Wrist vein verification method with lightweight attention mechanism deep learning model
By using a lightweight attention mechanism deep learning model to extract wrist vein features from visible light images, the accuracy and efficiency problems of existing wrist vein verification methods are solved, achieving low-cost and secure identity verification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2026-03-10
AI Technical Summary
Existing wrist vein verification methods are difficult to improve in terms of accuracy and efficiency, and existing devices are expensive and may pose health risks.
A lightweight attention-based deep learning model is used to extract wrist vein features from visible light images, perform region of interest detection and vascular image enhancement, and combine the lightweight attention-based deep learning model for feature extraction and comparison, avoiding the use of near-infrared sensors.
It reduces device costs, improves verification efficiency and accuracy, reduces health hazards, and is suitable for rapid identity verification using handheld devices such as smartphones.
Smart Images
Figure CN121640529A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a wrist vein verification method with a lightweight attention mechanism deep learning model, and particularly to a wrist vein verification method using visible light images. Background Technology
[0002] Biometric technology is the preferred technology for identity recognition or authentication today. Among various biometric technologies, veins have reliable identification characteristics and are unique to each individual. There have been relevant studies on the identification of hand veins such as palm veins, dorsal veins, and finger veins. However, among these detection locations, wrist veins provide highly accurate identification results because they are closer to the skin than other organs. Their unique visual vein feature pattern makes them the primary identification target when identity recognition is required for various electronic device logins, access control, transportation systems, and other applications.
[0003] Existing devices for wrist vein identification primarily utilize fixed near-infrared sensors to detect images of wrist veins. While near-infrared light is effective at acquiring images of wrist veins, the cost of devices with near-infrared sensors is too high, hindering widespread adoption. Furthermore, prolonged exposure to near-infrared radiation raises concerns about user health. Therefore, readily available smartphones and similar handheld devices have become a focus of research. If identification can be performed via handheld devices, their accessibility and ease of use would make them an important biometric identification method. However, handheld devices can only capture wrist images using visible light image acquisition, which presents challenges in image acquisition and processing. Additionally, the limited processing capabilities of handheld devices make it difficult to execute complex computational models, resulting in insufficient accuracy for verification.
[0004] In view of this, existing biometric technologies still have certain shortcomings in wrist vein verification devices and methods, making it difficult to complete the identification process accurately and quickly. To address this, the inventors of this invention have conceived and designed a wrist vein verification method with a lightweight attention mechanism deep learning model, improving upon the deficiencies of existing technologies and thereby enhancing its industrial application. Summary of the Invention
[0005] To address the aforementioned problems, the present invention aims to provide a wrist vein verification method with a lightweight attention mechanism deep learning model, thereby solving the problem that existing wrist vein verification methods are difficult to improve in terms of accuracy and verification efficiency.
[0006] According to one objective of the present invention, a wrist vein verification method with a lightweight attention mechanism deep learning model is proposed, comprising the following steps: acquiring a visible light image of the wrist vein of the subject through an image capturing device and storing it in an image processing device; performing a region of interest (ROI) detection procedure on the visible light image of the wrist vein using the image processing device to extract the ROI image, and performing a vascular image enhancement procedure on the ROI image to obtain the model input image; executing a lightweight attention mechanism deep learning model using the image processing device and inputting the model input image to extract feature vectors; comparing the feature vectors with the registered wrist feature vectors of the subject using the image processing device; if the comparison result matches, the verification is considered successful; if the comparison result does not match, the verification is considered unsuccessful.
[0007] Preferably, the visible light image of the wrist vein can be preprocessed to convert the red, green, and blue values (R, G, B values) of the hue attribute in the image space (HSL color space) into saturation values of the saturation attribute.
[0008] Preferably, the wrist region of interest detection procedure may include cropping and scaling the visible light image of the wrist vein to a predetermined size, segmenting the predetermined size image to generate a wrist region image, finding the maximum contour of the wrist region image, finding the convex hull and convexity defects of the maximum contour, determining the key vector by the convexity defects, locating the region of interest by the key vector, and then extracting the wrist region of interest image.
