Deep learning based k-nearest neighbor graph iterative vein recognition method and system
The deep learning-based K-nearest neighbor graph iterative vein recognition method utilizes the feature vector distance and correlation information of the vein image training set, solving the problem that existing models fail to fully utilize this information, and achieving higher vein recognition accuracy and faster computation speed.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING FINTECH INSTITUTE
- Filing Date
- 2022-12-21
- Publication Date
- 2026-05-01
AI Technical Summary
Existing deep learning models fail to fully utilize the feature vector distance and correlation information between samples in vein recognition, resulting in low accuracy in predicted labels.
A deep learning-based K-nearest neighbor graph iterative vein recognition method is adopted. By calculating the distance between the vein image to be tested and the feature vectors of the training set, the K-nearest neighbor feature vectors are selected, a similarity matrix and an iterative propagation matrix are established, the correlation between feature vectors is learned, and the label information of the vein image is predicted.
It improves the classification accuracy of vein recognition, reduces storage space requirements, increases processing speed, and simplifies the testing process.
Smart Images
Figure CN116092138B_ABST
Abstract
Description
A Deep Learning-Based K-Nearest Neighbor Graph Iterative Vein Recognition Method and System Technical Field
[0001] This invention belongs to the field of biometric recognition technology, specifically relating to a deep learning-based K-nearest neighbor graph iterative vein recognition method and system. Background Technology
[0002] Current deep learning models for vein recognition only input one training or test sample at a time. Therefore, these models struggle to learn the correlations between samples, which are often helpful for vein feature classification. These deep learning models aim to learn the transformation model between samples and labels. However, in the label prediction process, these methods do not fully utilize the distance and correlation information between the feature vectors corresponding to the samples, resulting in low label prediction accuracy and consequently affecting the accuracy of vein recognition. Summary of the Invention
[0003] The present invention aims to propose a deep learning-based K-nearest neighbor graph iterative vein recognition method to solve the technical problem that existing technologies do not fully utilize the distance and correlation information between the feature vectors of corresponding samples, resulting in low accuracy of predicted labels.
[0004] The deep learning-based K-nearest neighbor graph iterative vein recognition method of this invention includes:
[0005] Step 1) Obtain the feature vector of the vein image to be tested and the feature vector set of the vein image training set. Each vein image training sample in the vein image training set has a corresponding label information.
[0006] Step 2) Calculate the distance between the vein image to be tested and the feature vector and each feature vector in the feature vector set, and select k feature vectors from the feature vector set that have the smallest distance to the feature vectors of the vein image to be tested;
[0007] Step 3) Input the selected k+1 feature vectors and the vein image to be tested, as well as the label information of the vein image training samples corresponding to the selected k feature vectors, into the graph iterative model. Iteratively learn the correlation between the feature vectors and predict the label information of the vein image to be tested, thereby achieving the final prediction of the classification and recognition of the vein image to be tested.
[0008] Furthermore, step 3) includes establishing a similarity matrix A = [a...] based on the pairwise similarity between the selected and the k+1 feature vectors of the vein image to be tested. mn ] (k+1 ) ×(k+1) , where a mn For the eigenvector z mWith eigenvector z n Similarity;
[0009] The iterative propagation matrix D is calculated based on the similarity matrix A as follows:
[0010]
[0011] The diagonal matrix of Λ is: Λ=Diag[b1,b2,...b n ,... xk+1 ],
[0012] The label information of each vein image sample in the vein image training set is represented by a C×1 one-hot label vector, where C is the number of categories of the vein image sample. A matrix is constructed using the one-hot label vectors of the selected k vein image training samples. Let the label vector of the vein image to be tested be... Initially, it is a vector of all zeros. and Combined into the initial label propagation matrix
[0013] The label propagation matrix is iteratively updated in the following manner:
[0014] H t =μDH t-1 +(1-μ)H 0 ,
[0015] Where μ is a hyperparameter and t is the number of iterations;
[0016] Alternatively, the label propagation matrix H after multiple iterations can be obtained directly in the following manner. t :
[0017]
[0018] The label propagation matrix H obtained after multiple iterations and updates t The vector corresponding to the all-zero vector before the update The part is used as the prediction result of the label vector of the still image to be tested, and the classification and recognition result of the label vector of the still image to be tested is obtained based on the prediction result.
