Training Method and Device for Venous Recognition Model

Abstract

The present invention discloses a training method and device for a vein recognition model, which relates to the technical field of vein recognition. The preset angle range for collection is precisely set from 0 degrees to 90 degrees, and a quasi-vein recognition device is used to ensure consistent image resolution and illumination conditions at different angles. This ensures that the vein images collected at each angle have consistent quality, reducing problems such as image noise and feature loss caused by illumination changes or inconsistent device resolutions, thereby improving the clarity and stability of the images. The Gaussian filtering denoising algorithm is used to process the collected vein images to remove noise introduced by the device or the external environment. Gaussian filtering can smooth the image, and this method effectively reduces image noise and improves the quality of the images. Through multi-angle collection and consistent illumination conditions, the instability during the image collection process is reduced, enabling the collected images to maintain a high degree of consistency at different angles.

Description

Technical Field

[0001] The present invention relates to the technical field of vein recognition, and specifically to a method and device for training a vein recognition model. Background Art

[0002] Vein recognition, as a relatively advanced biometric technology, verifies identity by analyzing vein images of parts such as the palm or finger. The uniqueness of vein image recognition lies in its use of the internal vein blood vessels of the human body as the recognition basis, and these features are not easily forged and have high security. In practical applications, vein recognition mainly relies on high-precision vein image acquisition and analysis methods to achieve accurate identity verification.

[0003] Most current vein recognition systems rely on image acquisition methods from a single perspective. In the face of partial occlusion, angle changes, or poor image quality, this method is prone to a decrease in recognition accuracy. For example, when the hand posture changes or the vein image is occluded, a single-perspective image may not capture enough vein features, thus affecting the performance of the recognition system. In addition, single-perspective images often cannot provide comprehensive information about vein blood vessels, restricting the richness and accuracy of feature extraction.

[0004] Since vein image acquisition usually relies on fixed lighting and shooting angles, any slight posture change or uneven lighting may affect the image quality, resulting in the loss or error of vein features. In this case, the vein recognition model may be interfered with, leading to a high false recognition rate and false rejection rate. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method and device for training a vein recognition model, which solves the problems mentioned in the background art.

[0006] To achieve the above object, the present invention is realized through the following technical solutions: A method for training a vein recognition model includes the following steps:

[0007] Step 1: Use existing vein recognition devices to collect vein images from multiple angles , ensure the clarity of veins at each angle, and maintain the continuity of the images. The collected images are marked as ;

[0008] Step 2: Preprocess the vein images at each angle , including denoising, enhancing contrast, and normalizing to obtain normalized vein images ;

[0009] Step 3: Use affine transformation image registration technology to register the preprocessed vein images from different perspectives Perform alignment to obtain the transformed image after alignment , ensuring that vein features can be accurately corresponding under different perspectives;

[0010] Step Four: Merge the aligned transformed image through the weighted fusion algorithm to generate the final vein image I fu ;

[0011] Step Five: Extract features from the final vein image I fu by using a convolutional neural network to extract the feature vector F;

[0012] Step Six: Input the extracted feature vector F into the vein recognition model for model training to obtain the final recognition result Y. The model trained by the fused images can adapt to vein images at different angles, significantly improving the robustness of the model for vein recognition, especially in dealing with complex situations such as partial occlusion and angle changes.

[0013] Preferably, determine the preset angle of the collected image through the vein recognition device , select the angle range between 0 degrees and 90 degrees. By using a quasi-vein recognition device, ensure that the resolution and lighting conditions of the images at different angles are consistent. After the device calibration is completed, perform the angle setting, and the device takes pictures according to the preset angle;

[0014] Collect vein images at the preset angle , record the images collected at each angle , and continuously collect images from multiple angles in a short time by using an automated image collection mode. Reduce the influence of vein shape and pose changes on the image quality.

[0015] Preferably, perform denoising processing on the vein image by using the Gaussian filtering denoising algorithm to remove the image noise introduced by the collection device and the external environment, and obtain the denoised vein image ;

[0016] The denoised vein image is obtained through the following formula:

[0017] * ;

[0018] In the formula, represents the image pixel value, represents the central position of the pixel, represents the standard deviation of the Gaussian distribution;

[0019] Enhance the denoised vein image by using histogram equalization The contrast is used to emphasize the local contrast of the venous structure, highlight the edges and textures of the venous image, and obtain an enhanced venous image ;

[0020] The enhanced venous image is obtained through the following formula:

[0021] ;

[0022] In the formula, represents the cumulative distribution function of the image pixel values, M*N represents the image size, L represents the maximum gray level of the pixel values, represents the minimum non-zero value of the cumulative distribution function;

[0023] The enhanced venous image is normalized to uniformly map the image pixel values to the standard range, and a normalized venous image is obtained. Normalization can eliminate the image intensity differences caused by the acquisition device or lighting conditions, ensure that the image feature values at different angles are in the same scale range, and facilitate subsequent feature extraction and fusion.

[0024] Preferably, key point P is extracted from the normalized venous image by using a feature point detection algorithm j , and the key point P j is integrated to obtain a key point set , where each key point P j includes a position and a descriptor d j ; Through key point detection and matching, the corresponding points of the venous structure in the images at different angles are found, providing a reference for the calculation of subsequent affine transformation.

