A finger vein recognition method and system based on improved MobileNetV2

Abstract

A finger vein recognition method based on improved MobileNetV2, comprising the following steps: S1, obtaining a finger vein image, and preprocessing the finger vein image to serve as an input image; S2, constructing an improved MobileNetV2 network, including an initial convolutional layer, a plurality of reverse residual block stages, a final convolutional layer and a classifier; S3, embedding an SE attention mechanism in the last four reverse residual blocks of the improved MobileNetV2 network to form SE reverse residual blocks, and keeping the first three reverse residual block stages as standard reverse residual blocks; S4, inputting the preprocessed finger vein image in step S1 into the improved MobileNetV2 network for feature extraction and classification recognition, and outputting a recognition result. The present application embeds an SE attention mechanism in the deep network selectively, thereby significantly improving the finger vein recognition accuracy while maintaining the lightweight characteristics, and enhancing the robustness to noise, illumination changes, local occlusions and other interferences.

Description

Technical Field

[0001] This invention belongs to the field of biometric recognition and deep learning technology, and in particular relates to a finger vein recognition method and system based on an improved MobileNetV2. Background Technology

[0002] In the digital age, the dual demands for security and convenience in identity authentication technology are becoming increasingly prominent. Traditional identity authentication methods (such as passwords, keys, and access cards) have significant limitations and are difficult to adapt to the requirements of high-security application scenarios. While biometric features such as fingerprints, faces, voiceprints, and irises are widely used in personal identity authentication, the rapid development of artificial intelligence technology makes these external biometric features easily forged and copied, leading to identity fraud and posing extremely high security risks. Finger vein recognition, as an emerging technology in the field of biometrics, has received widespread attention in recent years. Compared to other biometric technologies, finger vein recognition possesses core advantages such as liveness detection, feature stability, and individual uniqueness.

[0003] Deep learning is a key technology in the field of machine learning and is currently widely used in image recognition, speech recognition, natural language processing, and many other fields. In recent years, driven by the rapid development of computer vision technology, deep learning has also been gradually introduced into the field of finger vein recognition. Existing deep learning-based finger vein recognition methods mainly include the following categories:

[0004] 1. Yang et al. used generative adversarial networks for finger vein research. Instead of relying on training to extract image features, they mined the intrinsic relationship between images and templates by generating similar images. Although the recognition effect was good, there were problems such as unstable training results and difficulty in predicting the time required for ideal results.

[0005] 2. Song et al. used a composite image of two finger vein images as input to the DenseNet-161 network and introduced shift matching to solve the image misalignment problem. However, due to the large number of layers and large size, although the network has excellent performance, it cannot be put into engineering practice, which also highlights the inevitable trend of lightweight deep learning models.

[0006] 3. Li Jinghui introduced transfer learning into the finger vein algorithm, compared the "small sample direct learning" and "pre-training-fine-tuning" modes of different networks, and proved the feasibility of transfer learning. It also shortened the model learning time and improved convergence and accuracy.

[0007] 4. Wang Zhilin et al. proposed a finger vein recognition network that integrates multiple semantic features and combines ArcFace loss to complete the recognition. It uses shallow networks to extract vein structure information and enhances the recognition ability of vein geometric information through cross-layer fusion.

[0008] 5. Shaheed et al. used a deep separable convolutional model to extract finger vein features, shortened the learning time of the target dataset by pre-training, and simplified the computational complexity by relying on the separable convolutional InceptionV3 structure, achieving good results.

[0009] 6. Mou Jiale et al. proposed a lightweight finger vein recognition algorithm based on deep residual networks. Based on ResNet34, they added depthwise separable convolution and SE attention mechanism to improve performance. Then, through knowledge distillation training, they solved the performance loss problem caused by the reduction of parameters.

[0010] MobileNetV2, as a classic lightweight convolutional neural network, achieves a good balance between accuracy and efficiency through inverse residual structures and depthwise separable convolutions. However, when directly applied to finger vein recognition tasks, the following technical problems arise:

[0011] 1. Insufficient learning of key features: MobileNetV2’s inverted residual blocks adopt an “expansion-convolution-compression” structure, which may lose fine-grained discriminative features such as vein branches and intersections during the compression stage.

[0012] 2. Limited generalization ability with small sample data: Finger vein datasets are usually small in size (only a few to a dozen images per class), and MobileNetV2 is prone to overfitting on such data.

