Vein recognition model training method and device

Abstract

The invention discloses a training method and device of a vein recognition model, and relates to the technical field of vein recognition, an acquired preset angle range is accurately set to be 0-90 degrees, and image resolution and illumination conditions at different angles are ensured to be consistent through a quasi vein recognition device. The vein images collected at each angle are ensured to have consistent quality, the problems of image noise and feature loss caused by illumination variation or inconsistent equipment resolution are reduced, the definition and stability of the images are improved, the collected vein images are processed by using a Gaussian filtering denoising algorithm, and the image quality is improved. And noise introduced by equipment or an external environment is removed. According to the method, the image can be smoothed through Gaussian filtering, the image noise is effectively reduced, the image quality is improved, the instability in the image collection process is reduced through multi-angle collection and consistent illumination conditions, and the collected images keep high consistency at different angles.

Description

Technical Field

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

[0002] Vein recognition, as a relatively advanced biometric technology, performs identity verification by analyzing vein images of parts such as the palm or fingers. The uniqueness of vein image recognition is that it uses the human body's internal veins as the basis for identification. These features are not easy to forge 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, which can easily lead to reduced recognition accuracy in the face of partial occlusion, angle changes, or poor image quality. For example, when the hand posture changes or the vein image is occluded, the single-view image may not capture enough vein features, thus affecting the performance of the recognition system. In addition, single-view images often cannot provide comprehensive information about veins, limiting 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 loss or error of vein features. In this case, the vein recognition model may be disturbed, resulting in a high misrecognition rate and false rejection rate. Summary of the invention

[0005] In view of 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 technology.

[0006] To achieve the above objectives, the present invention is implemented by the following technical scheme: a training method for a vein recognition model comprises the following steps:

[0007] Step 1: Collect vein images from multiple angles by using existing vein recognition equipment θi , ensuring the clarity of the veins at every angle and maintaining the continuity of the image. The acquired images are marked as I θ1 ,I θ2 ,...,I θn ;

[0008] Step 2: For each vein image I θi Preprocessing is performed, including denoising, contrast enhancement and normalization, to obtain the normalized vein image P(I θi );

[0009] Step 3: By using the affine transformation image registration technique, the preprocessed vein images P(I θ1 ),P(I θ2 ),...,P(I θn ) to align and obtain the aligned transformed image P′(I θi ), ensuring that vein features can be accurately matched at different viewing angles;

[0010] Step 4: Align the transformed image P′(I θi ) are combined through weighted fusion algorithm to generate the final vein image I fu ;

[0011] Step 5: Final vein image I fu Perform feature extraction by extracting feature vector F using convolutional neural network;

[0012] Step 6: Input the extracted feature vector F into the vein recognition model, train the model, and obtain the final recognition result Y. The model trained by fusion 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, the preset angles θ1, θ2, ..., θ collected are determined by the vein recognition device. n , select the angle range between 0 and 90 degrees, and use the quasi-vein recognition device to ensure that the resolution and lighting conditions of the image at different angles are consistent. After the device is calibrated, set the angle and the device will shoot at the preset angle;

[0014] At the preset angles θ1, θ2, ..., θ n The vein image is collected at each angle and the image collected is recorded. θ1 ,I θ2 ,...,I θn , by using the automated image acquisition mode, images are continuously acquired from multiple angles in a short time, reducing the impact of changes in vein shape and posture on image quality.

[0015] Preferably, the vein image I is denoised by using a Gaussian filter denoising algorithm. θi Perform denoising to remove image noise introduced by acquisition equipment and external environment, and obtain denoised vein image I′ θi ;

[0016] The denoised vein image I′ θi Obtained by the following formula:

[0017]

[0018] In the formula, I θi (x, y) represents the image pixel value, μx, μy represent the center position of the pixel, and σ represents the standard deviation of the Gaussian distribution;

[0019] Enhance the denoised vein image I′ by using histogram equalization θi The contrast of the vein structure is emphasized, the edge and texture of the vein image are highlighted, and the enhanced vein image I′ is obtained. θ ' i ;

[0020] The enhanced vein image I′ θ ' i Obtained by the following formula:

[0021]

[0022] In the formula, cdf(I′ θi (x,y)) represents the cumulative distribution function of the image pixel value, M*N represents the image size, L represents the maximum grayscale of the pixel value, cdf min represents the minimum non-zero value of the cumulative distribution function;

[0023] For the enhanced vein image I′ θ ' i Normalize and map the image pixel values ​​uniformly to the standard range to obtain the normalized vein image P(I θi ). Normalization can eliminate the image intensity differences caused by acquisition equipment or lighting conditions, ensuring that the image feature values ​​at different angles are in the same scale range, facilitating subsequent feature extraction and fusion.