[0025] For the normalized venous images from different perspectives , by performing key point matching, the corresponding point pairs between two images are determined by using the nearest neighbor algorithm, and the matching points are obtained;

[0026] According to the matching points , the affine transformation matrix is calculated by the least squares method, and the form of the affine transformation is:

[0027] ;

[0028] In the formula, represents the coordinates in the normalized venous image , represents the coordinates of the venous image after transformation The coordinates in, the affine transformation matrix The parameters a, b, c, d, e, and f in are obtained through matching points Obtained;

[0029] Perform an affine transformation on the normalized vein image And convert it to the coordinate system aligned with the transformed vein image To obtain the transformed image .

[0030] Preferably, perform clarity and quality evaluation on the aligned transformed images at each viewing angle . The quality evaluation is calculated through the image gradient to determine the quality of each image;

[0031] The quality evaluation formula is:

[0032] ;

[0033] In the formula, Represents the quality index of the i-th viewing angle image, M*N represents the image size, Represents the gradient of the image;

[0034] According to the quality evaluation of each viewing angle , calculate the weight value of each image ;

[0035] The weight value Is obtained through the following formula:

[0036] ;

[0037] The weight value Is directly proportional to the clarity of the image, and the sum of all weights is 1, satisfying

[0038] According to the weight value of each viewing angle image , perform pixel-level weighted fusion on the transformed image . For each pixel position , obtain the final vein image I fu ;

[0039] The final vein image I fu Is obtained through the following formula:

[0040] I fu ;

[0041] Wherein, I fu represents the pixel value of the final venous image at the position , and represents the value of the aligned image at this pixel position;

[0042] The final venous image I fu obtained through weighted fusion integrates venous information from multiple perspectives, reduces the occlusion and noise effects of a single perspective, and generates a clearer and more feature-rich venous image.

[0043] Preferably, the final venous image I fu is input into the convolutional layer of the convolutional neural network for a series of convolutional operations. The convolutional layer extracts local features in the image including edges, textures, and venous structures through different filters, and obtains the output of the convolutional layer:

[0044] ;

[0045] Wherein, represents the pixel value of the output feature map of the l-th convolutional layer in channel c, represents the convolutional kernel, represents the bias term, Na represents the size of the convolutional kernel, represents the movement of the convolutional window;

[0046] The feature map output by the convolutional layer is processed by the non-linear activation function ReLU to make the high-frequency information in the feature vector more prominent;

[0047] ; Subsequently, a pooling operation is performed to reduce the dimension of the feature map and retain the most significant venous features. The pooled feature map is .

[0048] Preferably, after passing through multiple convolutional layers, activation layers, and pooling layers, the output pooled feature map is , flattened into a one-dimensional vector, and input into the fully connected layer for feature mapping to generate the final high-dimensional feature vector F;

[0049] The high-dimensional feature vector F is obtained through the following formula:

[0050] F ;

[0051] Wherein, respectively represent the weight value and bias term of the fully connected layer, represents the activation function Sigmoid;

[0052] The high-dimensional feature vector F is regularized and normalized to unit length so that the feature vectors of different samples can be effectively compared in the subsequent recognition stage:

[0053] Fn ;

[0054] wherein, represents the norm of the feature vector.

[0055] Preferably, the extracted high-dimensional feature vector F is used as the input of the vein recognition model and passes through a series of fully connected layers, activation layers and classification layers to generate the output prediction value y of the model. The output prediction value y is obtained through the following formula: y = M(F);

[0056] The cross-entropy loss function is used to measure the difference between the output prediction value of the model and the true label ;

[0057] ;

[0058] wherein, represents the one-hot encoding of the true label, represents the probability that the model output belongs to class k; among them, the smaller the cross-entropy loss, the closer the model prediction is to the true label;

[0059] According to the calculation result of the loss function, the gradient of the loss with respect to the model parameters is calculated through the backpropagation algorithm, and the SGD is used to update the parameters of the model; the model gradually reduces the loss function value through multiple iterative learning;

[0060] ;

[0061] wherein, represents the model parameters, represents the learning rate, represents the gradient of the loss function with respect to the model parameters.

[0062] Preferably, during the training process, a part of the validation set data that has not participated in the training is used, input into the model for forward propagation, to obtain the final recognition result Y of the model, and the loss and accuracy of the validation set are calculated. The loss and accuracy of the validation set are used to monitor the overfitting situation of the model and are adjusted through early stopping and model regularization;

[0063] The calculation of the validation set loss is the same as the training loss;

[0064] ;

[0065] The formula for the validation set accuracy is:

[0066] ;

[0067] In the formula, represents the number of samples in the validation set, represents the model output whether it is consistent with the true label ;

[0068] After the training is completed, independent test set data is used for the final evaluation of the model to obtain the performance of the model on the test set, including the test set accuracy , confusion matrix, precision, recall, and F1-score metrics.

