Dimension-adaptive non-contact palm vein efficient identification method

Feature extraction through adaptive ROI technology and deep learning model, the problem of low processing efficiency of image processing in different sizes in traditional methods is solved, and efficient and accurate palm vein recognition is achieved to avoid the risk of contact infection.

CN118982850APending Publication Date: 2024-11-19GUILIN UNIV OF ELECTRONIC TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411048630.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

Traditional palm vein recognition methods need to zoom when processing images of different sizes, resulting in low recognition speed and accuracy, and fixed-size ROIs are easy to intercept irrelevant information and low image utilization.

Method used

Using the non-contact palm vein efficient recognition method with size adaptability, the palm profile is obtained through HOG features, combined with the SVM classifier to determine the pixel belongs to the foreground or background, perform binarization processing, determine the ROI center, find the maximum incision circle, and rotate correction ROI images are extracted using deep learning models based on ResNet and SppLayer, and dimensionality reduction is used to adopt Euro-style distance feature matching.

Benefits of technology

It is realized that images of different sizes are input to output images of the same size are output, which improves the recognition rate and accuracy, avoids the risk of contact infection, with an accuracy rate of up to 99.6%, and a low probability of error.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982850A_ABST
    Figure CN118982850A_ABST
Patent Text Reader

Abstract

The invention relates to a size-adaptive non-contact palm vein efficient recognition method. The method comprises the steps of S1, palm vein image acquisition; s2, extracting contour information to obtain a foreground region and a background region; s3, carrying out binarization processing on the palm vein image to obtain a binarized image; s4, carrying out corrosion processing and expansion processing on the binarized image, and solving the center of mass of the processed palm vein image as an ROI center; s5, solving the maximum inscribed circle of the palm as an ROI region; s6, rotating and correcting the ROI image; s7, extracting palm vein features of the ROI images of different sizes by adopting a deep learning model based on ResNet and SppLayer, and establishing a feature database; performing dimension reduction processing by using PCA, and then outputting feature vectors with the same size; and S8, adopting a feature matching strategy based on Euclidean distance, and taking the template with the minimum distance as an identification result. According to the method, the palm vein image utilization rate is high, and the recognition rate and accuracy are high.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of biometric identification and deep learning technology, and in particular to a size-adaptive non-contact palm vein efficient identification method. Background Art

[0002] Technology is a double-edged sword. While it provides convenience for human life, it also brings security risks: using "fingerprint film" to steal other people's fingerprint information for illegal activities; using "AI face-changing technology" to make video calls with relatives and friends to extort money; using "synthetic sound tracks" to create fake recordings and then defraud money, etc. The determination of identity has gradually become a top priority for research, and palm vein recognition as an emerging biometric recognition technology has received widespread attention.

[0003] The advantages of palm vein recognition technology are as follows: 1) Accuracy. By irradiating the hemoglobin in the body with near-infrared rays, the vein image is obtained after reflection and collection. If the body is not alive, the blood will not flow and the vein image cannot be obtained; 2) Concealment. Unlike face or fingerprint recognition, the palm vein features are located below the surface of the skin and can only be collected with near-infrared light, which is naturally concealed; 3) Safety. The vein features will no longer change after adulthood. Even for identical twins or the left and right hands of the same person, the palm vein features are different and will not be affected by the external state of the skin or skin color. In addition to the above advantages, the contactless palm vein also avoids bacterial infection caused by contact, ensuring the hygiene of the user who is undergoing identity recognition. Therefore, the identity recognition technology based on contactless palm vein has great potential, and has broad application prospects whether it is used in single mode or multi-mode.

[0004] Chinese patent document CN107341473A discloses a palm feature recognition method, a palm feature recognition device, and a storage medium. After receiving an image acquisition instruction, the first image information of the palm is acquired in real time, and a binary image of the palm is obtained according to the first image information. Then, according to the first image information and the binary image, the moving speed of the palm, the tilt angle and the distance relative to the palm feature recognition device are acquired. When the identifiable conditions are met, the second image information of the palm is acquired. After preprocessing the second image information, the palm shape feature, palm print feature and palm vein feature of the palm are proposed. Finally, the palm shape feature, palm print feature and palm vein feature are identified and compared with the registration template entered into the database.