[0024] Preferably, the normalized vein image P(I θi ) Extract key points P j , the key point P j Integrate and obtain the key point set K θi = {P1,P2,...,P m}, where each key point P j Include position (x j ,y j ) and descriptor d j ; Through key point detection and matching, the corresponding points of the vein structure in the image at different angles are found to provide a reference for the subsequent affine transformation calculation.

[0025] Normalized vein images P(I θi ), by performing key point matching, the corresponding point pairs between the two images are determined by using the nearest neighbor algorithm to obtain the matching points (P θi ,Pθj );

[0026] According to the matching point (P θi ,P θj ) Calculate the affine transformation matrix T by the least squares method θi→θj , the form of affine transformation is:

[0027]

[0028] Where (x, y) represents the normalized vein image P(I θi ), (x′, y′) represents the transformed vein image P(I θj ) in the coordinates, the affine transformation matrix T θi→θj The parameters a, b, c, d, e and f in the matching point (P θi ,P θj ) obtain;

[0029] For the normalized vein image P(I θi ) is affine transformed to convert it into the transformed vein image P(I θj ) in the aligned coordinate system, obtain the transformed image P′(I θi ).

[0030] Preferably, for each perspective, the aligned transformed image P′(I θi ) performs clarity and quality assessment, the quality assessment is calculated by image gradient to determine the quality of each image;

[0031] The quality assessment formula is:

[0032]

[0033] In the formula, Q θi represents the quality index of the i-th view image, M*N represents the image size, ▽P′(I θi )(x,y) represents the gradient of the image;

[0034] According to the quality evaluation Q of each view θi , calculate the weight value ω for each image i ;

[0035] The weight value ω i Obtained by the following formula:

[0036]

[0037] Weight value ω i It is proportional to the clarity of the image, and the sum of all weights is 1, satisfying

[0038] According to the weight value ω of each perspective image i , for the transformed image P′(I θi ) performs pixel-level weighted fusion, and for each pixel position (x, y), obtains the final vein image I fu ;

[0039] The final vein image I fu Obtained by the following formula:

[0040]

[0041] In the formula, I fu (x, y) represents the pixel value of the final vein image at position (x, y), Q[P′(I θi )(x,y)] represents the value of the aligned image at that pixel position;

[0042] The final vein image I obtained by weighted fusion fu The vein information from multiple perspectives is integrated to reduce the occlusion and noise effects of a single perspective, generating clearer and more feature-rich vein images.

[0043] Preferably, the final vein image I fu Input the convolutional layer of the convolutional neural network and perform a series of convolution 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:

[0044]

[0045] In the formula, F (l) (x, y, c) represents the pixel value of the output feature map of the lth convolutional layer in channel c, W (l) represents the convolution kernel, b (l) represents the bias term, K represents the size of the convolution kernel, Δx and Δy represent the movement of the convolution window;

[0046] The feature map F output by the convolutional layer (l) The high-frequency information in the feature vector is more prominent after being processed by the nonlinear activation function ReLU;

[0047] Then 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

[0048] Preferably, after multiple convolutional layers, activation layers and pooling layers, the output pooled feature map is Flatten it into a one-dimensional vector and input it 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 by the following formula:

[0050]

[0051] Where W (fc) and b (fc) Represent the weight value and bias term of the fully connected layer respectively, and σ 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]

[0054] Where ‖F‖ represents the norm of the eigenvector.

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

[0056] Using the cross entropy loss function To measure the output prediction value y of the model and the actual label The difference between

[0057]

[0058] In the formula, Represents the one-hot encoding of the true label, y k Represents the probability that the model output belongs to category k; the smaller the cross entropy loss, the closer the model's prediction is to the true label;

[0059] According to the loss function calculation results, the gradient of the loss to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using SGD. The model gradually reduces the loss function value through multiple iterative learning.

[0060]

[0061] In the formula, θ represents the model parameters, η represents the learning rate, Represents the gradient of the loss function with respect to the model parameters.