[0069] A training device for a vein recognition model, the training device includes:

[0070] A multi-view image acquisition module, which collects vein images from multiple angles by using existing vein recognition devices , ensuring the clarity of the veins at each angle and maintaining the continuity of the images. The collected images are marked as ;

[0071] An image preprocessing module, which preprocesses the vein images at each angle , including denoising, enhancing contrast, and normalization processing, to obtain normalized vein images ;

[0072] An image alignment and correction module, which aligns the preprocessed vein images from different perspectives by using affine transformation image registration technology to obtain the aligned transformed image , ensuring that the vein features can be accurately corresponding under different perspectives;

[0073] A multi-view image fusion module, which merges the aligned transformed images through a weighted fusion algorithm to generate the final vein image I fu ;

[0074] A feature extraction module, which extracts features from the final vein image I fu , and extracts the feature vector F by using a convolutional neural network;

[0075] A model training and evaluation module, which inputs the extracted feature vector F into the vein recognition model for model training to obtain the final recognition result Y.

[0076] The present invention provides a training method and device for a vein recognition model, having the following beneficial effects:

[0077] (1) The preset angle range for acquisition is accurately set to 0 degrees to 90 degrees, and the quasi-vein recognition device is used to ensure consistent image resolution and illumination conditions at different angles. This ensures that the vein images acquired at each angle have consistent quality, reducing image noise and feature loss problems caused by illumination changes or inconsistent device resolutions. As a result, the clarity and stability of the images are improved. The Gaussian filtering denoising algorithm is used to process the acquired vein images to remove noise introduced by the device or the external environment. Gaussian filtering can smooth the images, reducing noise interference while retaining important vein features. This method effectively reduces image noise, improves the quality of the images, makes the vein structure more obvious, and reduces the risk of misidentification caused by noise.

[0078] Through multi-angle acquisition and consistent illumination conditions, the instability during the image acquisition process is reduced, enabling the acquired images to maintain high consistency at different angles. This provides high-quality basic data for subsequent image fusion and alignment.

[0079] (2) Through accurate feature point matching, vein images from different perspectives can be efficiently corresponded, reducing the possibility of mismatching. Especially in the case where vein features undergo transformations such as rotation and scaling, the accuracy of image alignment is greatly improved. Affine transformation can effectively correct geometric distortions caused by different shooting angles, ensuring the consistent geometric structure of vein images, thereby reducing errors during the image fusion process and enhancing the robustness of vein recognition. Through quality assessment, high-quality images can be automatically selected and weighted, avoiding the negative impact of low-quality images on the final recognition result. The effectiveness of image fusion is improved, making the finally generated vein image clearer and more feature-rich. Pixel-level weighted fusion ensures that the values of each pixel are reasonably weighted among images from multiple perspectives, and the generated vein image I fu is clearer and contains feature information from multiple angles.

[0080] (3) Nonlinear processing is performed on the output of the convolutional layer by using the ReLU activation function to highlight the high-frequency information in the feature vector. Next, the pooling operation reduces the redundant information in the feature map through dimensionality reduction, preserves the most significant vein features, and reduces the computational amount. The ReLU activation function can effectively improve the sensitivity of the model to high-frequency vein features and increase the non-linear expression ability, while the pooling operation reduces the computational amount and maintains the key features at the same time. By reducing the redundant information after pooling, the model can identify more efficiently and is not easily interfered by noise and irrelevant information. The fully connected layer integrates the local features extracted by the convolutional layer into global features, enhancing the model's understanding of the overall pattern of vein images. By regularizing the feature vector, the influence of the difference in the length of the feature vector on comparison is avoided, making the model more robust when identifying different vein samples.

[0081] (4) Through multi-view image fusion, the vein recognition system can integrate information from multiple angles, significantly improving the accuracy of recognition. Fusing the vein images under each view can capture more comprehensive and detailed vein features, reducing the recognition error caused by a single view.

[0082] The fused vein image can effectively reduce the impact of partial occlusion, pose changes, and image quality problems on the recognition result. The adaptability of the system to different views and occlusion situations is significantly enhanced, improving the stability and reliability of the model in practical applications. By fusing the aligned multi-view images, the generated final vein image contains more information and features, enabling the feature extraction process to extract richer and more accurate vein features. This is crucial for training an efficient vein recognition model and helps to improve the final recognition performance. Description of the Drawings

[0083] Figure 1 It is a schematic flow chart of the steps of the training method of the vein recognition model of the present invention;

[0084] Figure 2 It is a schematic block diagram flow chart of the training device of the vein recognition model of the present invention. Detailed Embodiments

[0085] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0086] Embodiment 1

[0087] The present invention provides a training method and device for a vein recognition model. Please refer toFigure 1 , including the following steps:

[0088] Step 1: By using an existing vein recognition device, collect vein images from multiple angles , ensure the clarity of the veins at each angle and maintain the continuity of the images. The collected images are marked as ;

[0089] Step 2: Preprocess the vein images at each angle , including denoising, enhancing contrast, and normalization processing, to obtain normalized vein images ;

[0090] Step 3: By using affine transformation image registration technology, align the preprocessed vein images from different perspectives to obtain the aligned transformed images ;

[0091] Step 4: Merge the aligned transformed images through a weighted fusion algorithm to generate the final vein image I fu ;

[0092] Step 5: Extract features from the final vein image I fu , and extract the feature vector F by using a convolutional neural network;

[0093] Step 6: Input the extracted feature vector F into the vein recognition model for model training to obtain the final recognition result Y.

