A lightweight finger vein recognition method and system fusing multi-scale vein texture enhancement and coordinate attention

Abstract

The application discloses a kind of fusion multi-scale vein texture enhancement and coordinate attention light weight finger vein identification method and system, through embedding multi-scale vein texture enhancement module in MobileNetV3 style light weight main stem, and introducing coordinate attention mechanism, while keeping lower parameter amount and computational complexity, the expression ability of model to fine weak vein texture, multi-scale blood vessel structure and spatial direction information is improved, the generalization ability under small sample condition is improved by training strategy of original drawing and enhanced copy combination, can alleviate the influence of collection posture, illumination change and noise interference on identification performance.

Description

Technical Field

[0001] This invention belongs to the fields of biometric recognition, computer vision and deep learning technology, and in particular relates to a lightweight finger vein recognition method and system that integrates multi-scale vein texture enhancement and coordinate attention. Background Technology

[0002] With the development of applications such as mobile payment, access control, identity verification, and smart terminal security authentication, biometric identification technology has received widespread attention. Compared with external biometric features such as faces and fingerprints, finger vein features are located inside the human finger and usually require near-infrared imaging equipment for collection. They have advantages such as strong liveness detection, difficulty in forgery, non-contact or weak-contact collection, and high security.

[0003] Despite the promising application prospects of finger vein recognition, several challenges remain in practical applications. First, finger vein images are typically acquired using near-infrared devices, resulting in low overall contrast, weak vein texture, and susceptibility to noise interference. Second, the thickness, bifurcation, and direction of veins vary significantly among different fingers, requiring models with multi-scale texture extraction capabilities. Third, variations in finger placement, acquisition equipment, lighting conditions, and image preprocessing methods can also affect recognition performance. Finally, in mobile or embedded authentication scenarios, models not only need high recognition accuracy but also require small parameter counts, low computational cost, and fast inference speeds.

[0004] While existing deep convolutional neural networks have achieved some success in finger vein recognition tasks, traditional deep networks typically have a large number of parameters, making them unsuitable for deployment on resource-constrained devices. The MobileNet series of networks effectively reduces model complexity through depthwise separable convolutions and inverse residual structures, making them suitable for lightweight visual recognition tasks. MobileNetV3 further incorporates the H-Swish activation function, lightweight attention mechanisms, and a more efficient network structure based on MobileNetV2, achieving a better balance between accuracy and speed.

[0005] However, directly applying existing lightweight networks to finger vein recognition still has shortcomings: First, weak vein textures are easily weakened by shallow convolutions under low contrast conditions; second, a single convolution scale is insufficient to simultaneously model small blood vessels, medium-scale blood vessels, and the course of blood vessels over long distances; third, traditional SE attention mainly models channel relationships, failing to fully utilize the spatial location and orientation information of blood vessel textures; and fourth, small sample finger vein datasets are prone to model overfitting, affecting the stability of practical deployment. Therefore, a finger vein recognition method is needed that can enhance multi-scale vein texture representation and orientation awareness while maintaining lightweight design. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a lightweight finger vein recognition method and system that integrates multi-scale vein texture enhancement and coordinate attention. By embedding a multi-scale vein texture enhancement module into the MobileNetV3-style lightweight backbone and introducing a coordinate attention mechanism, the model's ability to express weak vein textures, multi-scale vascular structures and spatial orientation information is improved while maintaining a low number of parameters and computational complexity.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] 1. A lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention, characterized by comprising the following steps:

[0009] S1. Obtain a near-infrared finger vein image of the finger to be identified, and perform grayscale conversion, size normalization, and pixel normalization on the finger vein image to obtain an input image;

[0010] S2. Construct a lightweight finger vein recognition network. Input the input image obtained in step S1. The lightweight finger vein recognition network includes a Stem feature extraction layer, a MobileNetV3-style inverted residual backbone network, a multi-scale vein texture enhancement module, a coordinate attention module, a global pooling layer, an embedding layer, and a classifier.

[0011] S3. Multi-scale vein texture enhancement modules are embedded in the middle and high-level feature positions of the inverted residual backbone network, respectively, and fine-grained vein texture, medium-scale blood vessel texture and long-distance blood vessel orientation features are extracted through deep convolution branches with multiple different receptive fields.

[0012] S4. The features are aggregated along the height and width directions by the coordinate attention module to generate attention weights with directional position information, and the fused features are recalibrated using the attention weights.

[0013] S5. Input the recalibrated features into the global pooling layer, embedding layer and classifier, output the probability distribution of each identity category, and take the category with the highest probability as the finger vein recognition result.