[0005] Chinese patent document CN112200159A discloses a contactless palm vein recognition method based on an improved residual network, the contactless palm vein recognition method comprising the following steps: 1) collecting two infrared images of the same palm; 2) locating the ROI area of ​​the palm; 3) palm vein registration; 4) palm vein verification; 5) palm vein verification determination: setting a recognition threshold T, calculating the distance between the feature vector and the registration template, if the distance is less than the set recognition threshold T, it is determined that the palm vein verification is successful, otherwise, the verification fails; the present invention proposes an improved residual network structure, the network structure can extract texture features of different scales of input samples during training, so that the trained model has the ability to extract multi-scale texture information, the extracted feature vector better expresses the input sample information, and can solve the problem of multiple scaling and translation of the input sample to a certain extent.

[0006] Chinese patent document CN116524549A discloses a method for positioning key points and ROI of a back of hand or palm vein image based on improved UNet, comprising the following steps: S1: setting up the experimental system machine collection site; S2: building a self-built data set; S3: establishing a UNet model; S4: establishing a residual module; S5: an improved UNET model for positioning key points of back of hand or palm vein images; S6: using an ROI calculation method to calculate the extraction of a region of interest; S7: evaluating the model; and using near-infrared imaging technology to obtain back of hand or palm vein images.

[0007] Chinese patent document CN116978074A discloses a palm vein small sample recognition method based on a Siamese network, specifically: S1: collect palm vein images to construct an image data set and a database, and perform ROI extraction and preprocessing on all images in the image data set; S2: divide the processed images into a training set and a validation set, and adjust the size of all images; S3: select two palm vein images to form an image pair, and annotate the image pairs; S4: input the annotated image pairs into the Siamese network for feature extraction, and use the extracted image pair feature training to generate a model file based on palm vein data; S5: extract the palm vein image features to be matched, import the image features into the model file based on palm vein data, compare them with the features in the database, and check whether the input features are consistent with the existing features in the database, so as to determine the identity.

[0008] In traditional methods, it is usually necessary to scale images of different sizes to the same size and then use a fixed-size ROI for feature extraction. This processing will result in missing sample information. When combined with a deep learning framework, it not only fails to achieve high accuracy but also has a slow speed due to high time complexity.

[0009] Therefore, it is necessary to provide a size-adaptive non-contact palm vein efficient recognition method, as the existing palm vein images have low utilization rate, the fixed-size ROI easily captures irrelevant information, and images of different sizes need to be scaled to the same size, resulting in low recognition speed and accuracy. Summary of the invention

[0010] The present invention provides a size-adaptive non-contact palm vein efficient recognition method, which has a high utilization rate of palm vein images, realizes the output of images of the same size when images of different sizes are input, and has high recognition rate and accuracy.

[0011] In order to solve the above technical problems, the technical solution adopted by the present invention is that the size-adaptive non-contact palm vein efficient recognition method comprises the following specific steps:

[0012] S1 Palm vein image acquisition: collect the user's palm vein image through a non-contact acquisition device to obtain a palm vein image;

[0013] S2 extracts contour information: Use HOG features to obtain edge information of the palm contour, and then combine the SVM classifier to determine whether the pixel belongs to the foreground or background, and obtain the foreground area and background area;

[0014] S3 Binarization processing: performing binarization processing on the palm vein image to obtain a binarized image;

[0015] S4 determines the ROI center: the binary image is corroded and expanded in sequence to obtain a palm vein image after noise removal; the centroid of the processed palm vein image is then calculated as the ROI center;

[0016] S5 Find the maximum inscribed circle of the palm: Take the center of the ROI as the center of the circle, find the maximum inscribed circle that intersects the palm contour, and take the maximum inscribed circle as the ROI area;

[0017] S6: Rotate ROI: Use the line connecting the tangent point of the maximum inscribed circle in step S5 and the center of the circle as a reference, calculate the angle between the line and the horizontal line, and then rotate the original image by the corresponding angle, so as to obtain a rotation-corrected ROI image;

[0018] S7 feature extraction: A deep learning model based on ResNet and SppLayer is used to extract palm vein features of ROI images of different sizes and establish a feature database; PCA is used to perform dimensionality reduction and output feature vectors of the same size;

[0019] S8 feature matching: Adopting the feature matching strategy based on Euclidean distance, the template to be authenticated is matched with the template stored in the feature database in a 1-N manner, and the distance between the feature vectors of two adjacent templates is calculated, and the template with the smallest distance is taken as the recognition result.