[0009] Preferably, the vascular image enhancement procedure can perform local ridge enhancement (LRE) on the region of interest image of the wrist, including automatic gamma correction (AGC) on the region of interest image of the wrist, limiting contrast-limited adaptive histogram equalization (CLAHE) to increase image contrast, eliminating noise with a low-pass Gaussian filter, and highlighting edge details with a high-pass Laplacian operator to obtain the model input image.
[0010] Preferably, the lightweight attention mechanism deep learning model may include a convolutional module, four sandglass blocks, a mixed pooling coordinate attention block (MPCA block), and a parallel spatial pyramid pooling block (PSPP block).
[0011] Preferably, the convolution module may include a standard convolutional layer, a batch normalization layer, and a non-linear activation function layer.
[0012] Preferably, the four hourglass modules can be improved hourglass modules with a bottleneck structure, including two pointwise convolutional layers and two depthwise separable convolutional layers. The two pointwise convolutional layers are used for channel reduction and channel expansion, respectively. The bottleneck structure is located between the two pointwise convolutional layers. The two depthwise separable convolutional layers are respectively set at the input and output ends of the improved hourglass module, and a linear separable convolutional module is set to connect the two depthwise separable convolutional layers.
[0013] Preferably, a discard layer can be set between the four hourglass modules and the hybrid pooling coordinate attention module.
[0014] Preferably, the hybrid pooling coordinate attention module can perform global average pooling and max pooling along the horizontal and vertical directions respectively, and then combine them by adding operators before outputting.
[0015] Preferably, the parallel spatial pyramid pooling module can be configured with three parallel spatial pyramid pooling blocks, each with a different size of partitioned region (bins). The feature maps generated by the operation of the three parallel spatial pyramid pooling blocks are connected and flattened to generate an output feature map, from which feature vectors are extracted.
[0016] Preferably, the method for comparing the feature vector with the registered wrist feature vector is to calculate the cosine distance between the feature vector and the registered wrist feature vector. If the cosine distance is greater than the distance threshold, it is considered that the result is inconsistent and the verification is judged as failed. If the cosine distance is not greater than the distance threshold, it is considered that the result is consistent and the verification is judged as successful.
[0017] In summary, the wrist vein verification method of the present invention with a lightweight attention mechanism deep learning model has one or more of the following advantages:
[0018] (1) The wrist vein verification method of the present invention with a lightweight attention mechanism deep learning model can identify visible light images of wrist veins captured by a smartphone. It does not require the installation of a near-infrared sensing device, which reduces the cost of the device and increases the popularity of the verification device. It also avoids the health hazards to users caused by irradiating near-infrared light.
[0019] (2) The wrist vein verification method of the present invention with a lightweight attention mechanism deep learning model can improve verification efficiency by reducing the amount of system computation through the lightweight attention mechanism deep learning model, and improve the accuracy of identification by extracting more features through pyramid pooling blocks.
[0020] (3) The wrist vein verification method of the present invention with a lightweight attention mechanism deep learning model can capture images by detecting the region of interest in wrist images, avoiding the impact of differences in shooting angle and direction of non-contact detection on subsequent feature extraction and feature comparison results, reducing computation time and improving recognition rate. Attached Figure Description
[0021] To make the technical features, content, advantages, and effects of the present invention clearer, the present invention will now be described in detail below with reference to the accompanying drawings and by way of embodiments:
[0022] Figure 1 This is a flowchart of a wrist vein verification method with a lightweight attention mechanism deep learning model, according to an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of obtaining a visible light image of a wrist vein according to an embodiment of the present invention.
[0024] Figure 3 This is a flowchart of a wrist region of interest detection procedure according to an embodiment of the present invention.
[0025] Figure 4 This is a schematic diagram of the region of interest in the wrist according to an embodiment of the present invention.
[0026] Figure 5 This is a schematic diagram of a vascular image enhancement procedure according to an embodiment of the present invention.
[0027] Figure 6 This is an architecture diagram of a lightweight attention mechanism deep learning model according to an embodiment of the present invention.
[0028] Figure 7 This is a schematic diagram of the improved hourglass module according to an embodiment of the present invention.