[0014] Furthermore, the input image is a single-channel grayscale image with a size of 224×224; during the training phase, the training image is enhanced using at least one of the following methods: random ROI jitter, rotation, translation, scaling, brightness and contrast variation, Gamma variation, and slight noise perturbation.

[0015] Furthermore, the Stem feature extraction layer includes a 3×3 convolutional layer, a batch normalization layer, and an H-Swish activation function; the inverse residual backbone network is composed of depthwise separable convolution, linear projection, and residual connections, with a width factor of 0.50, 0.75, or 1.00.

[0016] Furthermore, the multi-scale vein texture enhancement module includes parallel 3×3 depth convolution branches, 5×5 depth convolution branches, and dilated 3×3 depth convolution branches. The outputs of each branch are spliced ​​by channels and fused by 1×1 convolution, and then output through residual connection.

[0017] Furthermore, the coordinate attention module performs feature aggregation along the height and width directions respectively, and after a shared 1×1 convolution transformation, it splits to generate height attention weights and width attention weights, which are then applied to the input features in the form of element-wise multiplication.

[0018] The present invention also provides a lightweight finger vein recognition system that integrates multi-scale vein texture enhancement and coordinate attention to implement the above method. The system includes an image acquisition and preprocessing module, a network construction module, a network training module, and a feature extraction and recognition module.

[0019] Compared with the prior art, the present invention has the following beneficial technical effects:

[0020] 1. This invention embeds a multi-scale vein texture enhancement module into a lightweight MobileNetV3 style backbone, which can simultaneously capture fine-grained local texture, medium-scale vascular texture and long-distance vascular direction information, thereby improving the discriminative feature expression ability in low-contrast finger vein images.

[0021] 2. This invention introduces a coordinate attention mechanism, which retains the positional encoding in the height and width directions in channel attention, enabling the network to better focus on blood vessel orientation, bifurcation locations, and local continuity, thus making up for the problem of insufficient utilization of spatial positional information in traditional SE attention.

[0022] 3. This invention improves recognition performance while maintaining a low number of parameters and computational complexity, making it suitable for deployment in resource-constrained scenarios such as mobile devices, embedded access control systems, and identity authentication terminals.

[0023] 4. This invention improves the generalization ability under small sample conditions by combining the original image with the enhanced copy training strategy, which can alleviate the impact of acquisition pose, illumination changes and noise interference on recognition performance. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall network structure of MSVTE-CA-MobileNetV3 in an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the multi-scale vein texture enhancement module structure in an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the coordinate attention module structure in an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of the training curve of the model on the SDUMLA dataset in an embodiment of the present invention;

[0028] Figure 5 This is a schematic diagram of the training curve of the model on the FVUSM dataset in an embodiment of the present invention. Detailed Implementation

[0029] 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.

[0030] Example 1:

[0031] This embodiment provides a lightweight finger vein recognition method that integrates multi-scale vein texture enhancement and coordinate attention, such as... Figure 1 As shown, it includes the following steps:

[0032] S1. Obtain the near-infrared finger vein image of the finger to be identified, convert it into a single-channel grayscale image, scale it to 224×224, and then perform pixel normalization to obtain the finger vein image to be input into the network.

[0033] S2. Construct a lightweight finger vein recognition network, abbreviated as MSVTE-CA-MobileNetV3. The network uses a MobileNetV3-Small style inverted residual structure as its backbone. A Stem layer is placed at the network input, consisting of 3×3 convolutions, batch normalization, and H-Swish activation, used for initial downsampling and low-level edge texture extraction. The backbone network reduces computational cost through depthwise separable convolutions and maintains lightweight feature representation through linear projection and residual connections.

[0034] S3. Insert two multi-scale vein texture enhancement modules at the mid-to-high-level feature locations of the backbone network. For example... Figure 2 The first multi-scale vein texture enhancement module works on higher resolution mid-level features to enhance weak vein edges, local textures, and short-distance bifurcation structures; the second multi-scale vein texture enhancement module works on deeper semantic features to capture thicker blood vessels and long-distance orientation information.

[0035] S4. The multi-scale vein texture enhancement module comprises three parallel branches: a 3×3 depth convolution branch, a 5×5 depth convolution branch, and a dilated 3×3 depth convolution branch. The 3×3 depth convolution branch extracts fine-grained local textures, the 5×5 depth convolution branch covers mid-scale vascular regions, and the dilated 3×3 depth convolution branch expands the receptive field without significantly increasing the number of parameters, enabling the perception of longer-distance vascular pathways. The three types of features are concatenated via channels and then compressed and fused using a 1×1 convolution. The output is then added to the module input via residual connections.