[0094] In this embodiment, an existing vein recognition device is used to collect vein images from multiple angles, ensuring the clarity of the veins at each angle and maintaining the continuity of the images. This step provides multi-dimensional information of the vein blood vessels through images from different perspectives, which helps to capture more comprehensive and detailed vein features.

[0095] By collecting images at multiple fixed angles with the device, the recognizability of the vein structure in each image is ensured, reducing image loss or distortion caused by angle changes or pose problems.

[0096] By using affine transformation image registration technology, align the preprocessed vein images from different perspectives. The aligned images ensure that the vein features at each perspective can accurately correspond, reducing feature deviation caused by perspective differences. Apply the affine transformation algorithm to adjust the position, rotation, and scaling of the images, so that the vein images from different perspectives are aligned in the same coordinate system, optimizing the registration accuracy of the images. Merge the aligned transformed images through a weighted fusion algorithm to generate the final vein image I fu, the fused image can integrate the features from multiple perspectives, reduce noise and occlusion problems, and improve the effect of feature extraction.

[0097] Use the weighted fusion algorithm to perform weighted averaging on multiple aligned images according to weights to obtain a vein image that combines multi-perspective features. For the final vein image I fu Perform feature extraction, and use a convolutional neural network to extract high-dimensional feature vectors F. The extracted feature vectors can accurately represent the vein structure in the vein image and provide rich feature information for the training of the model. Input the extracted feature vectors F into the vein recognition model for training to obtain the final recognition result Y. Through model training, the recognition performance can be optimized, making the model more accurate in classifying and recognizing vein images. Input the feature vectors F into the vein recognition model, and adjust the model parameters through backpropagation and optimization algorithms so that the finally output recognition result Y can accurately match the vein features.

[0098] Embodiment 2

[0099] This embodiment is an explanatory description based on Embodiment 1. Please refer to Figure 1 , specifically: Determine the preset angle of collection through the vein recognition device , select the angle range between 0 degrees and 90 degrees. By using the quasi-vein recognition device, ensure that the resolution and lighting conditions of the images at different angles are consistent. After the device calibration is completed, perform the angle setting, and the device takes pictures according to the preset angle;

[0100] At the preset angle Perform vein image collection, and record the images collected at each angle , and continuously collect images from multiple angles in a short time by using an automated image collection mode.

[0101] Perform denoising processing on the vein image by using the Gaussian filtering denoising algorithm to remove the image noise introduced by the collection device and the external environment, and obtain the denoised vein image ;

[0102] The denoised vein image is obtained through the following formula:

[0103] * ;

[0104] In the formula, represents the image pixel value, represents the central position of the pixel, represents the standard deviation of the Gaussian distribution;

[0105] Enhancing denoised vein images by using histogram equalization to enhance the contrast, emphasize the local contrast of the vein structure, highlight the edges and textures of the vein image, and obtain an enhanced vein image ;

[0106] The enhanced vein image is obtained by the following formula:

[0107] ;

[0108] wherein, represents the cumulative distribution function of the image pixel values, M*N represents the image size, L represents the maximum gray level of the pixel values, represents the minimum non-zero value of the cumulative distribution function;

[0109] Normalize the enhanced vein image to uniformly map the image pixel values to the standard range to obtain a normalized vein image .

[0110] In this embodiment, the preset angle range collected is accurately set from 0 degrees to 90 degrees, and the image resolution and illumination conditions at different angles are ensured to be consistent through the quasi-vein recognition device. This ensures that the vein images collected at each angle have consistent quality, reduces the image noise and feature loss problems caused by illumination changes or inconsistent device resolutions, thereby improving the clarity and stability of the images. The Gaussian filtering denoising algorithm is used to process the collected vein images to remove the noise introduced by the device or the external environment. Gaussian filtering can smooth the image, reduce noise interference while retaining important vein features. This method effectively reduces the image noise, improves the image quality, makes the vein structure more obvious, and reduces the risk of misidentification caused by noise.

[0111] Through multi-angle collection and consistent illumination conditions, the instability in the image collection process is reduced, so that the collected images maintain high consistency at different angles. This provides high-quality basic data for subsequent image fusion and alignment.

[0112] Through Gaussian filtering denoising and histogram equalization enhancement, the details of the vein structure, especially the edges and textures of the veins, can be effectively highlighted. The enhanced image enables the subsequent feature extraction process to capture more useful information, significantly improving the recognition accuracy. Gaussian filtering and normalization processing help eliminate the image quality problems caused by device noise and environmental illumination, enabling the model to be based on cleaner and standardized image data during training and inference. This reduces the complexity during model training and the misidentification rate caused by image quality differences.

[0113] Example 3

[0114] This example is an explanation given in Example 2. Please refer to Figure 1 , specifically: By using the feature point detection algorithm on the normalized vein image Extract key point P j , and integrate the key point P j to obtain a set of key points , where each key point P j includes a position and a descriptor d j ;

[0115] For normalized vein images from different perspectives , by performing key point matching, use the nearest neighbor algorithm to determine the corresponding point pairs between two images and obtain the matching points ;

[0116] According to the matching points calculate the affine transformation matrix by the least squares method , and the form of the affine transformation is:

[0117] ;

[0118] In the formula, represents the coordinates in the normalized vein image , represents the coordinates in the vein image after transformation , and the parameters a, b, c, d, e, and f in the affine transformation matrix are obtained from the matching points ;

[0119] Perform an affine transformation on the normalized vein image and convert it to a coordinate system aligned with the vein image after transformation to obtain the transformed image .