[0029] Explanation of reference numerals in the attached figures:
[0030] 11: Smartphone; 12: Wrist; 13: Wrist image; 14: Preprocessing; 15: Visible light image of wrist veins; 21: Wrist region image; 22: Model input image; 30: Lightweight attention mechanism deep learning model; 31: Convolution module; 32: Hourglass module; 32a: First hourglass module; 32b: Second hourglass module; 32c: Third hourglass module; 32d: Fourth hourglass module; 33: Dropout layer; 34: Hybrid pooling coordinate attention module; 35: Parallel spatial pyramid pooling module; 41: First depthwise separable convolutional layer; 42: First pointwise convolutional layer; 43: Second pointwise convolutional layer; 44: Second depthwise separable convolutional layer; 45: Linear separable convolutional module. Detailed Implementation
[0031] To facilitate understanding of the technical features, content, advantages, and effects of the present invention, the present invention will now be described in detail below with reference to the accompanying drawings and embodiments. The accompanying drawings are for illustrative purposes only and are not necessarily the actual proportions and precise configurations of the present invention after implementation. Therefore, the proportions and configurations of the accompanying drawings should not be used to interpret or limit the scope of the claims in actual implementation of the present invention. This is explained in advance.
[0032] Please see Figure 1 This is a flowchart of a wrist vein verification method with a lightweight attention mechanism deep learning model according to an embodiment of the present invention. As shown in the figure, the wrist vein verification method includes the following steps (S1-S6):
[0033] Step S1: Acquire a visible light image of the wrist vein of the subject using an image capturing device and store it in an image processing device. When identifying vein features, the wrist vein is closer to the skin than veins in other organs, making it easier to obtain an image of the vein for detection and verification. Therefore, an image capturing device can be used to photograph the wrist area to obtain images for analysis and verification. Existing technologies use near-infrared cameras to detect wrist vein images, but the cost of setting up near-infrared cameras and the potential harm to human health make existing detection methods quite flawed. Therefore, the image capturing device of this invention uses a visible light camera or video camera to capture a visible light image of the wrist vein, using this visible light image as the image for subsequent analysis and verification.
[0034] Please also refer to Figure 2This is a schematic diagram illustrating the acquisition of a visible light image of a wrist vein according to an embodiment of the present invention. As shown in the figure, the subject takes a picture of their wrist 12 using the camera or video camera of a smartphone 11, obtaining a wrist image 13. This wrist image 13 is a visible light image taken using a general camera. To make the vascular pattern of the wrist vein more obvious, a preprocessing program 14 converts the red, green, and blue values (R, G, B values) of the hue attribute in the image space (HSL color space) into saturation values, forming a visible light image 15 of the wrist vein with more obvious features. This visible light image 15 of the wrist vein is then used for subsequent processing and verification. In this embodiment, the wrist image 13 can be stored in the memory of the smartphone 11. The processor of the phone accesses the memory to execute the preprocessing program 14 to obtain the visible light image 15 of the wrist vein, which is also stored in the memory.
[0035] Step S2: The visible light image of the wrist vein is processed using an image processing device to perform a region of interest (ROI) detection procedure, extracting the ROI image. A vascular image enhancement procedure is then performed on the ROI image to obtain the model input image. Since the visible light image of the wrist vein is acquired by the subject or the examiner using an image capturing device, such as a smartphone, the distance, angle, and direction of the wrist are not fixed in a non-contact environment. To improve the final recognition rate, the visible light image of the wrist vein is processed using an image processing device to perform a region of interest detection procedure, extracting the ROI image. A further vascular image enhancement procedure is then performed to enhance the ROI image to obtain the model input image for model analysis.
[0036] Please also refer to Figure 3 This is a flowchart of the wrist region of interest detection procedure according to an embodiment of the present invention. As shown in the figure, the image processing procedure includes the following steps (S21-S27):
[0037] Step S21: Crop and scale the visible light image of the wrist vein to a predetermined size. To reduce processing time, the visible light image of the wrist vein is cropped and scaled to a predetermined size, for example, cropped and scaled to a central square image of 160×160 pixels.
[0038] Step S22: Segment the image to a predetermined size to generate a wrist region image. In this step, the wrist region is segmented from the background. To avoid undersegmentation and oversegmentation, the area of the wrist region must be maximized, but not larger than 0.8 times the area of the entire image. In this case, the selected channel has the most compact pixel intensity distribution. Next, the wrist region is segmented from the background using Otsu's thresholding method to form a wrist region image, which is then used to find the contour.