[0036] S5. Introduce a coordinate attention module in the network Head section. For example... Figure 3 The coordinate attention module aggregates features along both the height and width directions, concatenates the features from both directions, transforms them using a shared 1×1 convolution, and then splits them into height and width attention weights. Finally, the input features are recalibrated. Compared to channel attention that only uses global average pooling, coordinate attention preserves directional encoding when generating channel weights, making it more suitable for processing long, continuous, and directional vascular textures in finger vein images.

[0037] S6. Input the recalibrated high-level features into the global average pooling layer and the embedding layer to obtain the feature vector for identity classification. Then, output the probability distribution of each identity category through the classifier and take the category corresponding to the highest probability as the finger vein recognition result.

[0038] Example 2:

[0039] This embodiment provides a lightweight finger vein recognition system that integrates multi-scale vein texture enhancement and coordinate attention to implement the method in Embodiment 1. The system includes:

[0040] Image acquisition and preprocessing module: used to acquire finger vein images and perform grayscale conversion, size normalization, pixel normalization, and data augmentation processing on the finger vein images during the training phase;

[0041] Network building module: used to build the MSVTE-CA-MobileNetV3 lightweight finger vein recognition network. The network includes a Stem layer, a MobileNetV3 style inverse residual backbone, a multi-scale vein texture enhancement module, a coordinate attention module, a global pooling layer, an embedding layer, and a classifier.

[0042] Network training module: used to train the lightweight finger vein recognition network on a public or self-built finger vein dataset. During training, the cross-entropy loss function or the labeled smooth cross-entropy loss function is used. The optimizer is AdamW, with an initial learning rate of 0.001 and a weight decay of 0.0001. A cosine annealing learning rate scheduling strategy can also be used.

[0043] Feature extraction and recognition module: This module takes the input image to be recognized and inputs it into a pre-trained lightweight finger vein recognition network. The network then performs multi-scale texture enhancement, coordinate attention recalibration, and a classifier to obtain the identity recognition result.

[0044] Example 3: Experiment and Analysis

[0045] To verify the effectiveness of the method of this invention, experiments were conducted on two publicly available finger vein datasets, SDUMLA and FVUSM. The experiments employed a closed-set identity classification approach, meaning the training, validation, and test sets shared the same set of identity categories. The input image size was 224×224, the batch size was 64, the number of training epochs was 100, the default width factor was 0.75, and the embedding dimension was 576.

[0046] For the SDUMLA dataset, due to the small number of samples per class, this embodiment adopts a "original image + augmented copy" training strategy. That is, the original images are retained in the training set, while additional randomly augmented images are added. Augmentation methods include slight ROI jitter, rotation, translation, scaling, brightness / contrast changes, gamma changes, and slight noise. Random augmentation is not used on the validation and test sets. For the FVUSM dataset, data augmentation is not used by default to maintain stable data distribution. Specific experimental data and comparisons are shown in Tables 1 to 5, and appendices are also provided. Figure 4 , 5 .

[0047] Table 1 Dataset Partitioning Methods

[0048]

[0049] Table 2 Ablation Experiment Results on SDUMLA Dataset

[0050]

[0051] Table 3 Experimental results of the model under different enhancement strategies

[0052]

[0053] Table 4. Experimental results of the model under different width factors.

[0054]

[0055] Table 5 Experimental results for the FVUSM dataset

[0056]

[0057] It should be noted that standard MobileNetV3 typically integrates the SE module in some inverted residual blocks. To avoid conceptual confusion and simplify the description, in Tables 2 and 5, the MobileNetV3 style backbone without SE / CA is denoted as "Baseline"; "+SE", "+CA", and "+MSVTE" respectively indicate the addition of the corresponding module to the baseline; "+MSVTE+CA" indicates the simultaneous introduction of a multi-scale vein texture enhancement module and a coordinate attention mechanism.

[0058] The experimental results show that the MSVTE-CA-MobileNetV3 achieves a test accuracy of 98.27% on the SDUMLA dataset using the "original image + 2 enhanced copies" strategy, and 98.58% on the dataset with a width factor of 1.0; on the FVUSM dataset, it achieves a test accuracy of 99.59% and an EER of 0.51%. These results demonstrate that the multi-scale vein texture enhancement module and the coordinate attention module can improve the finger vein feature discrimination ability with a relatively small increase in complexity.