[0120] For each aligned transformed image from each perspective perform sharpness and quality assessment. The quality assessment is calculated through image gradients to determine the quality of each image;

[0121] The quality assessment formula is:

[0122] ;

[0123] In the formula, represents the quality index of the i-th perspective image, M*N represents the image size, Represents the gradient of the image;

[0124] Based on the quality evaluation of each perspective , calculate the weight value of each image ;

[0125] The said weight value is obtained through the following formula:

[0126] ;

[0127] The weight value is directly proportional to the clarity of the image, and the sum of all weights is 1, satisfying

[0128] Based on the weight value of each perspective image , perform pixel-level weighted fusion on the transformed image . For each pixel position , obtain the final vein image I fu ;

[0129] The said final vein image I fu is obtained through the following formula:

[0130] I fu ;

[0131] In the formula, I fu represents the pixel value of the final vein image at the position , represents the value of the aligned image at this pixel position;

[0132] The final vein image I obtained through weighted fusion fu integrates the vein information of multiple perspectives, reduces the occlusion and noise effects of a single perspective, and generates a clearer and more feature-rich vein image.

[0133] In this embodiment, through accurate feature point matching, vein images from different perspectives can be efficiently corresponded, reducing the possibility of mismatching. Especially in the case of transformations such as rotation and scaling of vein features, the accuracy of image alignment is greatly improved. Affine transformation can effectively correct the geometric distortion caused by different shooting angles, ensuring the consistency of the geometric structure of vein images, thereby reducing the error in the image fusion process and enhancing the robustness of vein recognition. Through quality assessment, high-quality images can be automatically selected and weighted, avoiding the negative impact of low-quality images on the final recognition result. The effectiveness of image fusion is improved, making the finally generated vein image clearer and more feature-rich. Pixel-level weighted fusion ensures that the values of each pixel are reasonably weighted among images from multiple perspectives, generating the vein image I fu is clearer and contains feature information from multiple angles. Through the fused image, the integrity of the vein structure is enhanced, improving the accuracy and robustness of the vein recognition model.

[0134] By performing quality assessment on vein images from different perspectives and performing weighted fusion according to the assessment results, the negative impact of low-quality images on the final recognition result can be minimized, thereby obtaining a clearer vein image.

[0135] The weighted mechanism of multi-perspective image fusion ensures the integration of information from multiple angles, reduces the occlusion or noise problems brought by a single angle, and the generated image is more complete and has richer details. This significantly improves the accuracy and reliability of the model in the vein recognition process. Since the fused image contains vein information from multiple angles, the finally generated vein image I fu is clearer and more feature-rich. This makes the vein recognition model more robust in dealing with different angles, illumination changes, and partial occlusions, significantly improving the recognition accuracy of the system.

[0136] Embodiment 4

[0137] This embodiment is an explanatory description based on Embodiment 3. Please refer to Figure 1 , specifically: Input the final vein image I fu into the convolutional layer of the convolutional neural network for a series of convolutional operations. The convolutional layer extracts local features in the image including edges, textures, and vein structures through different filters, and obtains the output of the convolutional layer:

[0138] ;

[0139] In the formula, represents the pixel value of the output feature map of the l-th convolutional layer in channel c, represents the convolutional kernel, denotes the bias term, and Na denotes the size of the convolutional kernel. represents the movement of the convolutional window;

[0140] The feature map output by the convolutional layer is processed by the non-linear activation function ReLU to make the high-frequency information in the feature vector more prominent;

[0141] ; Subsequently, a pooling operation is performed to reduce the dimension of the feature map and retain the most significant vein features. The pooled feature map is .

[0142] Use a convolutional neural network to extract features from the final vein image I fu For feature extraction, through multiple convolutional operations, the network can capture local features in the image, including edges, textures, and vein structures. By operating with different convolutional kernels, the detailed information of the vein image is extracted layer by layer, making the high-dimensional feature vector F more refined. The convolutional neural network can automatically learn more abundant and efficient vein features from the image, reducing the dependence on manually designed features. The layer-by-layer extraction of the convolutional layer enables accurate capture of vein detail features at different levels, contributing to improving the accuracy and robustness of vein recognition.

[0143] After passing through multiple convolutional layers, activation layers, and pooling layers, the output pooled feature map is , flattened into a one-dimensional vector, and input into the fully connected layer for feature mapping to generate the final high-dimensional feature vector F;

[0144] The high-dimensional feature vector F is obtained through the following formula:

[0145] F ;

[0146] In the formula, respectively represent the weight values and bias terms of the fully connected layer, represents the activation function Sigmoid;

[0147] Perform regularization processing on the high-dimensional feature vector F, normalize it to unit length, and the feature vectors of different samples can be effectively compared in the subsequent recognition stage:

[0148] Fn

[0149] In the formula, represents the norm of the feature vector.