[0039] Step S23: Find the maximum contour of the wrist region image. A contour is defined as a curve connecting all points with the same intensity along the image boundary, and it is very useful in shape analysis, object size determination, and object detection. The contour is extracted from the wrist region image using the OpenCV function findContour(), and the maximum contour is used to find the convex hull for the next step.
[0040] Step S24: Find the convex hull of the largest contour and the convexity defect. The smallest contour that completely encloses the object is called the convex hull of the object, and the convexity defect is the deviation between the contour and the convex hull. The convexity defect is obtained using the OpenCV function convexityDefects(). In this embodiment, only the largest convexity defect is retained.
[0041] Step S25: Determine the key vector from the convexity defect. The convexity defect includes a series of attributes: start point, end point, farthest point, and approximate distance to the farthest point. Please refer to the schematic diagram of the detection of the wrist region of interest in an embodiment of the present invention. As shown in the figure, after detecting the contour of the wrist region image 21, the convex hull and convexity defect information are obtained. The points from P1 to P8 are located on the contour. P1 and P2 are the start and end points of the edge of the convex hull P1P2 covering the deepest valley. P3 is the farthest point from P1P2. The line segment P3P4 is perpendicular to the line segment P3P2. The angle θ is composed of two vectors P3P1 and P3P4. If the angle θ is clockwise (θ<0), then P3 is on the left. Here, the vector P3P1 is the key vector indicating the direction of the hand.
[0042] Step S26: Locate the region of interest using key vectors. See also... Figure 4Let P5 be the midpoint of P2P3, and P5P6 be parallel to P3P4. P5P6 can be considered the wrist scale. M1 is the midpoint of P3P4, M2 is the midpoint of P5P6, M3 lies on M1M2 and M1M3 = P5P6, and P7P8 passes through M3 and is parallel to P3P4. In this embodiment, the vertices of the region of interest must be moved inward to avoid clipping to the background. An edge reduction ratio is set, for example, T. L T R B L and B R These are the upper left, upper right, lower left, and lower right regions of interest, respectively, where T L P3 = T R P4 = P3P4 × σ, B L P7 = B R P8 = P7P8 × σ, σ = 0.1, and then the region of interest of the wrist is located using the four corner positioning points.
[0043] Step S27: Extract the region of interest (ROI) image of the wrist. Since the wrist may rotate in three-dimensional space in a non-contact environment, the operation of the above four positioning points projects the ROI of the wrist into two-dimensional space, forming an ROI image of the wrist.
[0044] The region of interest (ROI) image of the wrist can be further enhanced using vascular imaging techniques to obtain the model input image for model analysis. Please refer to [link to relevant documentation]. Figure 5 This is a schematic diagram illustrating the detection of the region of interest on the wrist according to an embodiment of the present invention. Figure 5 As shown in section (a), after the visible light image of the wrist vein is processed by the wrist region of interest (ROI) detection procedure, the original wrist ROI image can be extracted. The vascular image enhancement procedure performs local ridge enhancement (LRE) on this wrist ROI image, including automatic gamma correction (AGC), such as... Figure 5 As shown in section (b), image contrast is increased using Contrast Limited Adaptive Histogram Equalization (CLAHE), as shown below. Figure 5 As shown in section (c). Noise is eliminated by a low-pass Gaussian filter, such as... Figure 5 As shown in section (d), edge details are highlighted using the High-pass Laplacian operator, as... Figure 5As shown in section (e), the final image of the enhanced wrist region of interest is obtained and used as the input image for the model.
[0045] Step S3: Execute a lightweight attention-based deep learning model using an image processing device, inputting the model image to extract feature vectors. As described in the preceding steps, after obtaining the region of interest image of the wrist from the visible light image of the wrist acquired by the image capturing device and performing image enhancement, the obtained model input image can be input into the deep learning network model to extract image features. However, in order to perform the computation on smartphones or other embedded devices, this invention designs a lightweight attention-based deep learning model.