[0013] 3. Global embedding of attention mechanism leads to redundancy: If the SE attention mechanism is simply embedded in all inverted residual blocks, although it can improve the feature representation ability, it will introduce a large number of additional parameters and computational overhead. Moreover, the shallow feature semantics are simple and the redundancy is high, so the effect of the attention mechanism is limited and may even interfere with the original feature extraction.

[0014] Therefore, developing a lightweight finger vein recognition algorithm that balances recognition accuracy and operational efficiency to address the above issues and ensure high accuracy and rapid recognition capability of finger vein recognition systems in resource-constrained scenarios is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0015] The purpose of this invention is to overcome the shortcomings of the prior art and provide a finger vein recognition method and system based on an improved MobileNetV2. By selectively embedding the SE attention mechanism deep into the network, the accuracy of finger vein recognition is significantly improved while maintaining lightweight characteristics, and the robustness to interference such as noise, lighting changes, and local occlusion is enhanced.

[0016] To achieve the objectives of this invention, the technical solution adopted is as follows:

[0017] A finger vein recognition method based on an improved MobileNetV2 includes the following steps:

[0018] S1. Obtain a finger vein image, preprocess it, and use it as the input image;

[0019] S2. Construct an improved MobileNetV2 network, including an initial convolutional layer, several inverted residual block stages, a final convolutional layer, and a classifier;

[0020] S3. In the improved MobileNetV2 network, the SE attention mechanism is embedded in the last four inverted residual blocks to form SE inverted residual blocks, while the first three inverted residual block stages remain as standard inverted residual blocks;

[0021] S4. Input the preprocessed finger vein image from step S1 into the improved MobileNetV2 network for feature extraction and classification, and output the recognition results.

[0022] Furthermore, the size of the input image in step S1 is 224×224×3.

[0023] Furthermore, in step S2, the width multiplier of the improved MobileNetV2 network is set to 1.0; the kernel size of the initial convolutional layer is 3×3, the stride is 2, and the number of output channels is 32; the classifier is a fully connected layer, and the number of output channels is equal to the number of finger vein identity categories N.

[0024] Furthermore, in step S3, the last four SE inverted residual blocks include an extended convolutional layer, a depthwise separable convolutional layer, an SE attention layer, and a projective convolutional layer connected in sequence; the extended convolutional layer is used to increase the channel dimension; the depthwise separable convolutional layer is used to extract spatial features; the SE attention layer is used to adaptively calibrate the channel weights; and the projective convolutional layer is used to reduce the high-dimensional features to the number of output channels.

[0025] Furthermore, in step S3, the feature map sizes corresponding to the last four inverted residual blocks are 28×28, 14×14, 14×14 and 7×7, respectively.

[0026] Further, step S4 specifically involves: inputting the preprocessed finger vein image from step S1 into the improved MobileNetV2 network; finally, the convolutional layer maps the high-dimensional feature map to the classification feature space; the classifier outputs the probability distribution of each identity category; and the category corresponding to the maximum value is taken as the recognition result.

[0027] The present invention also provides a finger vein recognition system based on an improved MobileNetV2 for implementing the above method, characterized in that the system comprises:

[0028] Image acquisition and preprocessing module: used to acquire finger vein images and preprocess them into images to be recognized with a size of 224×224×3;

[0029] Network building module: used to build an improved MobileNetV2 network, which embeds the SE attention mechanism in the last four inverted residual blocks of MobileNetV2;

[0030] Network training module: used to train the network on a public or self-built finger vein dataset, and update the network parameters using the Adam optimizer and cross-entropy loss function;

[0031] Feature extraction and recognition module: This module takes the preprocessed image to be recognized and inputs it into the trained network. After forward propagation, it outputs the identity recognition result.

[0032] Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0033] 1. It provides a new approach for finger vein recognition. Compared with the original MobileNetV2 network, the improved MobileNetV2 solves the problems of insufficient focus on key vein features and easy overfitting in the original network.

[0034] 2. The improved MobileNetV2 in this invention embeds the SE attention module in the semantically rich Stage of the network back end. Through adaptive calibration of channel weights, it strengthens the expression of discriminative features such as vein branches and intersections, further optimizes the attention module and feature fusion logic, and improves robustness in complex scenarios.