[0150] The output of the convolutional layer is non-linearly processed by using the ReLU activation function to highlight the high-frequency information in the feature vector. Next, the pooling operation reduces the redundant information in the feature map by dimensionality reduction, retains the most significant vein features, and reduces the computational amount. The ReLU activation function can effectively enhance the sensitivity of the model to high-frequency vein features and increase the non-linear expression ability, while the pooling operation reduces the computational amount and maintains the key features at the same time. By reducing the redundant information after pooling, the model can identify more efficiently and is not easily interfered by noise and irrelevant information.

[0151] The feature map after the convolutional layer and the pooling layer is flattened into a one-dimensional vector and input into the fully connected layer for feature mapping to generate a high-dimensional feature vector F. The high-dimensional feature vector F is regularized and normalized to unit length to ensure that the feature vectors of different samples can be effectively compared in the subsequent recognition stage. The fully connected layer integrates the local features extracted by the convolutional layer into global features, enhancing the model's understanding of the overall pattern of vein images. By regularizing the feature vector, the influence of the difference in the length of the feature vector on comparison is avoided, making the model more robust when identifying different vein samples.

[0152] The extracted high-dimensional feature vector F is used as the input of the vein recognition model After passing through a series of fully connected layers, activation layers and classification layers, the output prediction value y of the model is generated. The output prediction value y is obtained through the following formula: y = M(F);

[0153] The cross-entropy loss function is used to measure the difference between the output prediction value of the model and the true label ;

[0154] ;

[0155] In the formula, represents the one-hot encoding of the true label, represents the probability that the model output belongs to class k; among them, the smaller the cross-entropy loss, the closer the prediction of the model is to the true label;

[0156] According to the calculation result of the loss function, the gradient of the loss with respect to the model parameters is calculated by the backpropagation algorithm, and the parameters of the model are updated using SGD; the model gradually reduces the value of the loss function through multiple iterations of learning;

[0157] ;

[0158] In the formula, represents the model parameters, represents the learning rate, Represents the gradient of the loss function with respect to the model parameters.

[0159] During the training process, a part of the validation set data that has not participated in the training is used. It is input into the model for forward propagation to obtain the final recognition result Y of the model, and the loss of the validation set is calculated. And accuracy , the loss of the validation set And accuracy Are used to monitor the overfitting situation of the model and are adjusted through early stopping and model regularization;

[0160] The calculation of the validation set loss is the same as that of the training loss;

[0161] ;

[0162] The formula for the validation set accuracy is:

[0163] ;

[0164] In the formula, Represents the number of samples in the validation set, Represents the model output Whether it is consistent with the true label ;

[0165] After training is completed, independent test set data is used for the final evaluation of the model to obtain the performance of the model on the test set, including the test set accuracy , confusion matrix, precision, recall, and F1-score metrics.

[0166] In this embodiment, the cross-entropy loss function is used to measure the difference between the model output prediction value and the true label. The gradient of the loss with respect to the model parameters is calculated through the backpropagation algorithm, and SGD is used to update the model parameters. This iterative optimization method can gradually reduce the value of the loss function and improve the recognition accuracy of the model.

[0167] The cross-entropy loss function is very effective for classification tasks, which can accelerate the convergence of the model. At the same time, the way of combining the backpropagation algorithm with SGD to update parameters can optimize the model performance in each iteration, gradually approaching the best parameter combination, and improving the accuracy and robustness of vein recognition. During the training process, the validation set is used to monitor the model, calculate the loss and accuracy of the validation set, and avoid model overfitting through methods such as early stopping and regularization, thereby improving the generalization ability of the model.

[0168] Using a validation set can monitor the performance of the model on unseen data in real time, preventing the model from overfitting to the training data and performing poorly in actual applications. Through techniques such as early stopping and regularization, the generalization ability of the model can be effectively improved, making the model more stable in actual vein recognition applications.

[0169] Example 5

[0170] For the training method and device of the vein recognition model, please refer to Figure 2 , specifically: The training device includes:

[0171] A multi-view image acquisition module, by using existing vein recognition devices, collects vein images from multiple angles , ensuring the clarity of the veins at each angle and maintaining the continuity of the images. The collected images are labeled as ;

[0172] An image preprocessing module preprocesses the vein images at each angle , including denoising, enhancing contrast, and normalization processing, to obtain normalized vein images ;

[0173] An image alignment and correction module aligns the preprocessed vein images from different perspectives by using affine transformation image registration technology to obtain the aligned transformed images ;

[0174] A multi-view image fusion module merges the aligned transformed images through a weighted fusion algorithm to generate the final vein image I fu ;

[0175] A feature extraction module extracts features from the final vein image I fu and extracts feature vectors F by using a convolutional neural network;

[0176] A model training and evaluation module inputs the extracted feature vectors F into the vein recognition model for model training to obtain the final recognition result Y.

[0177] In this embodiment, through multi-view image fusion, the vein recognition system can integrate information from multiple angles, significantly improving the recognition accuracy. Fusing the vein images at each perspective can capture more comprehensive and detailed vein features, reducing the recognition errors caused by a single perspective.

[0178] The fused vein image can effectively reduce the impact of partial occlusion, pose variation, and image quality problems on the recognition result. The adaptability of the system to different perspectives and occlusion situations is significantly enhanced, improving the stability and reliability of the model in practical applications.