[0046] Please also refer to Figure 6 This is an architecture diagram of the lightweight attention mechanism deep learning model according to an embodiment of the present invention. As shown in the figure, the lightweight attention mechanism deep learning model 30 may include a convolutional module 31, four hourglass modules 32, a dropout layer 33, a hybrid pooling coordinate attention module 34, and a parallel spatial pyramid pooling module 35. The input image 22 is input into the lightweight attention mechanism deep learning model 30, and the feature vector of the wrist image is obtained after multi-layer operations.
[0047] The convolutional module 31 may include standard convolutional layers, batch normalization layers, and non-linear activation function layers. The standard convolutional layers can be used to learn and retain information-rich wrist vein features.
[0048] The four hourglass modules 32 include a first hourglass module 32a, a second hourglass module 32b, a third hourglass module 32c, and a fourth hourglass module 32d. These four hourglass modules 32 can be improved hourglass modules with a bottleneck structure. Please refer to [link / reference]. Figure 7 This is a schematic diagram of the improved hourglass module according to an embodiment of the present invention. As shown in the figure, the hourglass module 32 includes two pointwise convolutional layers and two depthwise separable convolutional layers, which are, in sequence, a first depthwise separable convolutional layer 41, a first pointwise convolutional layer 42, a second pointwise convolutional layer 43, and a second depthwise separable convolutional layer 44. The first pointwise convolutional layer 42 is used for channel reduction, and the second pointwise convolutional layer 43 is used for channel expansion. The bottleneck structure is located between the first pointwise convolutional layer 42 and the second pointwise convolutional layer 43. The first depthwise separable convolutional layer 41 is disposed at the input end of the improved hourglass module 32, and the second depthwise separable convolutional layer 44 is disposed at the output end of the improved hourglass module 32. A linear separable convolutional module 45 is provided to connect the first depthwise separable convolutional layer 41 and the second depthwise separable convolutional layer 44.
[0049] Pointwise convolutions are used to encode inter-channel information but cannot learn spatial information. To encode spatial information and improve network efficiency in mobile devices, depthwise separable convolutional layers are used instead of standard convolutional layers. In the improved hourglass module 32, a deep convolutional layer is followed by a batch normalization layer and a non-linear activation function layer. Two deep convolutional layers are used to capture more spatial information and place it at the end of the residual path instead of adding it to the bottleneck structure. The 3×3 kernel size in the deep convolutional layer is lightweight, reducing the computational burden. In addition, the connection identity mapping is placed in the high-dimensional representation instead of in the bottleneck structure. Therefore, compared with the inverse residual module, more information is transferred from input to output, and more gradients propagate across layers.
[0050] In this embodiment, a nonlinear activation function is chosen because it is a smooth, continuous, self-regulating, and non-monotonic activation function, which facilitates easier optimization, better generalization, a strong normalization effect, and reduces overfitting. Furthermore, the linearly separable convolutional module 45 is used for fast connections to avoid the vanishing gradient problem.
[0051] A dropout layer 33 is set after the hourglass module 32, and the dropout rate can be set to 0.2. Setting the dropout layer 33 can reduce overfitting and improve the generalization ability of the model.
[0052] The hybrid pooling coordinate attention module 34 can perform global average pooling and max pooling along the horizontal and vertical directions respectively, and then combine the global average pooling and max pooling by adding operators before outputting.
[0053] The parallel spatial pyramid pooling module 35 can be configured with three parallel spatial pyramid pooling blocks, each with a different size of partitioned region. The feature maps generated by the operation of the three parallel spatial pyramid pooling blocks are connected and flattened to generate an output feature map, from which feature vectors are extracted.
[0054] Step S4: Compare the feature vector with the registered wrist feature vector of the test subject using the image processing device. If the comparison results match, proceed to step S5: the verification is successful. If the comparison results do not match, proceed to step S6: the verification fails.
[0055] After obtaining the feature vector of the visible light image of the wrist veins captured by the test subject, the result is compared with the registered wrist feature vector of the test subject in the database. The comparison method can be to calculate the cosine distance between the feature vector and the registered wrist feature vector. If the cosine distance is greater than a distance threshold, the result is considered inconsistent and the verification is judged as failed. If the cosine distance is not greater than the distance threshold, the result is considered consistent and the verification is judged as successful. Since the registered wrist feature vector is a comparison standard pre-established by the user, when the comparison result is consistent, it means that the wrist image features captured by the test subject match the registered image features, that is, it is determined to belong to the same wrist as the user, and the user's identity is verified through the verification process. Conversely, if the comparison result is inconsistent, it is considered a different wrist, and the captured wrist image is determined not to be the wrist of the original registered user, thus the verification is judged as failed, and the access or operation of the device cannot continue.