[0035] 3. The improved network achieved recognition accuracies of 99.49% and 97.06% on the FVUSM and SDUMLA datasets, respectively, which are improvements compared to the basic network. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the SE attention mechanism structure in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the SE inverted residual block in an embodiment of the present invention;

[0038] Figure 3 These are partial images from the FVUSM dataset in this embodiment of the invention;

[0039] Figure 4 These are partial images of the SDUMLA dataset in an embodiment of the present invention;

[0040] Figure 5This is a graph showing the accuracy and loss of each model in this embodiment on two public dataset validation sets. Detailed Implementation

[0041] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0042] Example 1:

[0043] This embodiment provides a finger vein recognition method based on an improved MobileNetV2, including the following steps:

[0044] S1. Obtain a finger vein image, preprocess it, and use it as the input image with a size of 224×224×3.

[0045] S2. Construct an improved MobileNetV2 network, including an initial convolutional layer, several inverted residual block stages, a final convolutional layer, and a classifier; the width multiplier of the improved MobileNetV2 network is set to 1.0; the kernel size of the initial convolutional layer is 3×3, the stride is 2, and the number of output channels is 32; the classifier is a fully connected layer, and the number of output channels is equal to the number of finger vein identity categories N.

[0046] S3. In the improved MobileNetV2 network, the SE attention mechanism is embedded in the last four inverted residual blocks to form SE inverted residual blocks. The first three inverted residual blocks remain as standard inverted residual blocks. The last four SE inverted residual blocks include an extended convolutional layer, a depthwise separable convolutional layer, an SE attention layer, and a projective convolutional layer connected in sequence. The extended convolutional layer is used to increase the channel dimension. The depthwise separable convolutional layer is used to extract spatial features. The SE attention layer is used to adaptively calibrate the channel weights. The projective convolutional layer is used to reduce the dimensionality of high-dimensional features to the number of output channels. The feature map sizes corresponding to the last four inverted residual blocks are 28×28, 14×14, 14×14, and 7×7, respectively.

[0047] S4. Input the preprocessed finger vein image from step S1 into the improved MobileNetV2 network for feature extraction and classification, and output the recognition results.

[0048] Example 2:

[0049] This embodiment provides a finger vein recognition system based on an improved MobileNetV2 to implement the method in Embodiment 1. The system includes:

[0050] Image acquisition and preprocessing module: used to acquire finger vein images and preprocess them into images to be recognized with a size of 224×224×3;

[0051] Network building module: used to build an improved MobileNetV2 network, which embeds the SE attention mechanism in the last four inverted residual blocks of MobileNetV2;

[0052] Network training module: used to train the network on a public or self-built finger vein dataset, and update the network parameters using the Adam optimizer and cross-entropy loss function;

[0053] Feature extraction and recognition module: This module takes the preprocessed image to be recognized and inputs it into the trained network. After forward propagation, it outputs the identity recognition result.

[0054] Combining Examples 1 and 2, this paper addresses the issues encountered by MobileNetV2 in finger vein recognition tasks, such as insufficient learning of key features and limited generalization ability with small sample data. The improvement focuses on embedding an attention mechanism. Firstly, the inverted residual block of the MobileNetV2 network is improved, as shown below... Figure 1 The SE attention mechanism embedding shown yields the improved inverted residual blocks (SE-Inverted Residuals), with the structure as follows: Figure 2 As shown.

[0055] The improved SE-Inverted Residuals blocks are integrated into the existing MobileNetV2 network. Inverted residual blocks with SE modules are only used when there are 4 stages. The improved network is abbreviated as PSE-MobileNetV2 (PartialSE-MobileNetV2). The overall structure parameters are shown in Table 1. The input image size is 224×224×3, and the width multiplier (width_mult) is set to 1.0 to preserve sufficient feature channels. The network consists of an initial convolutional layer, 7 stages of inverted residual blocks (the first 3 stages are Inverted Residuals blocks, and the last 4 stages are SE-Inverted Residuals blocks), a final convolutional layer, and a classifier. This paper chooses to embed SE attention into the last 4 inverted residual blocks of MobileNetV2. The core theoretical basis is as follows: shallow layers focus on basic feature extraction, requiring no additional attention; deep layers have clear semantics and less redundancy, and attention intervention can accurately enhance effective features while controlling computational overhead, achieving an optimal balance between accuracy and efficiency.

[0056] Table 1 Overall structural parameters of PSE-MobileNetV2

[0057]

[0058] Note: N is the number of classification categories (SDUMLA dataset N=636, FV-USM dataset N=492).

