convolutional output and pooling input of the first and second convolutional blocks. That is, the output of Conv1_1 first enters ParametricReLU1 and then is fed into Conv1_2, and the output of Conv1_2 enters Parametric ReLU1_1 and then is fed into pool1; similarly, the output of Conv2_1 enters Parametric ReLU2_1 and then is fed into Conv2_2. This replacement does not change the tensor size flow relationship; that is, the spatial size marked by brackets before and after the activation layer remains consistent with the number of channels. However, in terms of numerical mapping, it ensures that the negative half-axis still retains a non-zero gradient, thus maintaining more stable gradient propagation after network pruning, which is beneficial for training stability under small sample conditions.

[0133] In one embodiment, the negative half-axis slope parameter α of Parametric ReLU is initialized to, for example, 0.25. Unlike activation functions with a fixed slope, this α value is adaptively updated during training as the finger vein feature distribution iterates, thus allowing gradients to propagate at an optimal scale even in regions where the input is negative.

[0134] The second convolutional block includes Conv2_1, Parametric ReLU, Conv2_2, Parametric ReLU, and pool2. This convolutional block is used to extract local texture and shape information at a deeper level, and the output feature map size is further downsampled from the previous layer.

[0135] The third convolutional block includes Conv3_1, Conv3_2, Conv3_3, Conv3_4, and their corresponding activation layers and pool3. This convolutional block is an important layer for extracting vein pattern features, effectively capturing branching structures, linear texture directions, and local topological relationships.

[0136] The fourth convolutional block in the original VGG-19 typically includes four convolutional layers: Conv4_1, Conv4_2, Conv4_3, Conv4_4, and pool4. This invention has performed a structured pruning at this point, such as... Figure 9As shown in the dashed box ②, the corresponding layers of Conv4_3 and Conv4_4 were deleted. That is, while retaining the main structure of the first three convolutional blocks (the structure up to conv3_4), the 12th and 13th layers (i.e., conv4_3 and conv4_4) in the fourth convolutional block were further deleted to reduce redundant parameters and computational burden in the deep layers. At the same time, Conv4_1 and Conv4_2 were retained for further extraction of higher-level texture combination features. This modification effectively reduced a large number of parameters while avoiding the loss of feature granularity that might be caused by removing the entire deep block, achieving a better balance between model capacity and computational efficiency.

[0137] Volume 5 blocks in Figure 9 The diagram still retains some deep convolutional and pooling structures to form the final high-level semantic representation. This part can be adapted to different hardware platforms and recognition accuracy requirements during implementation, but the overall design principle follows the principle of reducing deep redundancy and preserving the expressive power of key layers through pruning.

[0138] From an overall structural perspective, this invention does not simply scale the number of channels in each layer proportionally. Instead, it selectively deletes and retains layers of the original VGG-19 network to address the characteristics of fine-grained linear texture tasks such as finger veins. This allows the model to reduce complexity while maintaining a strong ability to represent texture details.

[0139] Figure 9 The feature extraction part of the LA-VGG network consists of multiple convolutional blocks and pooling layers cascaded together, following a forward connectivity relationship from left to right. That is, the input image first enters the first convolutional block to complete shallow texture extraction, then enters the second and third convolutional blocks to extract richer texture combination features, and then enters the fourth and fifth convolutional blocks to extract higher-level features. Finally, after the last pooling, a high-level feature tensor for classification is obtained and sent to the subsequent attention module and classification head.

[0140] The input image size corresponding to the first convolutional block is... Figure 9The input image is labeled (112, 112, 3). After passing through Conv1_1, the feature map is obtained as (112, 112, 64). Then, it enters the parametric activation layer Parametric ReLU1 for adaptive nonlinear mapping, maintaining the output size as (112, 112, 64). It then enters Conv1_2 to obtain (112, 112, 64), and after passing through Parametric ReLU1_1, the output is still (112, 112, 64). This output then enters pool1 for downsampling, resulting in (56, 56, 64). Thus, the connection relationship within the first convolutional block is that the input is sequentially connected in series with Conv1_1, Parametric ReLU1_1, Conv1_2, Parametric ReLU1_2, and then connected to pool1. The convolutional layers mainly improve the channel representation capability, while the pooling layers reduce the spatial resolution from 112×112 to 56×56 to expand the receptive field and reduce the subsequent computational cost.