[0062] Preferably, during the training process, a portion of the validation set data that is not involved in the training is used to input the model for forward propagation to obtain the final recognition result Y of the model, and the loss L of the validation set is calculated. val And the accuracy Acc val , the loss of the validation set is Lval And the accuracy Acc val Used to monitor model overfitting and adjust through early stopping and model regularization;

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

[0064]

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

[0066]

[0067] Where N val represents the number of samples in the validation set, Represents the model output y i Is it consistent with the true label Consistency;

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

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

[0070] The multi-view image acquisition module uses the existing vein recognition equipment to collect vein images from multiple angles. θi , ensuring the clarity of the veins at every angle and maintaining the continuity of the image. The acquired images are marked as I θ1 ,I θ2 ,...,I θn ;

[0071] Image preprocessing module, for each angle of the vein image I θi Preprocessing is performed, including denoising, contrast enhancement and normalization, to obtain the normalized vein image P(I θi );

[0072] The image alignment and correction module uses the affine transformation image registration technology to align the preprocessed vein images P(I θ1 ),P(I θ2 ),...,P(I θn ) to align and obtain the aligned transformed image P′(I θi ), ensuring that vein features can be accurately matched at different viewing angles;

[0073] The multi-view image fusion module transforms the aligned transformed image P′(I θi) are combined through weighted fusion algorithm to generate the final vein image I fu ;

[0074] Feature extraction module, for the final vein image I fu Perform feature extraction by extracting feature vector F using convolutional neural network;

[0075] The model training and evaluation module inputs the extracted feature vector F into the vein recognition model, trains the model, and obtains the final recognition result Y.

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

[0077] (1) The preset angle range of the acquisition is precisely set to 0 to 90 degrees, and the quasi-vein recognition device is used to ensure that the image resolution and lighting conditions at different angles are consistent. This ensures that the vein images acquired at each angle have consistent quality, reduces image noise and feature loss problems caused by changes in lighting or inconsistent device resolution, thereby improving image clarity and stability. The collected vein images are processed using a Gaussian filter denoising algorithm to remove noise introduced by the device or external environment. Gaussian filtering can smooth the image, reduce noise interference, and retain important vein features. This method effectively reduces image noise, improves image quality, makes the vein structure more obvious, and reduces the risk of misidentification due to noise.

[0078] Through multi-angle acquisition and consistent lighting conditions, the instability in the image acquisition process is reduced, so that the acquired images maintain a high degree of 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 matched, reducing the possibility of mismatching, especially when the vein features are rotated, scaled, and other transformations, which greatly improves the accuracy of image alignment. Affine transformation can effectively correct the geometric distortion caused by different shooting angles, ensuring the consistency of the geometric structure of the vein image, thereby reducing errors in the image fusion process and enhancing the robustness of vein recognition. Through quality assessment, high-quality images can be automatically selected and weighted to avoid the negative impact of low-quality images on the final recognition results. The effectiveness of image fusion is improved, making the final generated vein image clearer and richer in features. Pixel-level weighted fusion ensures that the value of each pixel is reasonably weighed in multiple perspective images, and the generated vein image I fu It is clearer and contains feature information from multiple angles.

[0080] (3) The convolutional layer output is processed nonlinearly using the ReLU activation function to highlight the high-frequency information in the feature vector. Next, the pooling operation reduces the redundant information of the feature map by reducing the dimension, retains the most significant vein features, and reduces the amount of calculation. The ReLU activation function can effectively improve the model's sensitivity to high-frequency vein features and increase nonlinear expression capabilities, while the pooling operation reduces the amount of calculation while maintaining key features. By reducing redundant information after pooling, the model can perform recognition more efficiently and is not easily disturbed 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 the vein image. By regularizing the feature vector, the impact of feature vector length differences 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 and significantly improve recognition accuracy. By fusing vein images from each perspective, more comprehensive and detailed vein features can be captured, reducing recognition errors caused by a single perspective.

[0082] The fused vein image can effectively reduce the impact of partial occlusion, posture changes and image quality issues on the recognition results. The system's adaptability to different viewing angles and occlusion conditions is significantly enhanced, which improves the stability and reliability of the model in practical applications. By fusing the aligned multi-view images, the final vein image generated contains more information and features, allowing the feature extraction process to extract richer and more accurate vein features. This is critical for training an efficient vein recognition model and helps improve the final recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 Schematic diagram of the training method steps of the vein recognition model of the present invention;

[0084] Figure 2 The figure is a flowchart diagram of the training device for the vein recognition model of the present invention. DETAILED DESCRIPTION

[0085] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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 ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0086] Example 1

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

[0088] Step 1: Collect vein images from multiple angles by using existing vein recognition equipment θi , ensuring the clarity of the veins at every angle and maintaining the continuity of the image. The acquired images are marked as I θ1 ,I θ2 ,...,I θn ;

[0089] Step 2: For each vein image I θi Preprocessing is performed, including denoising, contrast enhancement and normalization, to obtain the normalized vein image P(I θi );

[0090] Step 3: By using the affine transformation image registration technique, the preprocessed vein images P(I θ1 ),P(I θ2 ),...,P(I θn ) to align and obtain the aligned transformed image P′(I θi );

[0091] Step 4: Align the transformed image P′(I θi ) are combined through weighted fusion algorithm to generate the final vein image I fu ;

[0092] Step 5: Final vein image I fu Perform feature extraction by extracting feature vector F using convolutional neural network;

[0093] Step 6: Input the extracted feature vector F into the vein recognition model, train the model, and obtain the final recognition result Y.