[0179] By fusing the aligned multi-perspective images, the generated final vein image contains more information and features, enabling the feature extraction process to extract richer and more accurate vein features. This is crucial for training an efficient vein recognition model and helps to improve the final recognition performance.

[0180] The vein image after fusing multi-perspective images provides more useful information, enabling the model to converge to the optimal solution faster during training, reducing the consumption of training time and computing resources.

[0181] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A training method for a vein recognition model, characterized in that: Including the following steps: Step 1: Use existing vein recognition devices to collect vein images from multiple angles , ensure the clarity of the veins at each angle, and maintain the continuity of the images. The collected images are marked as ; Step 2: For the venous images at each angle perform preprocessing, including denoising, contrast enhancement, and normalization, to obtain normalized venous images ; Step 3: By using the affine transformation image registration technique, align the preprocessed vein images from different perspectives to obtain the transformed image after alignment ; Step 4: The aligned transformed images are merged through a weighted fusion algorithm to generate the final vein image I fu ; for the aligned transformed images under each perspective, sharpness and quality evaluation are performed. The quality evaluation is calculated through image gradients to determine the quality of each image; The quality assessment formula is: ； In the formula, represents the quality index of the i-th perspective image, M*N represents the image size, represents the gradient of the image; Quality evaluation according to each perspective , calculate the weight value of each image ; The weight value is obtained through the following formula: ； Weight value is directly proportional to the clarity of the image, and the sum of all weights is 1, satisfying ; According to the weight value of each perspective image , perform pixel-level weighted fusion on the transformed image . For each pixel position , obtain the final vein image I fu ; The final venous image I fu is obtained by the following formula: I fu ； where I fu represents the pixel value of the final venous image at the position , and represents the value of the aligned image at this pixel position; The final vein image I obtained through weighted fusion fu Integrates vein information from multiple perspectives, reduces the occlusion and noise effects of a single perspective, and generates a clearer and more feature-rich vein image; Step Five: Perform feature extraction on the final vein image I fu by using a convolutional neural network to extract the feature vector F; input the final vein image I fu into the convolutional layer of the convolutional neural network for a series of convolutional operations. The convolutional layer extracts local features in the image, including edges, textures, and vein structures, through different filters, and obtains the output of the convolutional layer: ； In the formula, represents the pixel value of the output feature map of the l-th convolutional layer in channel c, represents the convolutional kernel, represents the bias term, Na represents the size of the convolutional kernel, represents the movement of the convolutional window; Feature map output by the convolutional layer The high-frequency information in the feature vector is made more prominent by processing through the non-linear activation function ReLU; ; Subsequently, a pooling operation is performed to reduce the dimension of the feature map and retain the most significant vein features. The feature map after pooling is ; After passing through multiple convolutional layers, activation layers, and pooling layers, the output pooled feature map will be , flattened into a one-dimensional vector, and input into a fully connected layer for feature mapping to generate the final high-dimensional feature vector F; The high-dimensional feature vector F is obtained through the following formula: F ； In the formula, respectively represent the weight value and bias term of the fully connected layer, represents the activation function Sigmoid; Perform regularization processing on the high-dimensional feature vector F and normalize it to unit length so that the feature vectors of different samples can be effectively compared in the subsequent recognition stage: Fn ; In the formula, represents the norm of the eigenvector; Step 6: Input the extracted feature vector F into the vein recognition model for model training to obtain the final recognition result Y.

2. The training method of the vein recognition model according to claim 1, characterized in that: Determine the preset angle of acquisition through the vein recognition device , select an angle range between 0 degrees and 90 degrees. By using a quasi-vein recognition device, ensure that the resolution and lighting conditions of the images are consistent at different angles. After the device calibration is completed, perform the angle setting, and the device takes pictures according to the preset angle; At a preset angle venous images are collected, and the images collected at each angle are recorded , and by using an automated image acquisition mode, images are continuously collected from multiple angles within a short period of time.

3. The training method of the vein recognition model according to claim 1, wherein: By using the Gaussian filtering denoising algorithm for the vein image to perform denoising processing, removing the image noise introduced by the acquisition device and the external environment, and obtaining the denoised vein image ; The denoised vein image is obtained by the following formula: * ； In the formula, represents the image pixel value, represents the central position of the pixel, represents the standard deviation of the Gaussian distribution; Enhancing and denoising vein images by using histogram equalization to enhance the contrast, emphasize the local contrast of the vein structure, highlight the edges and textures of the vein image, and obtain enhanced vein images ; The enhanced vein image is obtained by the following formula: ； wherein, represents the cumulative distribution function of the image pixel values, M*N represents the image size, L represents the maximum gray level of the pixel values, represents the minimum non-zero value of the cumulative distribution function; Enhance the venous image Perform normalization to uniformly map the image pixel values to the standard range and obtain the normalized venous image .