[0019] Furthermore, in step 3), the final label propagation matrix H is obtained in the following manner. Δ :
[0020] H Δ = (1-μ)(E-μD) -1 H 0 ,
[0021] And the label propagation matrix H Δ The vector corresponding to the all-zero vector before the update The predicted label vector of the still image to be tested is used as the part of the prediction result. Based on the predicted label vector, the classification result of the label vector of the still image to be tested is obtained, and then the identification of the label of the still image to be tested is completed.
[0022] Furthermore, similarity a mn It is calculated as follows:
[0023]
[0024] Furthermore, the distance mentioned in step 2) is the Manhattan distance.
[0025] Furthermore, the Manhattan distance L(z) between feature vectors m , z n Calculated as follows:
[0026]
[0027] Wherein, the feature vector z m =[z m1 , z m2 ,…,z md ] T , z n =[z n1 , z n2 ,…,z nd ] T d represents the dimension of the feature vector.
[0028] Furthermore, the same deep learning model is used to extract the feature vector set of the vein image to be tested and the vein image training set mentioned in step 1).
[0029] Furthermore, the deep learning model is a ResNet-50, PV-CNN, or FVRAS-Net model.
[0030] Another objective of this invention is to provide a deep learning-based K-nearest neighbor graph iterative vein recognition system, comprising an encoder module, a recognition module, and a training sample module, wherein the recognition module is connected to both the encoder module and the training sample module.
[0031] The encoder module is used to extract the feature vector of the input vein image to be tested;
[0032] The training sample module is used to store the feature vector set of the vein image training sample set, as well as the label information of the vein image training sample that corresponds one-to-one with the feature vector in the feature vector set;
[0033] The recognition module is used to provide the classification and recognition results of the vein image to be tested according to the aforementioned method.
[0034] Furthermore, the training sample module is also connected to a labeling module, and the training sample module is also connected to an encoder module.
[0035] The encoder module is also used to extract the feature vector of the input vein image training sample;
[0036] The annotation module is used to provide label information for each vein image training sample according to the input;
[0037] The training sample module is also used to obtain the feature vectors of each vein image training sample from the encoder module and the label information of each vein image training sample from the annotation module, and save the feature vectors and label information of each vein image training sample in a one-to-one correspondence.
[0038] This invention proposes a novel vein recognition algorithm based on "copying learning." This method first fully utilizes the distance information between the feature vectors corresponding to sample images and the feature vectors of the vein image to be tested, identifying the K nearest neighbors of the vein image in the vein image training set. Then, through graph iteration, it fully leverages the correlation between these sample images and the feature vectors corresponding to the vein image to be tested, propagating the label information of the samples in the vein image training set multiple times to predict the label of the image to be tested, effectively improving the classification accuracy of vein images. When applied to a system, this method only requires storing the feature vectors and label information of the vein image training samples; no new model training is needed during testing. The classification result is obtained by calculating the correlation of feature vectors, reducing storage space and improving computational speed. Attached Figure Description
[0039] Figure 1 is a flowchart illustrating a deep learning-based K-nearest neighbor graph iterative vein recognition method according to an embodiment of the present invention.
[0040] Figure 2 is a schematic block diagram of a deep learning-based K-nearest neighbor graph iterative vein recognition system according to an embodiment of the present invention.
[0041] Figure 3 is a schematic block diagram of a deep learning-based K-nearest neighbor graph iterative vein recognition system according to another embodiment of the present invention.
[0042] Figure 4 is a schematic block diagram of a deep learning-based K-nearest neighbor graph iterative vein recognition system in another embodiment of the present invention. Detailed Implementation
[0043] The deep learning-based K-nearest neighbor graph iterative vein recognition method in one embodiment of the present invention is basically shown in Figure 1, and includes...
[0044] Step 1) Obtain the feature vector of the vein image to be tested and the feature vector set of the vein image training set. Each vein image training sample in the vein image training set has a corresponding label information, and the feature items in the feature vector set also have the same label information.
[0045] In this embodiment, an encoder module is used to obtain the feature vector set of the vein image to be tested and the vein image training set.
[0046] In a more specific embodiment, the encoder module extracts feature vectors from the vein image to be tested and the vein images in the training set based on a deep learning model. Specific models can be, but are not limited to, ResNet-50, PV-CNN, or FVRAS-Net models.