[0141] The input to the second convolutional block comes from the output of pool1 (56, 56, 64), which is then passed through Conv2_1 to obtain (56, 56, 128), and then through Parametric ReLU2_1 to maintain the output at (56, 56, 128). This output then enters Conv2_2 to obtain (56, 56, 128), and then through Parametric ReLU2_2 to maintain the output at (56, 56, 128). This output is further passed to pool2 for downsampling, resulting in (28, 28, 128). Therefore, the internal connection relationship of the second convolutional block is that the output of pool1 is sequentially connected in series with Conv2_1, Parametric ReLU2_1, Conv2_2, and ReLU2_2, and then connected in parallel with pool2, achieving an expansion from 64 channels to 128 channels while reducing the spatial size from 56×56 to 28×28.

[0142] The input to the third convolutional block comes from the output of pool2 (28, 28, 128), and is sequentially passed through Conv3_1 to obtain (28, 28, 256), and then passed through Parametric ReLU3_1 to maintain the same size. Subsequently, it passes through Conv3_2, Parametric ReLU3_2, Conv3_3, Parametric ReLU3_3, Conv3_4, and Parametric ReLU3_4. The spatial size of the output of each convolutional layer within the third convolutional block remains (28, 28), and the number of channels remains 256. After four layers of convolution and activation, the output is downsampled into pool3 to obtain (14, 14, 256). Thus, the third convolutional block enhances the expressive power of texture patterns through continuous convolution stacking, and further compresses the spatial size through pool3 to form more stable mid-to-high-level features.

[0143] The input to the fourth convolutional block comes from the output of pool3 (14, 14, 256), which is then processed by Conv4_1 and ParametricReLU4_1 to obtain (14, 14, 512), and then processed by Conv4_2 and Parametric ReLU4_2 to maintain (14, 14, 512). In the original VGG-19, the fourth convolutional block usually also includes two more layers, Conv4_3 and Conv4_4, to further deepen semantic abstraction, but this invention... Figure 9 As shown in section ②, the block is structurally pruned by deleting Conv4_4 and Conv4_3 layers, thus reducing deep redundant computations while maintaining the necessary representation capabilities of the fourth convolutional block. The output of the fourth convolutional block is downsampled by pool4 to obtain (7, 7, 512) and then fed into the fifth convolutional block.

[0144] The input to the fifth convolutional block comes from the output of pool4 (7, 7, 512), which is then processed by Conv5_1 and ParametricReLU5_1, resulting in the same output (7, 7, 512). This is followed by Conv5_2 and Parametric ReLU5_2, and then by Conv5_3 and Parametric ReLU5_3, again yielding the same output (7, 7, 512). The result is then downsampled in pool5 to (3, 3, 512). Thus, the feature extraction part completes the step-by-step mapping from the input (112, 112, 3) to the higher-level feature (3, 3, 512), gradually compressing the spatial size and expanding the number of channels. This results in output features with both strong discriminative power and low computational cost, facilitating subsequent attention enhancement and classification.

[0145] Regarding the improvement of the LA-VGG network over the classic VGG-19 network in this application (section 2), the ReLU activation function is replaced with the Parametric ReLU activation function. While the standard ReLU activation function in the existing VGG19 network model can alleviate the gradient vanishing problem, its "hard" zero region characteristic may cause negative gradients to permanently fail, i.e., the "DyingReLU" problem, especially after network depth pruning, where the stability of the gradient flow becomes even more important. We uniformly replace all standard ReLU activation functions in the network with Parametric ReLU. This invention replaces the standard ReLU activation function in the network with the Parametric ReLU activation function: This function retains a relatively small slope even when the input is negative, thus preventing the gradient from vanishing completely in the negative interval.

[0146] In one embodiment, the negative half-axis slope parameter α of Parametric ReLU is initialized to, for example, 0.25. Unlike activation functions with a fixed slope, Parametric ReLU dynamically optimizes the α value of each channel during training through backpropagation. This ensures that gradients can still propagate according to feature requirements within the negative interval, guaranteeing that all neurons are effectively updated throughout the training cycle. Allowing negative activations to exist in an optimal proportion provides a richer feature flow and finer feature discrimination for the lightweight model, especially in regions with extremely low contrast between finger vein patterns and the background, enabling adaptive capture of weak vascular signals. This parameterization mechanism allows the network to maintain a stable gradient distribution even after significant structural pruning, greatly improving the convergence speed and feature representation depth of the model under small sample conditions.