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

[0095] The device collects images at multiple fixed angles to ensure the recognizability of the venous structure in each image and reduce image loss or distortion caused by angle changes or posture problems.

[0096] The pre-processed vein images of different perspectives are aligned through affine transformation image registration technology. The aligned images ensure that the vein features under each perspective can accurately correspond to each other, reducing the feature deviation caused by the difference in perspective. The affine transformation algorithm is applied to adjust the position, rotation and scaling of the image so that the vein images under each perspective are aligned in the same coordinate system, optimizing the image registration accuracy. The aligned transformed images are merged through the weighted fusion algorithm to generate the final vein image I fu ,The fused image can integrate the features of multiple perspectives, reduce noise and occlusion problems, and improve the effect of feature extraction.

[0097] The weighted fusion algorithm is used to perform weighted averaging of multiple aligned images according to the weights to obtain a vein image that integrates multi-view features. fu Perform feature extraction and use a convolutional neural network to extract a high-dimensional feature vector F. The extracted feature vector can accurately characterize the vein structure in the vein image and provide rich feature information for model training. The extracted feature vector F is input into the vein recognition model for training to obtain the final recognition result Y. Model training can optimize the recognition performance, making the model more accurate in classifying and recognizing vein images. The feature vector F is input into the vein recognition model, and the model parameters are adjusted through back propagation and optimization algorithms so that the final output recognition result Y can accurately match the vein features.

[0098] Example 2

[0099] This embodiment is explained in Example 1, please refer to Figure 1 Specifically: the preset angles θ1, θ2, ..., θ are determined by the vein recognition device. n , select the angle range between 0 and 90 degrees, and use the quasi-vein recognition device to ensure that the resolution and lighting conditions of the image at different angles are consistent. After the device is calibrated, set the angle and the device will shoot at the preset angle;

[0100] At the preset angles θ1, θ2, ..., θ n The vein image is collected at each angle and the image collected is recorded. θ1 ,I θ2 ,...,I θn , by using automated image acquisition mode, images are continuously acquired from multiple angles in a short time.

[0101] By using Gaussian filtering denoising algorithm to denoise the vein image θi Perform denoising to remove image noise introduced by acquisition equipment and external environment, and obtain denoised vein image I′ θi ;

[0102] The denoised vein image I′ θi Obtained by the following formula:

[0103]

[0104] In the formula, I θi (x, y) represents the image pixel value, μx, μy represent the center position of the pixel, and σ represents the standard deviation of the Gaussian distribution;

[0105] Enhance the denoised vein image I′ by using histogram equalization θi The contrast of the vein structure is emphasized, the edge and texture of the vein image are highlighted, and the enhanced vein image I′ is obtained. θ ' i ;

[0106] The enhanced vein image I′ θ ' i Obtained by the following formula:

[0107]

[0108] In the formula, cdf(I′ θi (x,y)) represents the cumulative distribution function of the image pixel value, M*N represents the image size, L represents the maximum grayscale of the pixel value, cdf min represents the minimum non-zero value of the cumulative distribution function;

[0109] For the enhanced vein image I′ θ ' i Normalize and map the image pixel values ​​uniformly to the standard range to obtain the normalized vein image P(I θi ).

[0110] In this embodiment, the preset angle range of the acquisition is precisely set to 0 degrees to 90 degrees, and the image resolution and lighting conditions at different angles are guaranteed to be consistent through the quasi-vein recognition device. It ensures that the vein images acquired at each angle have consistent quality, reduces the image noise and feature loss problems caused by changes in lighting or inconsistent device resolution, thereby improving the clarity and stability of the image, and uses the Gaussian filter denoising algorithm to process the acquired vein images to remove the noise introduced by the device or external environment. Gaussian filtering can smooth the image, reduce noise interference, and retain important vein features. This method effectively reduces image noise, improves image quality, makes the vein structure more obvious, and reduces the risk of misidentification due to noise.