[0047] Let X = [x1, x2, ..., xn] s ] is the corresponding label Y = [y1, y2, ..., y s The vein image training set consists of s related vein image training samples. Specifically, x0 and y0 represent the vein image to be tested and its label information, respectively. The encoder module is denoted as φw, where W represents the trainable weights. Therefore, the feature vector z0 of the vein image to be tested and the feature vector set Z = [z1, z2, ..., zn] of the s vein image training samples are... s It can be calculated using formula (1):
[0048] z i =φw(x i ), i = 0, 1, 2, 3...s, (1)
[0049] Here, φw(·) can specifically be various deep learning models used to extract feature vectors from images, such as ResNet-50, PV-CNN, or FVRAS-Net models. After feature extraction is completed for all vein image training samples in the vein image training set, the extracted feature vectors form a feature vector set.
[0050] In this step, in order to classify the vein features in the vein image under test by utilizing the relevant information between the vein image training samples and the vein image under test, the vein image is mapped from the high-dimensional feature space to the low-dimensional feature space using formula (1) so that the feature space-based algorithm can be directly used for classification in the future.
[0051] Step 2) Calculate the distance between the vein image to be tested and the feature vector and each feature vector in the feature vector set, and use the K nearest neighbor method to select k feature vectors from the feature vector set that have the smallest distance to the feature vectors of the vein image to be tested.
[0052] This step is based on the K-Nearest Neighbor (KNN) method. Specifically, the feature vector z0 of the vein image to be tested x0 is obtained through feature extraction. Then, based on the feature vector z0 corresponding to each vein image training sample in the vein image training set... m The distance between the vein image and the feature vector z0 is used to find the k vein image training samples that are most similar to the vein image to be tested in the training set, forming a set N. k (x0) represents the vein image training samples corresponding to the k feature vectors with the smallest Manhattan distance from z0 in the feature vector set.
[0053] In this embodiment, the Manhattan distance is preferably used to calculate the distance between feature vectors.
[0054] The d-dimensional feature vectors of the two images are represented by z. m =[z m1 , z m2 ,…,z md ] T and z n =[z n1 , z n2 ,…,z nd ] T In this embodiment, the Manhattan distance L(z) between them m , z n ) Calculated using formula (2):
[0055]
[0056] In other embodiments, the use of other distances or distance calculation methods is not excluded, such as, but not limited to, Euclidean distance or Mahalanobis distance.
[0057] Step 3) Input the selected k+1 feature vectors and the vein image to be tested, as well as the label information of the vein image training samples corresponding to the selected k feature vectors, into the graph iterative model. Iteratively learn the correlation between the feature vectors and predict the label information of the vein image to be tested, thereby achieving the final prediction of the classification and recognition of the vein image to be tested.
[0058] In existing technologies, by analyzing set Nk Cluster analysis is performed on (x0) to determine the label information y0 of the vein image x0 to be tested. In this invention, the relationships between features are learned through graph iteration, narrowing down from global relationships to closely related local relationships, which not only reduces the amount of computation but also improves the testing accuracy.
[0059] In this embodiment, the feature vectors and label information of the k approximate images are used to construct relevant matrices, which are then input into the graph iterative model and propagated multiple times. In this way, the feature information and label information of the k approximate images can be used to predict the label information of the vein image to be tested, thereby achieving the purpose of classification.
[0060] Specifically, k+1 feature vectors are obtained from k approximate images and the vein image to be tested. The similarity between these feature vectors has an impact on the propagation process. Feature vector z m With z n similarity a mn It can be calculated using formula (3):
[0061]
[0062] The similarity calculation results between all k+1 feature vectors constitute the similarity matrix A = [a mn ] (k+1)×(k+1) The iterative propagation matrix D is calculated from the similarity matrix A using formula (4):
[0063]
[0064] The diagonal matrix of Λ is: Λ=Diag[b1,b2,...b n ...b k+1 ],
[0065] Given a training set of vein images with C categories, each vein image in the training set is labeled using a C×1 one-hot label vector. A matrix is constructed using the one-hot label vectors of the k nearest neighbor training images of the vein image to be tested. Let the label vector of the still image to be tested be... The initial vector is all zeros, thus obtaining the initialized label propagation matrix. Where || is the connector symbol.