[0147] The improvement of the LA-VGG network over the classic VGG-19 network in this application (section 3) is the introduction of a cross-layer multi-scale feature fusion path to compensate for the loss of detail information. For example... Figure 9 As shown in number ③, this invention establishes a skip connection from Block 2 to the deep feature layer. Since finger vein recognition highly depends on subtle texture topology, and depth convolution and downsampling processes easily lead to the loss of shallow spatial information, this embodiment extracts the feature map (56, 56, 128) containing rich edge details from the output of Block 2, performs channel alignment via 1×1 convolution, and then concatenates it spatially with the deep semantic features output from the fifth convolution block. This multi-scale fusion mechanism achieves feature complementarity between shallow subtle textures and deep global semantics, significantly enhancing the model's robustness to low-quality finger vein images.

[0148] Regarding the improvement of the LA-VGG network over the classic VGG-19 network in this application (4th aspect): a coordinate attention module is embedded in the preserved convolutional block structure to enhance the representation ability of key finger vein features. For example... Figure 9 As shown in number ④, in order to compensate for the potential decrease in representation ability after network pruning and to improve the model's ability to focus on key vein patterns, that is, to enable the network to further focus on the key channels and key spatial regions of finger vein patterns before the output enters the classification head, this invention introduces a coordinate attention module, namely the CA module, into the classic VGG-19 network. Specifically, the CA module is added between the feature extraction output and the classification input, that is, embedded between the two.

[0149] Figure 9 The connection relationships of the coordinate attention module are shown. The high-level feature tensor (3,3,512) output by pool5 first enters the coordinate attention module for feature enhancement. This module introduces spatial location information encoding during channel attention modeling, performs global information aggregation on the feature map in the horizontal and vertical directions respectively, and generates attention weights for the corresponding directions, thereby achieving joint modeling of feature channels and spatial location information. Since both channel attention and spatial attention are recalibration operations on the weights of the original features, the input and output brackets of the attention module have the same size, but it redistributes the effective information distribution of the features, making the network more inclined to retain vein pattern-related responses and suppress background or irrelevant texture responses.

[0150] In one embodiment, the CA module is embedded between the retained convolutional block output and the classification input. That is, after the convolutional block completes feature extraction and outputs a feature map, the features are first enhanced by the coordinate attention module before entering the subsequent classification structure.

[0151] The Coordinate Attention (CA) module performs directional decomposition encoding on the feature map. While maintaining the modeling of inter-channel dependencies, it embeds positional information into the channel attention representation, thereby more accurately capturing the spatial distribution information of the target structure. In the finger vein recognition task, this mechanism effectively highlights the features of the vein pattern region while suppressing background regions and noise information, thus improving the model's ability to express key vein features.

[0152] like Figure 9As shown in number ⑤, the classification part of the classic VGG-19 network in existing technologies typically includes multiple high-dimensional fully connected layers, with the second fully connected layer having a large number of parameters. The classic VGG-19 network usually has three fully connected layers at the classification end: fc6 (4096), fc7 (4096), and fc8 (number of categories, 1000 for ImageNet), with fc8 being the third fully connected layer, followed by a Softmax layer. To reduce model complexity, this invention removes the second fully connected layers fc7 and fc8 from the classic VGG-19 network in existing technologies, retaining only layer fc6 as the final mapping layer for feature vectors. At the classification output end, this invention introduces an ArcFace loss function layer with additive angular margins, which replaces the original combination of fc8 and simple Softmax by imposing angular interval constraints on finger vein categories in the feature space. This structural modification of the classification head, combined with a metric learning strategy, not only significantly reduces the number of parameters and memory usage but also greatly improves the model's ability to distinguish similar finger vein textures under small sample conditions.