[0056] The wrist vein verification method with a lightweight attention mechanism deep learning model allows users to take an image of their wrist veins using a handheld device and then perform the verification process directly on the device. The lightweight model improves the processing efficiency of the verification process, making it more convenient for users to verify their wrist veins and thus achieve identity verification.
[0057] The above description is merely illustrative and not restrictive. Any equivalent modifications or alterations made without departing from the spirit and scope of this invention should be included within the scope of the claims of this application.
Claims
1. A wrist vein verification method with a lightweight attention mechanism deep learning model, characterized in that, The method comprises the following steps: Obtaining a visible light image of a wrist vein of a subject by an image acquisition device and storing the image in an image processing device; Performing wrist region of interest detection on the visible light image of the wrist vein by the image processing device to extract a wrist region of interest image, and performing a blood vessel image enhancement process on the wrist region of interest image to obtain a model input image; Inputting the model input image into a lightweight attention mechanism deep learning model to extract a feature vector by the image processing device; Comparing the feature vector with a registered wrist feature vector of the subject by the image processing device, and determining that the verification is successful if the comparison result is consistent, or determining that the verification fails if the comparison result is inconsistent.
2. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 1, wherein the visible light image of the wrist vein is converted from a hue attribute with red, green and blue values to a saturation attribute with a saturation value by a preprocessing process. wherein 3. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 1, wherein the wrist region of interest detection process comprises cutting and scaling the visible light image of the wrist vein into a predetermined size image, segmenting the predetermined size image to generate a wrist region image, finding a maximum contour of the wrist region image, finding a convex hull and a convexity defect of the maximum contour, determining a key vector from the convexity defect, locating a region of interest through the key vector, and extracting the wrist region of interest image.
4. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 1, wherein the blood vessel image enhancement process performs local ridge enhancement on the wrist region of interest image, which comprises automatically gamma correcting the wrist region of interest image to limit the contrast adaptive histogram equalization to increase the image contrast, removing noise by a low-pass Gaussian filter, and highlighting edge details by a high-pass Laplacian operator to obtain the model input image. wherein, 5. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 1, wherein the lightweight attention mechanism deep learning model comprises a convolution module, four hourglass modules, a hybrid pooling coordinate attention module, and a parallel spatial pyramid pooling module.
6. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 5, wherein the convolution module comprises a standard convolution layer, a batch normalization layer, and a nonlinear activation function layer. wherein, 7. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 5, wherein wherein wherein wherein The four hourglass modules are improved hourglass modules with a bottleneck structure, including two point-wise convolution layers and two depth-wise separable convolution layers, the two point-wise convolution layers are respectively used for channel reduction and channel expansion, the bottleneck structure is located between the two point-wise convolution layers, the two depth-wise separable convolution layers are respectively arranged at the input end and the output end of the improved hourglass module, and a linear separable convolution module is arranged to connect the two depth-wise separable convolution layers.
8. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 5, wherein, wherein A dropout layer is arranged between the four hourglass modules and the mixed pooling coordinate attention module.
9. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 5, wherein, wherein, The mixed pooling coordinate attention module performs global average pooling and maximum pooling along the horizontal direction and the vertical direction respectively, and then outputs after combining through an addition operator.
10. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 5, wherein, wherein The parallel spatial pyramid pooling module is provided with three parallel spatial pyramid pooling blocks, each having a partition region of different size, and the feature maps generated by the three parallel spatial pyramid pooling blocks are connected and flattened to generate an output feature map, and the feature vector is extracted from the output feature map.
11. The wrist vein verification method with a lightweight attention mechanism deep learning model according to claim 1, wherein, wherein The way of comparing the feature vector with the registered wrist feature vector is to calculate the cosine distance between the feature vector and the registered wrist feature vector, if the cosine distance is greater than a distance threshold, it is considered that the results are inconsistent, and it is judged as verification failure, if the cosine distance is not greater than the distance threshold, it is considered that the results are consistent, and it is judged as verification success.