[0066] From the aforementioned technical principles, those skilled in the art can clearly understand that each tag vector (including all-zero vectors) The position of the zero vector in H0 should correspond to the order of the corresponding feature vectors when calculating the similarity matrix; in this embodiment, for example, but not limited to, as shown in the "Label Concatenation" and "Feature Concatenation" sections of Figure 1, the zero vector After concatenating with the other k label vectors, the corresponding index (z-index) of the feature vector of the vein image to be tested in calculating the similarity matrix should be k+1. However, in some other embodiments, the all-zero vector can actually be... The label propagation matrix H is constructed by interpolating k label vectors. 0 The feature vectors of the vein image under test are naturally indices similar to those of the all-zero vectors when calculating the similarity matrix. The corresponding position in H0.
[0067] In this embodiment, the aforementioned process of inputting a total of k+1 feature vectors (including the selected vein image itself and the vein image to be tested) and the label information of the vein image training samples corresponding to the selected k feature vectors into the graph iterative model for iterative learning, i.e., iteratively learning the label propagation matrix H... 0 The process of iterative updates can be expressed as follows:
[0068] H t =μDH t-1 +(1-μ)H 0 (5)
[0069] Where μ is a hyperparameter, and H is the iterative propagation matrix with respect to the initial state. 0 The label propagation matrix in the final result H t It has an impact, where t is the number of iterations.
[0070] The label matrix H after multiple iterations t It can be represented as:
[0071]
[0072] Therefore, the zero vectors in the label matrix Hw obtained after iterative update can be used to represent the vectors that were all zero before the update. The portion is used as the prediction result of the label vector of the still image to be tested.
[0073] According to the principle of graph iterative model, the more iterations there are, the more accurate the prediction results will be. Therefore, when the number of iterations t approaches infinity, more accurate prediction results should be obtained.
[0074] Therefore, the label propagation matrix is first diagonally decomposed into H = Q. T Λ H Q, where Q is an orthogonal matrix, Λ HIt is a diagonal matrix containing the eigenvalues of H. Letting t approach infinity, we can further obtain formula (7):
[0075]
[0076] In addition, it can be considered that The proof is as follows:
[0077]
[0078] Therefore, after calculation, the limit of the second term in formula (6) can be expressed as follows:
[0079]
[0080] Therefore, according to formulas (6), (7) and (9), the label matrix H obtained when t approaches infinity is... Δ It can be represented as follows:
[0081] H Δ = (1-μ)(E-μD) -1 H0 (10)
[0082] Therefore, in some embodiments, the tag propagation matrix H in formula (10) can be directly used. Δ The vector corresponding to the all-zero vector before the update The predicted label vector of the still image to be tested is used as the prediction result. Then, the classification result of the label vector of the still image to be tested is obtained based on the predicted label vector, thereby completing the identification of the label of the still image to be tested. In these embodiments, iterative calculation is eliminated, further simplifying the calculation process, reducing storage space, and improving the operation speed.
[0083] After obtaining the predicted label vector of the still image to be tested, the classification result of the label vector of the still image to be tested can be obtained based on the prediction result, thereby completing the identification of the label of the still image to be tested. For example, the classification corresponding to the largest value in the predicted label vector is taken as the classification and identification result of the still image to be tested.
[0084] In other embodiments of the present invention, a deep learning-based K-nearest neighbor graph iterative vein recognition system is also provided. The system is basically as shown in Figure 2, including an encoder module, a recognition module and a training sample module, wherein the recognition module is connected to the encoder module and the training sample module respectively.
[0085] The encoder module is used to extract the feature vector of the input vein image to be tested. In a more specific embodiment, the encoder module is equipped with a pre-trained deep learning model for extracting image feature vectors, such as, but not limited to, Resnet-50, PV-CNN or FVRAS-Net models.
[0086] The training sample module is used to store the feature vector set of the vein image training sample set, as well as the label information of the vein image training sample that corresponds one-to-one with the feature vector in the feature vector set;
[0087] The recognition module is used to provide the classification and recognition results of the vein image to be tested according to the process in the aforementioned method embodiments.