[0111] Through multi-angle acquisition and consistent lighting conditions, the instability in the image acquisition process is reduced, so that the acquired images maintain a high degree of 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 accuracy of recognition. Gaussian filtering and normalization help eliminate image quality issues caused by device noise and ambient lighting, allowing the model to be based on cleaner, standardized image data during training and inference. This reduces the complexity of model training and reduces the misrecognition rate caused by differences in image quality.

[0113] Example 3

[0114] This embodiment is explained in Example 2. Please refer to Figure 1 Specifically: By using the feature point detection algorithm to normalize the vein image P(I θi ) Extract key points P j , the key point P j Integrate and obtain the key point set K θi = {P1,P2,...,P m}, where each key point P j Include position (x j ,y j ) and descriptor d j ;

[0115] Normalized vein images P(I θi ), by performing key point matching, the corresponding point pairs between the two images are determined by using the nearest neighbor algorithm to obtain the matching points (P θi ,P θj );

[0116] According to the matching point (P θi ,P θj ) Calculate the affine transformation matrix T by the least squares method θi→θj , the form of affine transformation is:

[0117]

[0118] Where (x, y) represents the normalized vein image P(I θi ), (x′, y′) represents the transformed vein image P(I θj ) in the coordinates, the affine transformation matrix T θi→θj The parameters a, b, c, d, e and f in the matching point (Pθi ,P θj ) obtain;

[0119] For the normalized vein image P(I θi ) is affine transformed to convert it into the transformed vein image P(I θj ) in the aligned coordinate system, obtain the transformed image P′(I θi ).

[0120] For each perspective, the aligned transformed image P′(I θi ) performs clarity and quality assessment, the quality assessment is calculated by image gradient to determine the quality of each image;

[0121] The quality assessment formula is:

[0122]

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

[0124] According to the quality evaluation Q of each view θi , calculate the weight value ω for each image i ;

[0125] The weight value ω i Obtained by the following formula:

[0126]

[0127] Weight value ω i It is proportional to the clarity of the image, and the sum of all weights is 1, satisfying

[0128] According to the weight value ω of each perspective image i , for the transformed image P′(I θi ) performs pixel-level weighted fusion, and for each pixel position (x, y), obtains the final vein image I fu ;

[0129] The final vein image I fu Obtained by the following formula:

[0130]

[0131] In the formula, I fu (x, y) represents the pixel value of the final vein image at position (x, y), Q[P′(I θi )(x,y)] represents the value of the aligned image at that pixel position;

[0132] The final vein image I obtained by weighted fusion fu The vein information from multiple perspectives is integrated to reduce the occlusion and noise effects of a single perspective, generating clearer and more feature-rich vein images.

[0133] In this embodiment, through accurate feature point matching, vein images of different perspectives can be efficiently matched, reducing the possibility of mismatching, especially when the vein features are rotated, scaled, etc., which greatly improves the accuracy of image alignment. Affine transformation can effectively correct the geometric distortion caused by different shooting angles, ensure the consistency of the geometric structure of the vein image, thereby reducing the error in the image fusion process and enhancing the robustness of vein recognition. Through quality evaluation, high-quality images can be automatically selected and weighted to avoid the negative impact of low-quality images on the final recognition results. The effectiveness of image fusion is improved, making the final generated vein image clearer and richer in features. Pixel-level weighted fusion ensures that the value of each pixel is reasonably weighed in multiple perspective images, and the generated vein image I fu It is clearer and contains feature information from multiple angles. Through the fused images, the integrity of the vein structure is enhanced, which improves the accuracy and robustness of the vein recognition model.

[0134] By evaluating the quality of vein images at different viewing angles and performing weighted fusion based on the evaluation results, the negative impact of low-quality images on the final recognition results can be minimized, thereby obtaining a clearer vein image.

[0135] The weighted mechanism of multi-view image fusion ensures that information from multiple angles can be integrated, reducing the occlusion or noise problems caused by a single angle, and the generated image is more complete and richer in details. This significantly improves the accuracy and reliability of the model during vein recognition. Since the fused image contains vein information from multiple angles, the final generated vein image I fu Clearer and more feature-rich. This makes the vein recognition model more robust in dealing with different angles, lighting changes, and partial occlusion, significantly improving the recognition accuracy of the system.