[0020] By adopting the above technical scheme, palm vein images are collected by non-contact collection equipment, achieving touch-free safety and hygiene; then the palm vein images are processed, and the palm vein features of ROI images of different sizes are extracted by a deep learning model based on ResNet and SppLayer. After dimensionality reduction processing using PCA, feature vectors of the same size are output, and then feature matching is performed on feature vectors of the same size, so that images of different sizes can be input and images of the same size can be output, and the recognition rate and accuracy are both high.

[0021] Preferably, the non-contact data collection device in step S1 is a near-infrared device.

[0022] Preferably, the specific steps of obtaining the ROI center in step S4 are: firstly perform image erosion on the binary image, reduce and refine the foreground area, and then perform dilation operation to obtain the palm vein image after noise removal, and then calculate the centroid of the palm vein image after erosion and dilation, and use the centroid as the origin of the coordinate system, that is, as the ROI center, and obtain the ROI coordinate system at the same time; the centroid calculation formula of the binary image is:

[0023]

[0024] Among them, N p is the total number of pixels in the palm area, (x p ,y p ) is the pixel coordinate value of the palm area, (x c ,y c ) is the centroid coordinate value. In order to obtain the center point of the ROI area, the image corrosion technology is first used to make the ROI center as close to the center of the palm as possible instead of the center of the whole hand, so as to remove the edge information of the vein image; then the image expansion technology is used to solve the problem of image compression caused by the corrosion technology, which can not only remove noise but also increase the ROI area, obtain more effective sample information, and improve the accuracy.

[0025] Preferably, the specific steps of step S5 are:

[0026] S51: Draw a circle with the center of the ROI as the center, and continuously expand the radius until it intersects with the palm contour to obtain the maximum palm inscribed circle;

[0027] S52: After determining the ROI center, calculate the distance from the ROI center to the palm contour curve, take the minimum value of the distance as the radius of the minimum inscribed circle of the palm contour, and record the corresponding contour point as the minimum inscribed circle tangent point;

[0028] S53: Move the center of the circle along the direction of the ray between the tangent point of the minimum inscribed circle and the center of the circle, and expand the radius until the minimum inscribed circle intersects with the palm contour or the image boundary. The inscribed circle at this time is the largest, and the maximum inscribed circle is used as the ROI area. In order to obtain the maximized palm inscribed circle, after determining the ROI center, calculate the distance from the ROI center to the palm contour curve, and obtain the minimum value of the distance, which is recorded as the shortest line from the ROI center to the palm contour. With the length of the line as the radius and the ROI center as the center, draw a circle to obtain the minimum inscribed circle. Then move the center of the circle along the direction of the ray between the tangent point and the center of the circle, and expand the radius until the inscribed circle intersects with the palm contour or the image boundary. It is considered that the inscribed circle at this time is the largest.

[0029] Preferably, the palm vein feature extraction model based on ResNet and SppLayer in step S7, i.e., the model of palm vein recognition capability, includes a preprocessing layer and four convolutional layers, the convolutional layer has a residual block, and also includes a SppLayer and a fully connected layer. The palm vein ROI image is input into the feature extraction network, and features are extracted through the convolutional layer, and finally PCA is used for dimensionality reduction processing. For example, after inputting SppLayer pooling, 1×2560-dimensional data is obtained, and after the fully connected layer operation, a 1×512-dimensional feature vector is output; finally, PCA is used for dimensionality reduction processing to reduce time complexity and improve the recognition rate while ensuring high accuracy. Using PCA for dimensionality reduction processing can obtain feature vectors with lower dimensions while retaining the main feature information, thereby improving the recognition rate.