[0059] Example 3: Experiment and Analysis

[0060] 3.1 Experimental Environment and Dataset

[0061] The experiment was based on the PyTorch framework and ran on the PyCharm 2024 platform. The system used was Windows 11, the GPU was NVIDIA GeForce 4060ti 16GB, and the running memory was 16GB. The experiment used two internationally available finger vein datasets, namely the Universiti Teknologi Malaysia dataset (FVUSM)

[14] and the Shandong University dataset (SDUMLA)

[15] , covering different populations, collection devices and environments to ensure the generalization ability of the model was verified.

[0062] FVUSM collected data from 123 volunteers (aged 20-35, male-to-female ratio 2.075:1). Four fingers (index and middle fingers) from each volunteer's hands were collected, with each finger captured 12 times (in two phases), totaling 5904 images. The training, validation, and test sets were in a ratio of 8:1:1. Some finger vein images from the dataset are shown below. Figure 3 As shown.

[0063] The SDUMLA dataset consisted of 106 volunteers (aged 18-40, male-to-female ratio 1:1). Images were collected from six fingers (index, middle, and ring fingers) of each volunteer's hands, six times per finger, for a total of 3816 images. The training, validation, and test sets were arranged in a 4:1:1 ratio. Some finger vein images from this dataset are shown below. Figure 4 As shown.

[0064] 3.2 Experimental Design

[0065] During training, the Adam optimizer is used to update the model parameters. The initial learning rate is set to 0.001, a value commonly used in lightweight network finger vein recognition tasks, which can achieve fast convergence in the early stages of training while avoiding gradient oscillations. The weight decay is set to 0.0001, which, by applying L2 regularization constraints to the model parameters, suppresses excessive parameter growth and alleviates overfitting problems on small datasets. The ReduceLROnPlateau learning rate scheduling strategy is adopted to avoid model oscillations caused by excessively high learning rates in the later stages of training. The mode is set to monitoring mode='max', using the validation set accuracy (Val Acc) as the monitoring metric. Learning rate adjustment is triggered only when this metric stops improving. The decay factor is set to 0.5, which multiplies the current learning rate by 0.5 (i.e., halves the learning rate) when decay is triggered. The patience value is set to 10: if the validation set accuracy does not improve within 10 consecutive training epochs, learning rate decay is triggered to avoid misjudgments due to short-term fluctuations. The cross-entropy loss function is used.

[0066] 3.3 Results Analysis

[0067] 3.3.1 Comparison of accuracy of each model on the test set

[0068] This embodiment trains and compares PSE-MobileNetV2 with other mainstream convolutional neural networks on FVUSM and SDUMLA, with a batch size of 64 and epochs of 100 for each network. Tables 2 and 3 show the accuracy, precision, recall, and F1 score of each model on the two public test datasets. It is evident that PSE-MobileNetV2 significantly outperforms the other comparison models. Since some benchmark networks (such as ResNet50) can already effectively extract features, the introduction of the SE attention mechanism creates structural redundancy and interferes with the original feature representation. This situation also occurs in networks such as ShuffleNetV3 and MobileNetV3-Small, indicating that blindly stacking attention structures can limit model performance. Compared to FVUSM, the accuracy of all models on the SDUMLA dataset decreased because this dataset contains some low-quality images. These results can provide a reference for further model optimization and also provide strong evidence for the practical application of the models.

[0069] Table 2 Comparison of recognition performance of different models (FVUSM)

[0070]

[0071] Table 3 Comparison of recognition performance of different models (SDUMLA)

[0072]

[0073] Figure 5 The accuracy and loss curves of various models trained on two public dataset validation sets are shown.

[0074] 3.3.2 Comparison Table of Complexity and Inference Speed ​​of Various Models

[0075] To comprehensively evaluate the lightweight nature and inference efficiency of each model, this section compares and analyzes mainstream lightweight models with traditional deep networks from five dimensions: number of parameters (M), computational cost (FLOPs), model storage size (MB), single-image inference latency (ms), and frame rate (FPS).