[0153] Figure 9 The right side illustrates the structure of the classification section. The input is a feature map output from the feature extraction section, which contains multi-scale fusion information resulting from the concatenation of shallow features extracted by the second convolutional block and deep features from the fifth convolutional block. After mapping through the fully connected layer fc6 and subsequent classification layers, the final category result is output. Because this invention removes some fully connected layers from the original network, the classification head is more suitable for lightweight deployment. More specifically, Figure 9 As shown in number ④, the output of pool5 after attention enhancement is (3, 3, 512). This feature is first used as input to the classification head and is vectorized before entering the fully connected layer, so that the spatial and channel dimensions are organized into a feature representation that can be processed by the fully connected layer. Then, this feature enters fc6 ReLU6, with an output of (1, 1, 4096). The traditional Softmax classification layer is abandoned, and an ArcFace loss function layer with additive angular margins is used for identity classification, with an output of (1, 1, 64), where 64 corresponds to the number of target identity categories. By compressing intra-class distances and increasing inter-class angular margins in the feature space, an end-to-end recognition process from the input image to the high-discriminative feature output is completed.

[0154] In conclusion, Figure 9 The LA-VGG network shown balances the accuracy requirements under small-sample training conditions with the deployment requirements in resource-constrained platforms by means of activation function optimization, convolutional layer pruning, coordinate attention module embedding, multi-scale feature fusion, and application of metric learning loss function.

[0155] In other embodiments, a Coordinate Attention (CA) module is embedded in each retained convolutional block (conv1_x, conv2_x, conv3_x). This is an attention mechanism capable of simultaneously capturing cross-channel orientation awareness and precise location information, offering better spatial localization capabilities in lightweight models compared to CBAM. The processing mechanism of the Coordinate Attention (CA) module is as follows: Figure 10 As shown, its processing includes:

[0156] First, coordinate information embedding: global average pooling is performed on the input feature map along the horizontal (X-direction) and vertical (Y-direction) directions respectively to aggregate the spatial information of each dimension and form a pair of orientation-aware feature description vectors.

[0157] Secondly, coordinate mapping is generated: the two one-dimensional vectors are concatenated and input into a shared convolutional layer for dimensionality reduction mapping, and a non-linear transformation is performed using the Parametric ReLU activation function. Subsequently, the transformed feature vector is re-sliced ​​along the spatial dimension into two components: horizontal attention and vertical attention.

[0158] Finally, the output is weighted by location: the two components are converted into attention weights using the Sigmoid function and applied to the horizontal and vertical coordinates of the original input feature map, respectively. This bidirectional weighting mechanism not only emphasizes feature channels rich in vein information but also pinpoints the precise spatial coordinates of vein patterns, such as the bifurcation points and intersections of blood vessels.

[0159] This mechanism enables the network to automatically capture the topological information of finger veins while weakening noisy regions and irrelevant skin textures, significantly improving the discriminative power of features. The coordinate attention mechanism can more accurately identify the spatial location of vein patterns, allowing the model to focus on the main veins, branches, and intersections, and reducing the interference of uniform background regions on classification results. Through position-aware weighting, the spatial region containing vein patterns in the image is accurately located and enhanced, while uniform background regions are ignored.

[0160] In one embodiment, to further overcome the information loss caused by lightweight pruning, the LA-VGG model introduces a multi-scale feature fusion mechanism. For example... Figure 11 This demonstrates a multi-scale fusion path mechanism that extracts edge detail features from shallow convolutional blocks (Block 2) through skip connections and then concatenates them with deep semantic features processed by coordinate attention.

[0161] Multi-scale feature fusion can more accurately compensate for the lack of expression of microvascular terminals in deep networks, enabling the model to focus on the main vein while also taking into account branches and extremely fine texture regions, significantly reducing recognition errors in small sample scenarios. For example... Figure 11 As shown, in the multi-scale feature fusion path, the input is the shallow detail feature F. shallow With deep semantic features F deep First, shallow features are channel-aligned, and then the two results are concatenated along the channel dimension to generate a rich feature descriptor. Subsequently, through a feature integration layer, the model is guided to automatically learn the key representations of finger veins at different receptive field scales. It can accurately locate and enhance the spatial regions where vein patterns are located in an image, ignoring uniform background areas. The synergistic effect of multi-scale fusion and coordinate attention allows the model to focus on the main veins, branches, and intersections, while reducing the interference of uniform background areas on the classification results.

[0162] The synergistic effect of multi-scale fusion and coordinate attention enables the model to focus on the main vein, branches and intersection regions, and reduces the interference of uniform background regions on the classification results.