[0088] In a further embodiment, as shown in FIG3, the training sample module is also connected to a labeling module, and the training sample module is also connected to the encoder module.
[0089] The encoder module is also used to extract feature vectors from the input vein image training samples;
[0090] The annotation module is used to provide label information for each vein image training sample based on the input;
[0091] The training sample module is also used to obtain the feature vectors of each vein image training sample from the encoder module and the label information of each vein image training sample from the annotation module, and save the feature vectors and label information of each vein image training sample in a one-to-one correspondence.
[0092] In some embodiments, as shown in FIG4, the identification module includes a K-nearest neighbor module and a graph iteration module.
[0093] The K-nearest neighbor module is used to calculate the distance between the vein image to be tested and the feature vector and each feature vector in the feature vector set; and to select k feature vectors from the feature vector set that have the smallest distance between them and the feature vectors of the vein image to be tested; in some embodiments, the distance between the feature vectors calculated in this module is preferably, but not limited to, the Manhattan distance, and the Manhattan distance is calculated by formula (2);
[0094] The graph iteration module is used to input a total of k+1 feature vectors of the selected vein image and the vein image itself, as well as the label information of the vein image training samples corresponding to the selected k feature vectors, into the graph iteration model. Iteratively, it learns the correlation between the feature vectors and predicts the label information of the vein image to be tested, thereby achieving the final prediction and classification of the vein image to be tested.
[0095] In some preferred embodiments, the label information of each vein image sample in the vein image training set is represented by a C×1 one-hot label vector, where C is the number of categories of the vein image sample;
[0096] In these embodiments, the graph iteration module is used to construct a similarity matrix A = [a...]. mn ] (k+1)×(k+1) , where a mn For the eigenvector z m With eigenvector z n The similarity, similarity a mn Preferred, but not limited to, calculations can be performed using the method shown in formula (3);
[0097] Furthermore, the iterative propagation matrix D is calculated based on the similarity matrix A using the method shown in formula (4);
[0098] Furthermore, a matrix is constructed using the one-hot label vectors of the selected k vein image training samples. Let the label vector of the vein image to be tested be... Initially, it is a vector of all zeros. and Combined into the initial label propagation matrix H 0 :
[0099] And, the label propagation matrix is iteratively updated in the manner shown in Equation (5);
[0100] And, the label propagation matrix H obtained after multiple iterations and updates. t The vector corresponding to the all-zero vector before the update The part is used as the prediction result of the label vector of the still image to be tested, and the classification and recognition result of the label vector of the still image to be tested is obtained based on the prediction result.
[0101] In other embodiments, the graph iteration module directly obtains the label propagation matrix H after multiple iterations as shown in Equation (6). t .
[0102] In a further embodiment, the graph iteration module obtains the final label propagation matrix H as shown in Equation (10). Δ And the label propagation matrix H Δ The vector corresponding to the all-zero vector before the update The predicted label vector of the still image to be tested is used as the part of the prediction result. Based on the predicted label vector, the classification result of the label vector of the still image to be tested is obtained, and then the identification of the label of the still image to be tested is completed.
[0103] Experimental Example
[0104] To verify the effectiveness of the K-nearest neighbor graph iterative vein recognition algorithm in this invention, we conducted validation using palm vein datasets provided by Hong Kong Polytechnic University and Tongji University. In the experiments, all algorithms were implemented on the PyTorch deep learning platform, and their performance was verified on a high-performance computer. The computer's hardware included an i7-7700K processor, 16GB of RAM, and an NVIDIA GTX 1080Ti graphics card.
[0105] Dataset A: The Hong Kong Polytechnic University palm vein dataset consists of palm vein images provided by 250 volunteers from their left and right hands. The sample data was collected in two phases, approximately 9 days apart. Six vein images were collected from each palm in each phase, for a total of 12 vein images per palm across the two phases. Ultimately, the 250 participants provided 6000 (250 participants × 2 palms × 6 images × 2 phases) palm vein images. Each palm vein image has a resolution of 200 × 200 pixels and 256 gray levels.