[0136] Example 4

[0137] This embodiment is explained in Example 3, please refer to Figure 1 Specifically: the final vein image I fu Input the convolutional layer of the convolutional neural network and perform a series of convolution 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, F (l) (x, y, c) represents the pixel value of the output feature map of the lth convolutional layer in channel c, W (l) represents the convolution kernel, b (l) represents the bias term, K represents the size of the convolution kernel, Δx and Δy represent the movement of the convolution window;

[0140] The feature map F output by the convolutional layer (l) The high-frequency information in the feature vector is more prominent after being processed by the nonlinear activation function ReLU;

[0141] Then 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

[0142] The final vein image I is trained using a convolutional neural network fu After performing feature extraction and multiple layers of convolution operations, the network can capture local features in the image, including edges, textures, and vein structures. Through the operation of different convolution 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 richer and more efficient vein features from the image, reducing the reliance on manually designed features. The layer-by-layer extraction of convolution layers enables the detailed features of veins at different levels to be accurately captured, which helps to improve the accuracy and robustness of vein recognition.

[0143] After multiple convolutional layers, activation layers, and pooling layers, the output pooled feature map is Flatten it into a one-dimensional vector and input it 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 by the following formula:

[0145]

[0146] Where W (fc) and b (fc) Represent the weight value and bias term of the fully connected layer respectively, and σ represents the activation function Sigmoid;

[0147] 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:

[0148]

[0149] Where ‖F‖ represents the norm of the eigenvector.

[0150] The convolutional layer output is processed nonlinearly using the ReLU activation function to highlight the high-frequency information in the feature vector. Next, the pooling operation reduces the redundant information of the feature map by reducing the dimension, retains the most significant vein features, and reduces the amount of calculation. The ReLU activation function can effectively improve the model's sensitivity to high-frequency vein features and increase nonlinear expression capabilities, while the pooling operation reduces the amount of calculation while maintaining key features. By reducing redundant information after pooling, the model can perform recognition more efficiently and is not easily disturbed by noise and irrelevant information.

[0151] The feature map after the convolution 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 convolution layer into global features, which enhances the model's understanding of the overall pattern of the vein image. By regularizing the feature vector, the impact of the difference in feature vector length on the 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 M(.), and after 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 by the following formula: y = M(F);

[0153] Using the cross entropy loss function To measure the output prediction value y of the model and the actual label The difference between

[0154]

[0155] In the formula, Represents the one-hot encoding of the true label, y k Represents the probability that the model output belongs to category k; the smaller the cross entropy loss, the closer the model's prediction is to the true label;

[0156] According to the loss function calculation results, the gradient of the loss to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using SGD. The model gradually reduces the loss function value through multiple iterative learning.

[0157]

[0158] In the formula, θ represents the model parameters, η 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 was not involved in the training is used to input the model for forward propagation to obtain the final recognition result Y of the model and calculate the loss L of the validation set. val And the accuracy Acc val , the loss of the validation set is L val And the accuracy Acc val Used to monitor model overfitting and adjust through early stopping and model regularization;

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

[0161]

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

[0163]

[0164] Where N val represents the number of samples in the validation set, Represents the model output y i Is it consistent with the true label Consistency;

[0165] After training is completed, use the independent test set data to perform the final model evaluation to obtain the performance of the model on the test set, including the test set accuracy Acc val , 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 to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using SGD. This iterative optimization method can gradually reduce the loss function value and improve the recognition accuracy of the model.

[0167] The cross entropy loss function is very effective for classification tasks and can accelerate the convergence of the model. At the same time, the back propagation algorithm combined with the SGD parameter update method can optimize the model performance in each iteration, gradually approach the optimal parameter combination, and improve 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 overfitting of the model through early stopping and regularization, thereby improving the generalization ability of the model.

[0168] Using the validation set can monitor the performance of the model on unseen data in real time, preventing the model from overfitting 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 vein recognition model, please refer to Figure 2 Specifically: the training device includes:

[0171] The multi-view image acquisition module uses the existing vein recognition equipment to collect vein images from multiple angles. θi , ensuring the clarity of the veins at every angle and maintaining the continuity of the image. The acquired images are marked as I θ1 ,I θ2 ,...,I θn ;

[0172] Image preprocessing module, for each angle of the vein image I θi Preprocessing is performed, including denoising, contrast enhancement and normalization, to obtain the normalized vein image P(I θi );

[0173] The image alignment and correction module uses the affine transformation image registration technology to align the preprocessed vein images P(I θ1 ),P(I θ2 ),...,P(I θn ) to align and obtain the aligned transformed image P′(I θi );

[0174] The multi-view image fusion module transforms the aligned transformed image P′(I θi ) are combined through weighted fusion algorithm to generate the final vein image I fu ;

[0175] Feature extraction module, for the final vein image I fu Perform feature extraction by extracting feature vector F using convolutional neural network;

[0176] The model training and evaluation module inputs the extracted feature vector F into the vein recognition model, trains the model, and obtains the final recognition result Y.