[0163] In one embodiment of this disclosure, an electronic device for implementing finger vein recognition is also provided. The electronic device includes one or more processors and a memory. The processor may be a central processing unit (CPU) or other processing units with data processing and instruction execution capabilities, such as a graphics processing unit (GPU), a digital signal processor (DSP), or a field-programmable gate array (FPGA), used to control various functional modules in the electronic device and execute predetermined data processing flows. The memory may include one or more computer program products, which may contain various forms of computer-readable storage media, including volatile and non-volatile storage media. Volatile storage media include, for example, random access memory (RAM) and cache memory, while non-volatile storage media include, for example, read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD). Computer program instructions are stored on the computer-readable storage medium, and when the processor executes the computer program instructions, it implements the finger vein recognition method based on a few-sample lightweight attention model described in this disclosure.

[0164] In one embodiment, the processor, when executing the computer program instructions, can perform the following functional flow: First, it acquires raw image data of the finger veins and performs region of interest extraction processing on the raw image to obtain a region of interest image of the finger veins containing only the effective area of ​​the finger. Then, it performs image preprocessing on the region of interest image to enhance vein texture features and suppress noise and background interference. The preprocessing may include operations such as contrast enhancement, smoothing and denoising, and texture enhancement. Further, during the training phase, the processor can perform data augmentation on the training set images to increase the number and diversity of training samples. The data augmentation may include geometric transformation, pixel transformation, and noise injection, and may further include blurring and contrast attenuation processing to simulate the imaging degradation characteristics of low-temperature environments. Then, the processor constructs and trains a lightweight attention enhancement network. This network is based on a VGG-type backbone network, reducing the number of network parameters and computational cost through structural pruning, improving training stability by using an activation function with a negative half-axis slope, and capturing precise spatial location information of the vein texture by embedding a coordinate attention module. Simultaneously, it utilizes a cross-layer multi-scale feature fusion path to compensate for the loss of detail in deep features during the lightweighting process, thereby improving the representation ability of key vein topological features. During training, the processor employs a metric learning loss function to increase the inter-class discrimination distance. After training, the processor calls the trained finger vein recognition model to perform inference calculations on the finger vein image to be recognized, outputting the identity recognition result or category discrimination result, which can be used for subsequent identity authentication, access control, or security management processes.

[0165] In one embodiment, the electronic device may further include an input device and an output device, and the input device and output device may be communicatively connected to the processor and memory via a bus system or other connection mechanism. The input device may be used to receive image acquisition data or external control commands, and may include an image acquisition module, buttons, a touch unit, or other human-computer interaction components; the image acquisition module may be a near-infrared imaging device for acquiring images of finger vein patterns. The output device may be used to output recognition results, status information, or prompt information, and may include a display, indicator light, speaker, printer, or other output unit. Depending on the specific application scenario, the electronic device may also include a communication interface or communication module for uploading recognition results to a server or interacting with external terminals; the communication method may include wired communication or wireless communication.

[0166] In addition, depending on the specific application requirements, the electronic device may also include input / output interfaces, buses, clock units, hardware acceleration units or other appropriate functional components to support the implementation of functions such as image processing, model training, model inference and result output in the finger vein recognition process.

Claims

1. A finger vein recognition method based on a few-shot lightweight attention model, characterized in that, When applied to finger vein recognition, the method includes the following steps: S1. Obtain the original finger vein image and extract the region of interest from the original finger vein image to obtain a finger vein region of interest image that only contains the effective area of ​​the finger; S2. The image of the region of interest of the finger vein is preprocessed to enhance the vein pattern features and suppress background interference and noise, so as to obtain the preprocessed finger vein image. S3. Perform data augmentation on the preprocessed finger vein images in the training set to increase the number of training samples and sample diversity. The data augmentation includes at least one of geometric transformation and pixel perturbation, and includes degradation simulation augmentation for imaging characteristics in low temperature environments. S4. Construct a lightweight attention-enhanced finger vein recognition network, which is an improvement upon the VGG-19 network. The improvements include: The VGG-19 network is structurally pruned to reduce the number of parameters and computational cost of convolutional and fully connected layers. Replace the ReLU activation function with the Parametric ReLU activation function; A coordinate attention module is embedded after the convolutional block, which embeds spatial location information into the channel weights by performing global pooling in the horizontal and vertical directions; A multi-scale feature fusion path is established to spatially concatenate the edge features extracted by shallow convolutional blocks with deep semantic features to achieve feature compensation. S5. The lightweight attention-enhanced finger vein recognition network is trained using the amplified training set to obtain the trained finger vein recognition model. An additive angular margin loss function is introduced during training to enhance the model's discriminative performance under small sample conditions by increasing the angular distance between different categories of features on the hypersphere. S6. Input the finger vein image to be identified into the trained finger vein recognition model and output the corresponding identity recognition result.