[0030] Preferably, the specific steps of training the palm vein feature extraction model based on ResNet and SppLayer, i.e., the model of palm vein recognition capability in step S7 are:

[0031] S71: Use the public palm vein dataset as input image;

[0032] S72: using a ROI method based on maximizing the inscribed circle to convert all the original palm vein images in the palm vein dataset into palm vein ROI images;

[0033] S73: Perform data enhancement on the palm vein ROI image to obtain an enhanced data set; use random cropping and random rotation to enhance the inter-class differences and expand the data set;

[0034] S74: dividing the palm vein ROI image into a training set and a test set, setting hyperparameters, and training the feature extraction model;

[0035] S75: Train until the loss of the training model converges, and select the model with the highest recognition rate, that is, the model that obtains the palm vein recognition ability.

[0036] Preferably, the loss function of the palm vein feature extraction model based on ResNet and SppLayer in step S7 is a combination of a cross entropy loss function and a center loss function, and the loss function calculation formula is:

[0037] Loss = λ·L center (x)+L CEL (x);

[0038] Among them, L center (x) is the center loss function value, L CEL (x) is the cross entropy loss function value, and λ is a constant.

[0039] Preferably, the palm vein data set in step S71 has a total of 500 palms*12 images; the hyperparameters in step S74 are set to a learning rate α=0.001, the number of nodes in the fully connected layer is 500, and the feature vector dimension is 256.

[0040] Preferably, in step S8, Euclidean distance is used as a measure of feature distance. Euclidean distance represents the straight-line distance between two points in a three-dimensional or multi-dimensional space. The following formula is the definition of Euclidean distance. The smaller the Euclidean distance between two samples, the higher the similarity of the two samples. The formula for calculating the Euclidean distance between feature vector A and feature vector B is:

[0041]

[0042] Among them, a and b represent the eigenvectors in two directions, a i and b i Represents the components of two vectors in different dimensions.

[0043] Preferably, in step S7, the extracted feature information is stored to establish a feature database, and the feature database also stores unique identity information.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) Non-contact palm vein recognition avoids bacterial infection caused by contact. Experiments were conducted on a non-contact palm vein dataset with an accuracy rate of up to 99.6% and an average error probability as low as 0.030.

[0046] (2) The palm vein image is firstly obtained using HOG features to obtain the contour information of the palm, and then combined with the SVM classifier to determine which pixels belong to the foreground and which pixels belong to the background;

[0047] (3) Using binarization for processing, setting a threshold according to the pixel result, setting the foreground to 1 and the background to 0, and obtaining a binary image of the palm vein;

[0048] (4) Use image erosion and image dilation techniques to process the binary image, retain more effective information, improve recognition accuracy, and calculate the centroid of the processed binary image through the ordinary centroid algorithm;

[0049] (5) A size-adaptive ROI technology is proposed. The center of mass of the processed binary image is used as the center of the circle to draw an inscribed circle of the palm contour, and finally the ROI area is determined. The line connecting the tangent point of the inscribed circle and the center of the circle is used as the reference to calculate the angle between the line and the horizontal line. When the user's palm rotates a certain angle, the recognition accuracy will not be affected, and it has a certain degree of rotation robustness.

[0050] (6) Using the palm vein feature extraction model of SppLayer, the problem that the traditional method requires all images to be identified to be cropped to the same size is solved. The SppLayer model can convert the input ROI image of any size into an output feature vector of the same dimension;

[0051] (7) Use PCA to reduce dimensionality, reduce time complexity and improve recognition rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is the overall process framework of the size-adaptive non-contact palm vein efficient identification method of the present invention;

[0053] Figure 2 The unprocessed original palm vein image in the size-adaptive non-contact palm vein efficient recognition method of the present invention;

[0054] Figure 3 The comparison diagram is obtained by image corrosion and expansion processing in the size-adaptive non-contact palm vein efficient recognition method of the present invention; wherein (a) is the original image; (b) is the corroded image; (c) is the expanded image after corrosion;

[0055] Figure 4 The processed palm vein image in the size-adaptive non-contact palm vein efficient recognition method of the present invention;

[0056] Figure 5 Comparison of the centroid positions of the original image and the image after corrosion in the size-adaptive non-contact palm vein efficient recognition method of the present invention;

[0057] Figure 6 A schematic diagram of a scale-adaptive rotation angle in the size-adaptive non-contact palm vein efficient identification method of the present invention;