[0177] In this embodiment, through multi-view image fusion, the vein recognition system can integrate information from multiple angles and significantly improve the accuracy of recognition. By fusing the vein images at each perspective, more comprehensive and detailed vein features can be captured, reducing the recognition error caused by a single perspective.

[0178] The fused vein image can effectively reduce the impact of partial occlusion, posture change and image quality issues on the recognition results. The system's adaptability to different viewing angles and occlusion conditions is significantly enhanced, which improves the stability and reliability of the model in practical applications.

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

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

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

Claims

1. A method for training a vein recognition model, characterized in that: The following steps are involved: Step 1: Collect vein images from multiple angles by using existing vein recognition equipment θi , ensuring the clarity of the veins at every angle and maintaining the continuity of the image. The collected images are marked as I θ1 ,I θ2 ,...,I θn ; Step 2: For each vein image I θi Preprocessing is performed, including denoising, contrast enhancement and normalization, to obtain the normalized vein image P(I θi ); Step 3: By using the affine transformation image registration technique, the preprocessed vein images P(I θ1 ),P(I θ2 ),...,P(I θn ) to align and obtain the aligned transformed image P′(I θi ); Step 4: Align the transformed image P′(I θi ) are combined through weighted fusion algorithm to generate the final vein image I fu ; Step 5: Final vein image I fu Perform feature extraction by extracting feature vector F using convolutional neural network; Step 6: Input the extracted feature vector F into the vein recognition model, train the model, and obtain the final recognition result Y.

2. The method for training a vein recognition model according to claim 1, characterized in that: The preset angles θ1, θ2, ..., θ are determined by the vein recognition device. n , select the angle range between 0 and 90 degrees, and use the quasi-vein recognition device to ensure that the resolution and lighting conditions of the image at different angles are consistent. After the device is calibrated, set the angle and the device will shoot at the preset angle; At the preset angles θ1, θ2, ..., θ n The vein image is collected at each angle and the image collected is recorded. θ1 ,I θ2 ,...,I θn , by using automated image acquisition mode, images are continuously acquired from multiple angles in a short time.

3. The method for training a vein recognition model according to claim 1, characterized in that: By using Gaussian filtering denoising algorithm to denoise the vein image θi Perform denoising to remove image noise introduced by acquisition equipment and external environment, and obtain denoised vein image I′ θi ; The denoised vein image I′ θi Obtained by the following formula: In the formula, I θi (x, y) represents the image pixel value, μx, μy represent the center position of the pixel, and σ represents the standard deviation of the Gaussian distribution; Enhance the denoised vein image I′ by using histogram equalization θi The contrast of the vein structure is emphasized, the edge and texture of the vein image are highlighted, and the enhanced vein image I′ is obtained. θ ' i ; The enhanced vein image I′ θ ' i Obtained by the following formula: In the formula, cdf(I′ θi (x,y)) represents the cumulative distribution function of the image pixel value, M*N represents the image size, L represents the maximum grayscale of the pixel value, cdf min represents the minimum non-zero value of the cumulative distribution function; For the enhanced vein image I′ θ ' i Normalize and map the image pixel values ​​uniformly to the standard range to obtain the normalized vein image P(I θi ).

4. The method for training a vein recognition model according to claim 1, characterized in that: By using the feature point detection algorithm to normalize the vein image P(I θi ) Extract key points P j , the key point P j Integrate and obtain the key point set K θi = {P1,P2,...,P m }, where each key point P j Include position (x j ,y j ) and descriptor d j ; Normalized vein images P(I θi ), by performing key point matching, the corresponding point pairs between the two images are determined by using the nearest neighbor algorithm to obtain the matching points (P θi ,P θj ); According to the matching point (P θi ,P θj ) Calculate the affine transformation matrix T by the least squares method θi→θj , the form of affine transformation is: Where (x, y) represents the normalized vein image P(I θi ), (x′, y′) represents the transformed vein image P(I θj ) in the coordinates, the affine transformation matrix T θi→θj The parameters a, b, c, d, e and f in the matching point (P θi ,P θj ) obtain; For the normalized vein image P(I θi ) is affine transformed to convert it into the transformed vein image P(I θj ) in the aligned coordinate system, obtain the transformed image P′(I θi ).