2. The finger vein recognition method according to claim 1, characterized in that, Step S1 includes: Adaptive threshold segmentation or edge detection is performed on the original finger vein image to obtain finger contour information; The finger circumscribed region is determined based on the detection of the maximum connected component and / or convex hull. Based on the finger aspect ratio and position prior, the circumscribed region of the finger is fine-tuned and standardized for cropping to obtain the region of interest image of the finger vein.

3. The finger vein recognition method according to claim 1, characterized in that, Step S2 includes: Local contrast enhancement is performed on the region of interest image of the finger vein using contrast-limited adaptive histogram equalization. Gaussian filtering is used to smooth the enhanced image in order to suppress image noise; A multi-directional, multi-scale Gabor filter bank is used to enhance the vein patterns in the smoothed image, and the filtering response results of each direction are fused to obtain the preprocessed finger vein image.

4. The finger vein recognition method according to claim 3, characterized in that, The Gaussian filter has a kernel size of 5×5 or 7×7 and a standard deviation σ of 1.0 to 1.

5. The Gabor filter bank includes at least two of the four directions: 0°, 45°, 90°, and 135°, and uses one to two frequency scales to filter the image.

5. The finger vein recognition method according to claim 1, characterized in that, The data amplification in step S3 includes at least two of the following methods: Geometric transformations, including one or more of random rotations, random translations, and flips; Pixel transformation, including one or more of random brightness adjustment and random contrast adjustment; Noise injection includes adding one or both of Gaussian noise and salt-and-pepper noise.

6. The finger vein recognition method according to claim 5, characterized in that, The angle range of the random rotation is ±15°; The range of the random translation is ±10% of the image height and / or width; The random brightness adjustment uses a random scale factor of 0.8 to 1.2; The flipping is a horizontal flipping performed with a preset probability; The degradation simulation amplification for imaging characteristics in low-temperature environments involves applying Gaussian blur with a standard deviation σ of not less than 2.0 to the preprocessed finger vein image and simultaneously reducing the image contrast to simulate the blurred vein patterns and weakened edges caused by slowed blood flow and vasoconstriction in low-temperature environments.

7. The finger vein recognition method according to claim 1, characterized in that, In step S4: The structural pruning includes retaining the first three convolutional blocks of the VGG-19 network and removing the conv4_3 and conv4_4 layers from the fourth convolutional block; At the same time, the second fully connected layer in the VGG-19 network is removed to reduce the number of model parameters and computational complexity; The activation function employs Parametric ReLU, whose negative half-axis contains a learnable slope parameter α. The slope parameter α is automatically updated iteratively during model training to adaptively adjust the response intensity of different feature channels to negative information. The attention module is a coordinate attention module, which is embedded after the last convolutional layer and before the activation function in each of the retained convolutional blocks; The convolutional block attention module includes a channel attention submodule and a spatial attention submodule. The channel attention submodule generates channel attention weights based on global average pooling and global max pooling, and the spatial attention submodule generates spatial attention weights based on average pooling and max pooling along the channel dimension.

8. A finger vein recognition model, characterized in that, The finger vein recognition model is trained using the finger vein recognition method described in any one of claims 1-7. The finger vein recognition model is a lightweight attention enhancement network used to extract features from the input finger vein image, classify the categories, and output the identity recognition result.

9. An electronic device, characterized in that, include: One or more processors; Memory; And computer programs stored in the memory and capable of running on the one or more processors; When the one or more processors execute the computer program, they implement the finger vein recognition method as described in any one of claims 1 to 7, and / or call the finger vein recognition model as described in claim 8 to perform identity recognition on the finger vein image to be recognized.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the finger vein recognition method as described in any one of claims 1-7, and / or calls the finger vein recognition model as described in claim 8 to perform identity recognition on the finger vein image to be recognized.

Images