[0076] Table 4. Complexity and inference speed of different models

[0077]

[0078] The data in the table shows that ResNet50 and ResNet50SE, as traditional deep networks, have more than 24 M parameters, resulting in significantly higher computational costs than lightweight models. Their inference latency is close to 10 ms, placing high demands on hardware resources. ShuffleNetV3, with its lightweight structure and neural architecture search optimization, has the fewest parameters and the lowest computational cost among all models, occupying only 2.99 MB of storage. It also achieves the lowest inference latency of 1.36 ms and the highest frame rate of 735.3 FPS, but its recognition accuracy is lower on datasets with low image quality. MobileNetV2 and its improved versions, MobileNetV2SE and PSE-MobileNetV2, are at a moderate level in terms of parameter and computational costs, with all three having 3.04 M to 3.60 M parameters and similar computational costs. Their inference speed remains at 330 to 340 FPS, achieving a good balance between accuracy and efficiency. Although MobileNetV3 significantly reduces the number of parameters and computational costs, its recognition performance decreases noticeably. Overall, lightweight networks have significant advantages in resource-constrained scenarios. The PSE-MobileNetV2 proposed in this paper has a slightly increased number of parameters and computational cost after the introduction of the attention mechanism, but it can effectively improve the finger vein feature expression capability and recognition accuracy.

[0079] 3.4 Ablation Experiment

[0080] To verify the impact of different embedding strategies of the SE attention module on model performance, this paper uses MobileNetV2 as the baseline network, designs three sets of ablation experiments, and evaluates the accuracy on the FVUSM and SDUMLA datasets respectively. The results are shown in Table 5.

[0081] Table 5 Experimental results of different embedding methods of the SE module

[0082]

[0083] Globally embedding SE modules only achieves a slight improvement over the baseline model on the FVUSM dataset, but shows a performance decline on the SDUMLA dataset, indicating that global SE embedding is prone to introducing redundant parameters and overfitting risks. The partial SE embedding strategy proposed in this paper achieves state-of-the-art performance on both datasets, improving upon the baseline model by 0.31 and 0.84 percentage points respectively. This demonstrates that selectively embedding SE modules at key levels can more efficiently enhance feature channel attention, improving model representation ability and generalization performance.

[0084] 3.5 Interference Simulation Experiment

[0085] To verify the robustness of the proposed method in real-world data acquisition scenarios, this section constructs three typical types of interference: noise, illumination, and occlusion, and quantitatively evaluates the model's recognition performance under different levels of interference. In the experimental setup, the training set contains 50% interfered images, while the validation and test sets use clean, interference-free samples to meet the requirements of uncontrollable interference and stable, reliable recognition in real-world applications.

[0086] 3.5.1 Noise Simulation

[0087] During the acquisition process, finger vein images are susceptible to Gaussian noise and salt-and-pepper noise introduced by the sensor and the environment, resulting in degradation of texture details and blurring of features. To test the model's noise resistance, different intensities of Gaussian noise (σ=15, 25, 35) and salt-and-pepper noise (P=0.04, 0.06, 0.08) were applied, and the recognition accuracy on SDUMLA and FVUSM was compared. The results are shown in Table 6.

[0088] Table 6 Noise simulation experiment results

[0089]

[0090] Under Gaussian noise interference, the accuracy of FVUSM remained at 98.58%–99.37%, while that of SDUMLA was 96.28%–97.32%. Under salt-and-pepper noise interference, the accuracy of FVUSM was 98.27%–99.21%, and that of SDUMLA was 96.69%–96.85%. With increasing noise intensity, the performance of both methods generally decreased slightly, but under specific interference intensities, the accuracy improved slightly, indicating that training on appropriate interference datasets can improve recognition performance. Experimental results show that the attention mechanism can effectively suppress texture distortion caused by noise and improve the stability of feature extraction and recognition under noise contamination.

[0091] 3.5.2 Lighting Simulation

[0092] Uneven distribution of near-infrared light sources and exposure deviations can easily lead to underexposure, overexposure, and local shadows in finger vein images, significantly affecting the visibility of vein texture. Gamma correction was used to simulate different lighting conditions, with γ set to 0.5 (underexposure), 0.7 (dark), 1.5 (bright), and 2.0 (overexposure) to test the model's adaptability to lighting distortion. The experimental results are shown in Table 7.

[0093] Table 7 Results of Illumination Simulation Experiment

[0094]

[0095] Under a wide range of lighting conditions, FVUSM accuracy remained stable at 98.58%–99.52%, with fluctuations of less than 1%; SDUMLA accuracy ranged from 96.13% to 97.16%. Both methods exhibited good lighting robustness, with FVUSM reaching a peak accuracy of 99.52% at γ=1.5, maintaining high recognition accuracy even under underexposure and overexposure conditions, making it suitable for application environments with unstable lighting in real-world data acquisition scenarios.