[0106] Dataset B: The Tongji University palm vein dataset consists of palm vein images provided by 300 volunteers from their left and right hands. The sample data was collected twice, with 10 vein images collected from each palm each time, approximately 60 days apart. Therefore, each palm provided a total of 20 vein images from the two collections. In this way, 300 people can provide 12,000 (300 people × 2 palms × 10 images × 2 collections) palm vein images. Each palm vein image has a resolution of 128 × 128 and 256 gray levels. Some images from Dataset B are shown in Figure 3:
[0107] When collecting palm vein images from datasets A and B, the effects of factors such as illumination, temperature, and angle were fully considered to ensure greater diversity in the obtained vein images. To simulate real-world applications, palm vein images in the datasets were collected at different time periods. In the experiment, the vein images collected in the first phase were used as training images, and the vein images collected in the second phase were used as test images, resulting in training and test sets. To test the recognition performance of the method with different numbers of training samples, the training set was divided into multiple subsets. For example, in dataset A, since 3000 images were collected in each phase, the training and test sets each contained 3000 images. Then, 3, 4, 5, or 6 images were randomly selected from the 6 training images of each hand in the training set as training samples, resulting in four sub-training sets A3, A4, A5, and A6. Similarly, based on the two collection phases, dataset B can be split into a training set and a test set, each containing 6000 images. For each of the 10 images in the training set, 2, 4, 6, 8 or 10 images are randomly selected to train the classifier, thus obtaining 5 sub-training sets, denoted as B2, B4, B6, B8 and B10 respectively.
[0108] To effectively evaluate the performance of this invention, existing deep learning models such as ResNet-50, PV-CNN, and FVRAS-Net were used to extract features from vein images. Then, KNN was used to calculate k nearest neighbor images, and finally, a graph iteration method was used for label prediction to achieve classification and recognition. In addition, comparative experiments were conducted to compare the recognition performance of current state-of-the-art vein recognition algorithms such as Lightweight and SE-Net.
[0109] In formula (3), a value that is too small for the hyperparameter k will reduce the exchange of information, while a value that is too large for k will lead to over-smoothing. In formula (6), the choice of the hyperparameter μ can show that the iterative propagation matrix and the label propagation matrix have an impact on the exchange of information. If the value of μ is too small, the result will be too dependent on the initial state, and if the value of μ is too large, it will affect the exchange of information between the propagation matrices.
[0110] In this experiment, the parameter k was set to 15. Since each class has a maximum of 10 samples in datasets A and B, k > 10 can effectively learn the relationship between different categories, and the calculated labels of the vein images to be classified are more effective.
[0111] When parameter k = 15, different values of μ will also affect the test accuracy. In formula (6), if the value of μ is too small, the label prediction result will become overly dependent on the initial state H. 0Consequently, the relationships between categories cannot be learned. Conversely, a large μ value strengthens the relationship between the label prediction result and other labels, weakening the label information of that category. In order to fully learn the relationships between categories while preserving the category information of that category, μ = 0.7 is preferred in this experiment.
[0112] Here, Resnet-50+KNN+GI, PV-CNN+KNN+GI, and FVRAS-Net+KNN+GI represent vein recognition methods that extract vein features using Resnet-50, PV-CNN, or FVRAS-Net models and then implement the K-nearest neighbor graph iteration in this embodiment. In these methods, the final label propagation matrix is obtained in the manner shown in Equation (10). The recognition results of various methods on two public vein datasets are shown in Figure 1 and Table 2.
[0113] Table 1 shows the recognition accuracy of various methods on dataset A.
[0114]
[0115] Table 2 shows the recognition accuracy of various methods on dataset B.
[0116]
[0117] As can be seen from the experimental results in Tables 1 and 2, the method proposed in this embodiment achieves higher recognition accuracy than the two existing recognition methods, Lightweight and SE-Net, under different training samples. For example, on data A and data B, with 6 training samples per class, the method proposed in this embodiment achieves a maximum recognition rate of 99.67% and 92.72%, respectively, significantly outperforming existing methods. The main reason is that existing deep learning methods train on a single image each time, making it difficult to learn the relationships between different samples or categories. Conversely, the method proposed in this embodiment can learn the relationships between images of different categories, aiding in classification. During the testing phase, the method proposed in this embodiment requires simultaneous use of the image to be classified and training samples from each category. The model can automatically calculate the relationship between the image to be classified and other categories, achieving high-precision classification. Existing deep learning models mainly focus on "memorizing" the category of each image, using this memorized knowledge for classification during testing. However, the method of this invention stores training category information in a "copying" manner; during testing, it only needs to compare the image to be tested with the training image to achieve image classification. The recognition method in this embodiment is simpler, easier to understand and implement than other classification and recognition algorithms. It only requires saving the feature vectors and labels of the vein image training samples. No new model training is required during the testing process. The classification result is obtained by calculating the correlation of the feature vectors, which reduces storage space and improves the computing speed.