[0058] Figure 7 It is a process of gradually expanding the minimum inscribed circle to the maximum in the size-adaptive non-contact palm vein efficient identification method of the present invention;

[0059] Figure 8 The invention relates to a process of rotating the original image and then performing ROI in the size-adaptive non-contact palm vein efficient recognition method to obtain a final ROI image, and the final ROI image;

[0060] Fig. 9 This is a comparison diagram before and after PCA dimensionality reduction in the size-adaptive non-contact palm vein efficient identification method of the present invention; (a) is before dimensionality reduction; (b) is after dimensionality reduction. DETAILED DESCRIPTION

[0061] The technical solution is further described below in conjunction with the accompanying drawings.

[0062] Example: Figure 1 As shown, the size-adaptive non-contact palm vein efficient recognition method has the following specific steps:

[0063] S1 Palm vein image acquisition: The palm vein image of the user is acquired by a non-contact acquisition device, such as Figure 2 As shown; the non-contact acquisition device in step S1 is a near-infrared device; in this embodiment, a 1920*1080 near-infrared camera is placed in a well-lit environment to capture and acquire the user's palm vein image;

[0064] S2 extracts contour information: Use the HOG (Histogram of Oriented Gradient) feature to obtain the edge information of the palm contour, and then combine it with the SVM (Support Vector Machine) classifier to determine whether the pixel belongs to the foreground (i.e., the palm) or the background, and obtain the foreground area and background area, thereby improving the accuracy of the sample information;

[0065] S3 Binarization: Binarize the palm vein image to obtain a binary image; set the foreground to 1 and the background to 0, and transform the original palm vein image into a binary image; use a method combining OTSU and Gaussian filtering to transform the original palm vein image into a binary image, so that the pixels of the palm and the background are separated from each other; Figure 4 The binary image is shown after Gaussian filtering and then segmented using the OTSU algorithm;

[0066] S4 determines the ROI center: the binary image is corroded and expanded in sequence to obtain a palm vein image after noise removal; the centroid of the processed palm vein image is then calculated as the ROI center; Figure 5 The figure shows the positions of the centroid of the image after corrosion and the centroid of the image before corrosion. Since the influence of the finger part of the binary image needs to be removed when taking the centroid, the image corrosion technology is used to remove the finger part of the binary image.

[0067] The specific steps of obtaining the ROI center in step S4 are: firstly perform image corrosion on the binary image, reduce and refine the foreground area, and then perform dilation operation to obtain the palm vein image after noise removal, then calculate the centroid of the palm vein image after corrosion and dilation, take the centroid as the origin of the coordinate system, that is, as the ROI center, and obtain the ROI coordinate system at the same time; firstly use image corrosion technology to make the ROI center as close as possible to the center of the palm instead of the center of the whole hand, so as to remove the edge information of the vein image; then use image dilation technology to solve the problem of image compression caused by corrosion technology, which can not only remove noise but also increase the ROI area, obtain more effective sample information, and improve the accuracy; the centroid calculation formula of the binary image is:

[0068]

[0069] Among them, N p is the total number of pixels in the palm area, (x p ,y p ) is the pixel coordinate value of the palm area, (x c ,y c ) is the centroid coordinate value; in order to obtain the center point of the ROI area, the image corrosion technology is first used to make the ROI center as close to the center of the palm as possible instead of the center of the whole hand, so as to remove the edge information of the vein image; then the image expansion technology is used to solve the problem of image compression caused by the corrosion technology, which can not only remove noise but also increase the ROI area, obtain more effective sample information, and improve the accuracy; Figure 3 The following is a comparison of images after erosion and dilation. Figure 3 (a) is the original image; Figure 3 (b) in the figure is the eroded image; Figure 3 (c) in the figure is the image after corrosion and expansion. Figure 3 It can be seen that the image accuracy is obviously higher; Figure 5 The figure shows the comparison of the centroid positions of the original image and the processed image, using the centroid position of the image after corrosion and expansion as the origin of the coordinate system;

[0070] S5 Find the maximum inscribed circle of the palm: Take the center of the ROI as the center of the circle, find the maximum inscribed circle that intersects with the palm contour, and use the maximum inscribed circle as the ROI area; Figure 7 As shown, the adaptive ROI technology based on maximizing the inscribed circle is used to take the centroid of the eroded binary image as the center of the circle, make the inscribed circle of the palm contour, and then use the inscribed square of the circle as the final ROI area. It is applicable to palm images of different directions and sizes, and there is no need to manually mark auxiliary points.