[0096] 3.5.3 Occlusion Simulation

[0097] Factors such as finger misalignment, surface dirt, and skin folds often cause local occlusion of the vein area, leading to the loss of effective features. To simulate this type of interference, rectangular occlusion blocks with an area of ​​5% to 25% were randomly generated in the image to cover the key vein area. The model's recognition ability under local information loss was evaluated, and the results are shown in Table 8.

[0098] Table 8 Results of occlusion simulation experiment

[0099]

[0100] As the occlusion ratio increases, the accuracy of both FVUSM and SDUMLA shows a steady downward trend. At an occlusion intensity of 5%, FVUSM reaches 99.52%, and SDUMLA reaches 97.16%. When the occlusion intensity increases to 25%, FVUSM remains at 98.31%, and SDUMLA at 95.91%. This indicates that the proposed method has stronger feature compensation and discrimination capabilities for local occlusion and can maintain stable recognition performance even when key information is missing.

[0101] In summary, the PSE-MobileNetV2 finger vein recognition model proposed in the above embodiments provides a new approach to finger vein recognition. Compared with the original MobileNetV2 network, PSE-MobileNetV2 solves the problems of insufficient focus on key vein features and easy overfitting in the original network. PSE-MobileNetV2 embeds the SE attention module in the semantically rich Stage of the later stage of the network, and strengthens the expression of discriminative features such as vein branches and intersections through adaptive calibration of channel weights. It further optimizes the attention module and feature fusion logic, improving robustness in complex scenarios. The experimental results show that the recognition accuracy of the improved network on the FVUSM and SDUMLA datasets is 99.49% and 97.06%, respectively, which is an improvement compared with the basic network.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A finger vein recognition method based on an improved MobileNetV2, characterized in that, Includes the following steps: S1. Obtain a finger vein image, preprocess it, and use it as the input image; S2. Construct an improved MobileNetV2 network, including an initial convolutional layer, several inverted residual block stages, a final convolutional layer, and a classifier; S3. In the improved MobileNetV2 network, the SE attention mechanism is embedded in the last four inverted residual blocks to form SE inverted residual blocks, while the first three inverted residual block stages remain as standard inverted residual blocks; S4. Input the preprocessed finger vein image from step S1 into the improved MobileNetV2 network for feature extraction and classification, and output the recognition results.

2. The finger vein recognition method based on the improved MobileNetV2 according to claim 1, characterized in that, The size of the input image in step S1 is 224×224×3.

3. The finger vein recognition method based on the improved MobileNetV2 according to claim 1, characterized in that, In step S2, the width multiplier of the improved MobileNetV2 network is set to 1.0; the kernel size of the initial convolutional layer is 3×3, the stride is 2, and the number of output channels is 32; the classifier is a fully connected layer, and the number of output channels is equal to the number of finger vein identity categories N.

4. The finger vein recognition method based on the improved MobileNetV2 according to claim 1, characterized in that, In step S3, the last four SE inverted residual blocks include an extended convolutional layer, a depthwise separable convolutional layer, an SE attention layer, and a projective convolutional layer connected in sequence; the extended convolutional layer is used to increase the channel dimension; the depthwise separable convolutional layer is used to extract spatial features; The SE attention layer is used to adaptively calibrate channel weights; the projective convolutional layer is used to reduce the dimensionality of high-dimensional features to the number of output channels.

5. A finger vein recognition method based on an improved MobileNetV2 according to claim 4, characterized in that, In step S3, the feature map sizes corresponding to the last four inverted residual blocks are 28×28, 14×14, 14×14 and 7×7, respectively.

6. The finger vein recognition method based on the improved MobileNetV2 according to claim 1, characterized in that, Step S4 specifically involves inputting the preprocessed finger vein image from step S1 into the improved MobileNetV2 network. The final convolutional layer maps the high-dimensional feature map to the classification feature space, and the classifier outputs the probability distribution of each identity category. The category corresponding to the maximum value is taken as the recognition result.

7. A finger vein recognition system based on an improved MobileNetV2, for implementing the method as described in any one of claims 1 to 6, characterized in that, The system includes: Image acquisition and preprocessing module: used to acquire finger vein images and preprocess them into images to be recognized with a size of 224×224×3; Network building module: used to build an improved MobileNetV2 network, which embeds the SE attention mechanism in the last four inverted residual blocks of MobileNetV2; Network training module: used to train the network on a public or self-built finger vein dataset, and update the network parameters using the Adam optimizer and cross-entropy loss function; Feature extraction and recognition module: This module takes the preprocessed image to be recognized and inputs it into the trained network. After forward propagation, it outputs the identity recognition result.

Images