[0059] In summary, the lightweight finger vein recognition method and system proposed in this invention, which integrates multi-scale vein texture enhancement and coordinate attention, can effectively solve problems such as difficulty in extracting weak textures in low-contrast finger vein images, significant changes in blood vessel scale, insufficient utilization of spatial orientation information, and limitations in lightweight deployment. It improves recognition accuracy while maintaining low parameter quantity and computational complexity, and is suitable for scenarios such as access control, mobile terminals, and embedded identity authentication.

[0060] 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 lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention, characterized in that, Includes the following steps: S1. Obtain a near-infrared finger vein image of the finger to be identified, and perform grayscale conversion, size normalization, and pixel normalization on the near-infrared finger vein image to obtain an input image; S2. Construct a lightweight finger vein recognition network and input the input image obtained in step S1. The lightweight finger vein recognition network includes a Stem feature extraction layer, an inverted residual backbone network, a multi-scale vein texture enhancement module, a coordinate attention module, a global pooling layer, an embedding layer, and a classifier. S3. Multi-scale vein texture enhancement modules are embedded in the middle and high-level feature positions of the inverted residual backbone network, respectively. Features are extracted through deep convolution branches with multiple different receptive fields, including: fine-grained vein texture, medium-scale vascular texture and long-distance vascular orientation. S4. The features are aggregated along the height and width directions by the coordinate attention module to generate attention weights with directional position information, and the fused features are recalibrated using the attention weights. S5. Input the recalibrated features into the global pooling layer, embedding layer and classifier, output the probability distribution of each identity category, and take the category with the highest probability as the finger vein recognition result.

2. The lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention as described in claim 1, characterized in that, In step S1, the input image is a single-channel grayscale image with a size of 224×224. During the training phase, the training image is enhanced using at least one of the following methods: random ROI jitter, rotation, translation, scaling, brightness and contrast variation, Gamma variation, and slight noise perturbation.

3. The lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention as described in claim 1, characterized in that, In step S2, the Stem feature extraction layer includes a 3×3 convolutional layer, a batch normalization layer, and an H-Swish activation function; the inverse residual backbone network is composed of depthwise separable convolution, linear projection, and residual connections, with a width factor of 0.50, 0.75, or 1.

00.

4. A lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention as described in claim 1, characterized in that, The multi-scale vein texture enhancement module includes parallel 3×3 depth convolution branches, 5×5 depth convolution branches, and dilated 3×3 depth convolution branches; the 3×3 depth convolution branches are used to extract fine-grained local textures, the 5×5 depth convolution branches are used to extract mid-scale vascular region features, and the dilated 3×3 depth convolution branches are used to expand the receptive field and extract long-distance vascular orientation information.

5. A lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention as described in claim 4, characterized in that, The multi-scale vein texture enhancement module concatenates the output features of each branch, then compresses and fuses them through 1×1 convolution to restore the number of channels, and adds the fused features to the module input features through residual connection for output.

6. A lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention as described in claim 1, characterized in that, The coordinate attention module includes a height-direction pooling unit, a width-direction pooling unit, a shared 1×1 convolutional transformation unit, an attention splitting unit, and a feature recalibration unit; the height-direction pooling unit and the width-direction pooling unit retain the horizontal position information and the vertical position information, respectively.

7. A lightweight finger vein recognition method integrating multi-scale vein texture enhancement and coordinate attention as described in claim 1, characterized in that, The lightweight finger vein recognition network is trained for closed-set identity classification using the cross-entropy loss function or the labeled smooth cross-entropy loss function. The optimizer is AdamW with an initial learning rate of 0.001 and a weight decay of 0.0001.

8. A lightweight finger vein recognition system integrating multi-scale vein texture enhancement and coordinate attention, for implementing the method as described in any one of claims 1 to 7, characterized in that, The system includes: Image acquisition and preprocessing module: used to acquire finger vein images and process them into input images to be recognized; Network building module: Used to build a lightweight finger vein recognition network including an inverted residual backbone network, a multi-scale vein texture enhancement module, and a coordinate attention module; Network training module: used to train the lightweight finger vein recognition network based on public or self-built finger vein datasets, and to improve the generalization ability under small sample conditions through data augmentation strategies; Feature extraction and recognition module: This module takes the input image to be recognized and inputs it into a pre-trained lightweight finger vein recognition network. After multi-scale texture enhancement, coordinate attention recalibration, and a classifier, it outputs the identity recognition result.

Images