[0071] The specific steps of step S5 are:

[0072] S51: Draw a circle with the center of the ROI as the center, and continuously expand the radius until it intersects with the palm contour to obtain the maximum palm inscribed circle;

[0073] S52: After determining the ROI center, calculate the distance from the ROI center to the palm contour curve, take the minimum value of the distance as the radius of the minimum inscribed circle of the palm contour, and record the corresponding contour point as the minimum inscribed circle tangent point;

[0074] S53: moving the center of the circle along the ray direction between the tangent point of the minimum inscribed circle and the center of the circle, and enlarging the radius until the minimum inscribed circle intersects the palm contour or the image boundary. The inscribed circle at this time is the largest, and the largest inscribed circle is used as the ROI area;

[0075] When using the maximum inscribed circle for ROI, there are two methods to select the final ROI area. One is to select the inscribed rectangle of the inscribed circle as the final ROI area, and the other is to directly use the inscribed circle as the final ROI area. The method of selecting the inscribed rectangle is more in line with the habits of traditional image processing, but for deep learning methods, more effective information can improve the accuracy of recognition. The inscribed rectangle method will lose palm contour information due to shape restrictions, so consider using the inscribed circle as the final ROI area; in order to ensure the validity of the vein image information in the ROI, use a black mask to cover the image and only retain the vein image in the inscribed circle, such as Figure 8 Shown is the final ROI image;

[0076] S6 Rotate ROI: In order to make the ROI process rotationally robust, use the line between the tangent point and the center of the largest inscribed circle in step S5 as a reference, and calculate the angle θ between the line and the horizontal line, such as Figure 6 The angle θ shown in is used to rotate the original image by a corresponding degree, that is, using the hand shape information, different palm images of the same user can have the same rotation angle to reduce the intra-class difference; that is, the original image is rotated by a corresponding angle θ to obtain a rotation-corrected ROI image, as shown in Figure 6 As shown in the figure, when the user's palm rotates a certain angle, it will not affect the recognition accuracy, and has a certain degree of rotation robustness;

[0077] S7 feature extraction: A deep learning model based on ResNet and SppLayer is used to extract palm vein features of ROI images of different sizes and establish a feature database; in order to reduce time complexity and thus improve recognition speed, PCA (principal components analysis) is used to perform dimensionality reduction processing and then output feature vectors of the same size; in step S6, the extracted feature information is stored to establish a feature database, and the feature database also stores identity unique identification information; the palm vein feature extraction model based on ResNet and SppLayer in step S7, i.e., the palm vein recognition capability model, ResNet and SppLayer are from the literature Deep residual learning for image recognition and Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition; The structure of the network is shown in Table 1. Combining the characteristics of ResNet that can prevent gradient explosion and SppLayer that can receive input of any size, the network structure consists of two main parts, including a preprocessing layer and four convolutional layers. There are residual blocks in the convolutional layer, which greatly reduces the computational complexity of the neural network; it also includes a SppLayer and a fully connected layer, which realizes the conversion of input of any size into output of the same dimension; finally, PCA is used for dimensionality reduction processing to reduce time complexity and improve the recognition rate while ensuring high accuracy; the palm vein ROI image of any size is input into the feature extraction network, and the features are extracted by the convolutional layer. After input into the SppLayer pooling, 1×2560 dimensional data is obtained, and after the fully connected layer operation, a 1×512 dimensional feature vector is output; PCA (principal components analysis) is used for dimensionality reduction processing to output a feature vector of the same size; Fig. 9 The comparison chart before and after using PCA for dimensionality reduction is shown in the figure. Fig. 9 (a) is before dimensionality reduction. Fig. 9 (b) is after dimensionality reduction. It can be seen that while retaining the main feature information, a feature vector with lower dimension is obtained, which improves the recognition rate.