4. The training method of the vein recognition model according to claim 1, characterized in that: By using a feature point detection algorithm on the normalized vein image Extract the key point P j , and integrate the key point P j to obtain a set of key points , where each key point P j includes a position and a descriptor d j ; Normalized vein images from different perspectives , by performing key point matching, determining corresponding point pairs between two images using the nearest neighbor algorithm, and obtaining matching points ; According to the matching points Calculate the affine transformation matrix by the least squares method , and the form of the affine transformation is as follows: ； In the formula, represents the coordinates in the normalized vein image , represents the coordinates in the vein image after transformation . The parameters a, b, c, d, e, and f in the affine transformation matrix are obtained by matching points . Perform an affine transformation on the normalized vein image to transform it into a coordinate system aligned with the transformed vein image and obtain the transformed image .

5. The training method of the vein recognition model according to claim 4, characterized in that: Use the extracted high-dimensional feature vector F as the input of the vein recognition model After passing through a series of fully connected layers, activation layers and classification layers, the output prediction value y of the model is generated. The output prediction value y is obtained through the following formula: y = M(F); Use the cross-entropy loss function to measure the output prediction value of the model and the true label for the difference between them; ； wherein, represents the one-hot encoding of the true label, represents the probability that the model output belongs to class k; among them, the smaller the cross-entropy loss, the closer the model prediction is to the true label; According to the calculation result of the loss function, calculate the gradient of the loss with respect to the model parameters through the backpropagation algorithm, and use SGD to update the model parameters; the model learns through multiple iterations to gradually reduce the value of the loss function; ； In the formula, represents the model parameters, represents the learning rate, represents the gradient of the loss function with respect to the model parameters.

6. The training method of the vein recognition model according to claim 5, characterized in that: During the training process, a part of the validation set data that has not participated in the training is used and input into the model for forward propagation to obtain the final recognition result Y of the model, and the loss of the validation set is calculated. And accuracy , the loss of the validation set And accuracy are used to monitor the overfitting situation of the model and are adjusted through early stopping and model regularization; The calculation of the validation set loss is the same as that of the training loss; ； The formula for the validation set accuracy is: ； In the formula, represents the number of samples in the validation set, represents the model output whether it is consistent with the true label; After training is completed, the independent test set data is used to conduct the final evaluation of the model, and the performance of the model on the test set is obtained, including the test set accuracy , confusion matrix, precision, recall, and F1-score metrics.

7. A training device for a vein recognition model, characterized in that: The training device includes: Multi-view image acquisition module, by using existing vein recognition devices, collects vein images from multiple angles , ensuring the clarity of veins at each angle and maintaining the continuity of the images. The collected images are marked as ; The image preprocessing module preprocesses the vein images at each angle including denoising, enhancing contrast and normalizing to obtain normalized vein images ; Image alignment and calibration module, which uses affine transformation image registration technology to align the preprocessed vein images from different perspectives to obtain the transformed image after alignment ; Multi-view image fusion module, which fuses the aligned transformed images through a weighted fusion algorithm to generate the final vein image I fu ; For the aligned transformed images under each view perform clarity and quality assessment. The quality assessment is calculated through image gradients to determine the quality of each image; The quality assessment formula is: ； In the formula, represents the quality index of the i-th perspective image, M*N represents the image size, represents the gradient of the image; Quality evaluation according to each perspective , calculate the weight value of each image ; The weight value is obtained by the following formula: ； Weight value is directly proportional to the clarity of the image, and the sum of all weights is 1, satisfying ; According to the weight value of each perspective image , perform pixel-level weighted fusion on the transformed image . For each pixel position , obtain the final vein image I fu ; The final vein image I fu is obtained by the following formula: I fu ； Wherein, I fu represents the pixel value of the final venous image at the position , and represents the value of the aligned image at this pixel position; The final vein image I obtained through weighted fusion fu Integrates vein information from multiple perspectives, reduces the occlusion and noise effects of a single perspective, and generates a clearer and more feature-rich vein image; Feature extraction module, for the final vein image I fu Perform feature extraction, and extract the feature vector F by using a convolutional neural network; input the final vein image I fu Into the convolutional layer of the convolutional neural network for a series of convolutional operations. The convolutional layer extracts local features in the image including edges, textures, and vein structures through different filters, and obtains the output of the convolutional layer: ； In the formula, represents the pixel value of the output feature map of the l-th convolutional layer in channel c, represents the convolutional kernel, represents the bias term, Na represents the size of the convolutional kernel, represents the movement of the convolutional window; Feature map output by the convolutional layer The high-frequency information in the feature vector is made more prominent by processing through the non-linear activation function ReLU; ; Subsequently, a pooling operation is performed to reduce the dimension of the feature map and retain the most prominent vein features. The feature map after pooling is ; After passing through multiple convolutional layers, activation layers, and pooling layers, the output pooled feature map will be , flattened into a one-dimensional vector, and input into a fully connected layer for feature mapping to generate the final high-dimensional feature vector F; The high-dimensional feature vector F is obtained through the following formula: F ； In the formula, respectively represent the weight value and bias term of the fully connected layer, represents the activation function Sigmoid; Perform regularization processing on the high-dimensional feature vector F and normalize it to unit length so that the feature vectors of different samples can be effectively compared in the subsequent recognition stage: Fn ; In the formula, represents the norm of the eigenvector; A model training and evaluation module that inputs the extracted feature vector F into the vein recognition model for model training to obtain the final recognition result Y.

Images