[0118] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.
Claims
1. A deep learning-based K-nearest neighbor graph iterative vein recognition method, characterized in that, include: Step 1) Obtain the feature vector of the vein image to be tested and the feature vector set of the vein image training set. Each vein image training sample in the training set has a corresponding label. Step 2) Calculate the distance between the feature vector of the vein image to be tested and each feature vector in the feature vector set, and select k feature vectors from the feature vector set that have the smallest distance to the feature vector of the vein image to be tested. Step 3) Input the selected k+1 feature vectors (including those of the vein image to be tested itself) and the label information of the vein image training samples corresponding to the selected k feature vectors into a graph iterative model. Iteratively learn the correlation between the feature vectors and predict the label information of the vein image to be tested, thereby achieving the final prediction and classification of the vein image to be tested. This includes: establishing a similarity matrix A based on the pairwise similarity between the selected k+1 feature vectors (including those of the vein image to be tested). ,in For feature vectors The similarity is calculated; the iterative propagation matrix D is calculated based on the similarity matrix A as follows: The diagonal matrix of Ʌ is: The label information of each vein image sample in the vein image training set is used... The one-hot label vector representation, Let be the number of categories of the vein image samples, and let be the matrix constructed using the one-hot label vectors of the selected k vein image training samples. Let the label vector of the vein image to be tested be... Initially, it is a vector of all zeros. and Combined into the initial label propagation matrix The label propagation matrix is iteratively updated in the following manner: ,in, It is a hyperparameter, where t is the number of iterations; or, the label propagation matrix after multiple iterations can be obtained directly as follows: ： The label propagation matrix obtained after multiple iterations and updates The vector corresponding to the all-zero vector before the update The portion of the vector is used as the prediction result of the label vector of the vein image to be tested, and the classification and recognition result of the label vector of the vein image to be tested is obtained based on the prediction result.
2. The method according to claim 1, characterized in that, In step 3), the final label propagation matrix H is obtained in the following manner. Δ : and the label propagation matrix The vector corresponding to the all-zero vector before the update The predicted label vector of the vein image to be tested is used as the part of the prediction result. Based on the predicted label vector, the classification result of the label vector of the vein image to be tested is obtained, and then the identification of the label of the vein image to be tested is completed.
3. The method according to claim 1, characterized in that, Similarity It is calculated as follows: 。 4. The method according to claim 1, characterized in that, The distance mentioned in step 2) is the Manhattan distance.
5. The method according to claim 4, characterized in that, Manhattan distance between feature vectors Calculated as follows: , where the feature vector d represents the dimension of the feature vector.
6. The method according to claim 1, characterized in that, The same deep learning model is used to extract the feature vector set of the vein image to be tested and the vein image training set mentioned in step 1).
7. The method according to claim 6, characterized in that, The deep learning model is a Resnet-50, PV-CNN, or FVRAS-Net model.
8. A deep learning-based K-nearest neighbor graph iterative vein recognition system, characterized in that, It includes an encoder module, a recognition module, and a training sample module. The recognition module is connected to the encoder module and the training sample module, respectively. The encoder module is used to extract the feature vector of the input vein image to be tested. The training sample module is used to store the feature vector set of the vein image training sample set, as well as the label information of the vein image training samples that correspond one-to-one with the feature vectors in the feature vector set. The recognition module is used to give the classification and recognition result of the vein image to be tested in accordance with the method described in any one of claims 1-7.
9. The system according to claim 8, characterized in that, The training sample module is also connected to an annotation module and an encoder module. The encoder module is also used to extract the feature vectors of the input vein image training samples. The annotation module is used to provide label information for each vein image training sample according to the input. The training sample module is also used to obtain the feature vectors of each vein image training sample from the encoder module and the label information of each vein image training sample from the annotation module, and save the feature vectors and label information of each vein image training sample in a one-to-one correspondence.