[0078] Table 1 Feature extraction model structure based on ResNet and SppLayer

[0079]

[0080] The specific steps of training the palm vein feature extraction model based on ResNet and SppLayer, i.e., the palm vein recognition capability model in step S7 are as follows:

[0081] S71: Use a public palm vein dataset as an input image; the palm vein dataset in step S71 has 500 palms*12 images in total; the dataset used to train the deep learning model in this embodiment is the Tongji palm vein dataset (TJU-PV) (from the document Palmprint and palmvein recognition based on DCNN and a new large-scale contactless palmvein dataset);

[0082] S72: using the ROI method based on maximizing the inscribed circle, all 500 palm*12 palm vein original images in the palm vein dataset are converted into palm vein ROI images;

[0083] S73: performing data enhancement on the 500 palm*12 palm vein ROI images to obtain an enhanced data set; enhancing inter-class differences by random cropping and random rotation to expand the data set;

[0084] S74: Divide the 500 palm*12 palm vein ROI images into a training set and a test set, set reasonable hyperparameters, and train the feature extraction model; the hyperparameters are set to a learning rate α=0.001, the number of nodes in the fully connected layer is 500, and the feature vector dimension is 256; the best accuracy of the final model on the test set is 99.6%; the EER (average error rate) obtained by the model on the Tongji dataset is 0.030;

[0085] S75: training until the loss of the training model converges, and selecting the model with the highest recognition rate, that is, the model with palm vein recognition capability;

[0086] The loss function of the palm vein feature extraction model based on ResNet and SppLayer in step S7 is a combination of a cross entropy loss function and a center loss function. The loss function calculation formula is:

[0087] Loss = λ·L center (x)+L CEL (x);

[0088] Among them, L center (x) is the center loss function value, L CEL (x) is the cross entropy loss function value, and λ is a constant.

[0089] In this embodiment, it is set to 0.1;

[0090] S8 feature matching: adopt a feature matching strategy based on Euclidean distance, perform 1-N feature matching on the template to be authenticated and the template stored in the feature database, and calculate the distance between the feature vectors of two adjacent templates, and take the template with the smallest distance as the recognition result; in step S8, Euclidean distance is used as the measure of feature distance. Euclidean distance represents the straight-line distance between two points in three-dimensional or multi-dimensional space. The following formula is the definition of Euclidean distance. The smaller the Euclidean distance between two samples, the higher the similarity of the two samples can be considered; the formula for calculating the Euclidean distance between feature vector A and feature vector B is:

[0091]

[0092] Among them, a and b represent the eigenvectors in two directions, a i and b i Represents the components of two vectors in different dimensions.

[0093] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention, such as changes in certain materials or reaction conditions, should be included in the scope of protection of the present invention.

Claims

1. A size-adaptive non-contact palm vein efficient recognition method, characterized in that: The specific steps are as follows: S1 palm vein image acquisition: acquiring a palm vein image of a user through a non-contact acquisition device to obtain a palm vein image; S2 extracts contour information: Use HOG features to obtain edge information of the palm contour, and then combine the SVM classifier to determine whether the pixel belongs to the foreground or background, and obtain the foreground area and background area; S3 Binarization processing: performing binarization processing on the palm vein image to obtain a binarized image; S4 determines the ROI center: the binary image is corroded and expanded in sequence to obtain a palm vein image after noise removal; the centroid of the processed palm vein image is then calculated as the ROI center; S5 Find the maximum inscribed circle of the palm: Take the center of the ROI as the center of the circle, find the maximum inscribed circle that intersects the palm contour, and take the maximum inscribed circle as the ROI area; S6: Rotate ROI: Use the line connecting the tangent point of the maximum inscribed circle in step S5 and the center of the circle as a reference, calculate the angle between the line and the horizontal line, and then rotate the original image by the corresponding angle, so as to obtain a rotation-corrected ROI image; S7 feature extraction: A deep learning model based on ResNet and SppLayer is used to extract palm vein features of ROI images of different sizes and establish a feature database. PCA is then used to perform dimensionality reduction and output feature vectors of the same size. S8 feature matching: Adopting the feature matching strategy based on Euclidean distance, the template to be authenticated is matched with the template stored in the feature database in a 1-N manner, and the distance between the feature vectors of two adjacent templates is calculated, and the template with the smallest distance is taken as the recognition result.