5. The method for training a vein recognition model according to claim 4, characterized in that: For each perspective, the aligned transformed image P′(I θi ) performs clarity and quality assessment, the quality assessment is calculated by image gradient to determine the quality of each image; The quality assessment formula is: In the formula, Q θi represents the quality index of the i-th view image, M*N represents the image size, Represents the gradient of the image; According to the quality evaluation Q of each view θi , calculate the weight value ω for each image i ; The weight value ω i Obtained by the following formula: Weight value ω i It is proportional to the clarity of the image, and the sum of all weights is 1, satisfying According to the weight value ω of each perspective image i , for the transformed image P′(I θi ) performs pixel-level weighted fusion, and for each pixel position (x, y), obtains the final vein image I fu ; The final vein image I fu Obtained by the following formula: In the formula, I fu (x, y) represents the pixel value of the final vein image at position (x, y), Q[P′(I θi )(x,y)] represents the value of the aligned image at that pixel position; The final vein image I obtained by weighted fusion fu The vein information from multiple perspectives is integrated to reduce the occlusion and noise effects of a single perspective, generating clearer and more feature-rich vein images.

6. The method for training a vein recognition model according to claim 5, characterized in that: The final vein image I fu Input the convolutional layer of the convolutional neural network and perform a series of convolution 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, F (l) (x, y, c) represents the pixel value of the output feature map of the lth convolutional layer in channel c, W (l) represents the convolution kernel, b (l) represents the bias term, K represents the size of the convolution kernel, Δx and Δy represent the movement of the convolution window; The feature map F output by the convolutional layer (l) The high-frequency information in the feature vector is more prominent after being processed by the nonlinear activation function ReLU; Then 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 7. The method for training a vein recognition model according to claim 6, characterized in that: After multiple convolutional layers, activation layers, and pooling layers, the output pooled feature map is Flatten it into a one-dimensional vector and input it into the fully connected layer for feature mapping to generate the final high-dimensional feature vector F; The high-dimensional feature vector F is obtained by the following formula: Where W (fc) and b (fc) Represent the weight value and bias term of the fully connected layer respectively, and σ represents the activation function Sigmoid; 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: Where ‖F‖ represents the norm of the eigenvector.

8. The method for training a vein recognition model according to claim 7, characterized in that: The extracted high-dimensional feature vector F is used as the input of the vein recognition model M(.), and after 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 by the following formula: y = M(F); Using the cross entropy loss function To measure the output prediction value y of the model and the actual label The difference between In the formula, Represents the one-hot encoding of the true label, y k Represents the probability that the model output belongs to category k; the smaller the cross entropy loss, the closer the model's prediction is to the true label; According to the loss function calculation results, the gradient of the loss to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using SGD. The model gradually reduces the loss function value through multiple iterative learning. In the formula, θ represents the model parameters, η represents the learning rate, Represents the gradient of the loss function with respect to the model parameters.

9. The method for training a vein recognition model according to claim 8, characterized in that: During the training process, a part of the validation set data that was not involved in the training is used to input the model for forward propagation to obtain the final recognition result Y of the model and calculate the loss L of the validation set. val And the accuracy Acc val , the loss of the validation set is L val And the accuracy Acc val Used to monitor model overfitting and adjust through early stopping and model regularization; The calculation of validation set loss is the same as training loss; The formula for the validation set accuracy is: Where N val represents the number of samples in the validation set, Represents the model output y i Is it consistent with the true label Consistency; After training is completed, use the independent test set data to perform the final model evaluation to obtain the performance of the model on the test set, including the test set accuracy Acc val , confusion matrix, precision, recall and F1 score metrics.

10. A training device for a vein recognition model, characterized in that: The training device comprises: The multi-view image acquisition module uses the existing vein recognition equipment to collect vein images from multiple angles. θi , ensuring the clarity of the veins at every angle and maintaining the continuity of the image. The collected images are marked as I θ1 ,I θ2 ,...,I θn ; Image preprocessing module, for each angle of the vein image I θi Preprocessing is performed, including denoising, contrast enhancement and normalization, to obtain the normalized vein image P(I θi ); The image alignment and correction module uses the affine transformation image registration technology to align the preprocessed vein images P(I θ1 ),P(I θ2 ),...,P(I θn ) to align and obtain the aligned transformed image P′(I θi ); The multi-view image fusion module transforms the aligned transformed image P′(I θi ) are combined through weighted fusion algorithm to generate the final vein image I fu ; Feature extraction module, for the final vein image I fu Perform feature extraction by extracting feature vector F using convolutional neural network; The model training and evaluation module inputs the extracted feature vector F into the vein recognition model, trains the model, and obtains the final recognition result Y.

Images