2. The size-adaptive non-contact palm vein efficient recognition method according to claim 1, characterized in that: The non-contact collection device in step S1 is a near infrared device.

3. The size-adaptive non-contact palm vein efficient recognition method according to claim 1, characterized in that: The specific steps of obtaining the ROI center in step S4 are: firstly perform image erosion on the binary image, reduce and refine the foreground area, and then perform dilation operation to obtain the palm vein image after noise removal, and then calculate the centroid of the palm vein image after erosion and dilation, and use the centroid as the origin of the coordinate system, that is, as the ROI center, and obtain the ROI coordinate system at the same time; the centroid calculation formula of the binary image is: Among them, N p is the total number of pixels in the palm area, (x p ,y p ) is the pixel coordinate value of the palm area, (x c ,y c ) is the centroid coordinate value.

4. The size-adaptive non-contact palm vein efficient identification method according to claim 3, characterized in that: The specific steps of step S5 are: S51: Draw a circle with the center of the ROI as the center, and continuously expand the radius until it intersects with the palm contour to obtain the maximized palm inscribed circle; S52: After determining the ROI center, calculate the distance from the ROI center to the palm contour curve, take the minimum value of the distance as the radius of the minimum inscribed circle of the palm contour, and record the corresponding contour point as the minimum inscribed circle tangent point; S53: Move the center of the circle along the ray direction between the tangent point of the minimum inscribed circle and the center of the circle, and expand the radius until the minimum inscribed circle intersects the palm contour or the image boundary. The inscribed circle at this time is the largest, and the largest inscribed circle is used as the ROI area.

5. The size-adaptive non-contact palm vein efficient identification method according to claim 4, characterized in that: The palm vein feature extraction model based on ResNet and SppLayer in step S7, i.e., the model of palm vein recognition capability, includes a preprocessing layer and four convolutional layers, wherein the convolutional layers have residual blocks, and also include a SppLayer and a fully connected layer. The palm vein ROI image is input into the feature extraction network, features are extracted through the convolutional layers, and finally PCA is used for dimensionality reduction.

6. The size-adaptive non-contact palm vein efficient recognition method according to claim 5, characterized in that: The specific steps of training the palm vein feature extraction model based on ResNet and SppLayer, i.e., the palm vein recognition capability model in step S7 are as follows: S71: Use the public palm vein dataset as input image; S72: using a ROI method based on maximizing the inscribed circle to convert all the original palm vein images in the palm vein dataset into palm vein ROI images; S73: performing data enhancement on the palm vein ROI image to obtain an enhanced data set; S74: dividing the palm vein ROI image into a training set and a test set, setting hyperparameters, and training the feature extraction model; S75: Train until the loss of the training model converges, and select the model with the highest recognition rate, that is, the model that obtains the palm vein recognition ability.

7. The size-adaptive non-contact palm vein efficient identification method according to claim 6, characterized in that: The loss function of the palm vein feature extraction model based on ResNet and SppLayer in step S7 is a combination of a cross entropy loss function and a center loss function. The loss function calculation formula is: Loss=λ·L center (x)+L CEL (x); Among them, L center (x) is the center loss function value, L CEL (x) is the cross entropy loss function value, and λ is a constant.

8. The size-adaptive non-contact palm vein efficient identification method according to claim 6, characterized in that: The palm vein data set in step S71 has a total of 500 palms*12 images; the hyperparameters in step S74 are set to learning rate α=0.001, the number of nodes in the fully connected layer is 500, and the feature vector dimension is 256.

9. The size-adaptive non-contact palm vein efficient identification method according to claim 6, characterized in that: In step S8, Euclidean distance is used as a measure of feature distance. Euclidean distance represents the straight-line distance between two points in a three-dimensional or multi-dimensional space. The formula for calculating the Euclidean distance between feature vector A and feature vector B is: Among them, a and b represent the eigenvectors in two directions, a i and b i Represents the components of two vectors in different dimensions.

10. The size-adaptive non-contact palm vein efficient identification method according to claim 6, characterized in that: In step S7, the extracted feature information is stored to establish a feature database, and the feature database also stores unique identity information.