A Finger-Vein Hash Index Method Based on Positive-Negative Equilibrium Round Optimization and Multi-Table Complementation

Abstract

The present invention provides a finger vein hashing index method based on positive and negative balance round optimization and multi-table complementation. It includes: the training step of a multi-model finger vein indexing network and the finger vein image retrieval step based on multi-table code-level weighting; wherein: the training step of the multi-model finger vein indexing network includes constructing a multi-model training set, setting the multi-model network structure, data loading based on positive and negative balance round optimization, and multi-model complementary training; the finger vein image retrieval step based on multi-table code-level weighting includes registering a multi-table finger vein database and finger vein image query based on multi-table code-level weighting. The present invention proposes a hash code superiority and inferiority evaluation algorithm for abnormal hash codes, complementarily trains multiple hash tables, and performs multi-table code-level weighting in the matching stage to effectively reduce the influence caused by abnormal hash codes. At the same time, aiming at the problem of extremely unbalanced positive and negative samples of finger vein data, positive and negative balance round optimization is proposed to enable the network to use enough similar pairs for training.

Description

Technical Field

[0001] The present invention relates to the technical field of finger vein recognition. In particular, it relates to a finger vein hash indexing method based on positive and negative balance round optimization and multi-table complementarity. Background Art

[0002] In recent years, finger vein recognition has been gradually widely applied due to its advantages such as non-contact, living body recognition, and high security. High-dimensional features can be effectively applied to finger vein recognition, but it will lead to higher computational costs and storage costs, especially in an increasingly large finger vein database. Because of its powerful expression ability, deep learning has been applied in various fields. Therefore, usually, the deep hash method is used to map images into binary hash codes to obtain a fast query speed and reduce storage requirements. The finger vein data set has the characteristics of many categories and few samples in the same category. In the process of similarity learning, similar information is very sparse. It is very difficult for the network to learn the pulling force between samples in the same category, and it is easy to generate abnormal hash codes that deviate from the center of samples in the same category. The existing finger vein indexing algorithms do not pay enough attention to the problems of abnormal hash codes and extremely unbalanced positive and negative samples, and it is difficult to balance the recall rate and accuracy.

[0003] The existing methods are as follows: Obtain a finger vein training image library and perform data augmentation on it. Use a pair of similar / dissimilar images as input, extract features through a CNN based on a multi-scale attention mechanism, and map the finger vein features into hash codes through the collaborative supervision of a cross-entropy loss function, a contrast loss function, and a retrieval task loss function in the fully connected layer to construct an end-to-end finger vein image quantization coding network. In the registration stage, the quantized codes of the finger vein library images are stored in the finger vein coding database. In the query stage, the hash code of the query image is matched one by one with the hash codes of the library images through the Hamming distance, and the final query result is returned. The existing methods have the following defects:

[0004] (1) The finger vein data set has the characteristics of many categories and few samples in the same category. In the process of similarity learning of hash codes by the existing deep hash algorithms, the similarity comparison between similar and dissimilar pairs is much sparser. This extremely unbalanced situation of positive and negative samples will make it difficult for the network to learn the similarity between samples in the same category.

[0005] (2) Deep hashing based on mini-batch optimization is very easy to generate abnormal hash codes that deviate from the center of samples in the same category due to insufficient training. The existing finger vein indexing methods only use a single hash table, and are very sensitive to abnormal bad codes during the matching process, which is very easy to cause a large number of missed detections and false detections, and it is difficult to balance the recall rate and accuracy. Summary of the Invention

[0006] In order to meet the usage requirements of users in the case of large-scale finger vein data, the present invention provides a finger vein hashing index method based on positive and negative balance round optimization and multi-table complementarity, proposes a hash code superiority and inferiority evaluation algorithm for abnormal hash codes, complements and trains multiple hash tables, and performs multi-table code-level weighting in the matching stage to effectively reduce the impact caused by abnormal hash codes. At the same time, aiming at the problem of extremely unbalanced positive and negative samples in finger vein data, positive and negative balance round optimization is proposed to enable the network to use enough similar pairs for training.

[0007] The technical means adopted by the present invention are as follows:

[0008] A finger vein hashing index method based on positive and negative balance round optimization and multi-table complementarity, including: the training step of a multi-model finger vein index network and the finger vein image retrieval step based on multi-table code-level weighting; wherein:

[0009] The training step of the multi-model finger vein index network includes constructing a multi-model training set, setting the multi-model network structure, data loading based on positive and negative balance round optimization, and multi-model complementary training;

[0010] The finger vein image retrieval step based on multi-table code-level weighting includes registering a multi-table finger vein database and finger vein image query based on multi-table code-level weighting.

[0011] Further, in the training step of the multi-model finger vein index network, the construction of the multi-model training set specifically includes:

[0012] Randomly extract the finger vein image training set I n times, each time extracting 90% of the training set, to form n training subsets (T1, T2,..., T n ), which respectively correspond to the training sets of n models.

[0013] Further, in the training step of the multi-model finger vein index network, the setting of the multi-model finger vein index network structure specifically includes:

[0014] The network structure of each sub-model uses AlexNet as the CNN to extract the features of finger vein images, and maps the original features into finger vein hash codes by setting a fully connected layer;

[0015] The fully connected layer part includes a fully connected layer and a hash layer. The output unit number of the fully connected layer is 1000, relu is used as the activation function, and a dropout layer with a connection loss rate of 0.5 is connected. The hash layer is set with the unit number of k to output a hash code with a length of k;

[0016] Set the n models with the same network structure to form a multi-model network structure.

[0017] Further, in the training step of the multi-model finger vein indexing network, the data loading based on positive and negative balanced rounds optimization specifically includes:

[0018] Perform data loading in two ways for each training subset, namely randomly shuffled data loading method and positive sample maintained data loading method; among them, the randomly shuffled data loading method includes: randomly shuffling the training data and dividing the shuffled data into multiple batches; the positive sample maintained data loading method includes: loading the training data in the category order of finger vein images and then performing batching processing;

[0019] Set every c rounds as a training cycle, and determine the data loading method according to whether the number of training rounds e is a multiple of c. If so, perform the positive sample maintained data loading method, otherwise perform the randomly shuffled data loading method.

[0020] Further, in the training step of the multi-model finger vein indexing network, the multi-model complementary training includes: sequentially inputting the loaded training subsets into the corresponding models for multi-model complementary training, specifically including:

[0021] Train the first basic model, and the training process is as follows:

[0022] The first model is trained using the basic contrast loss function, and the basic contrast loss function is defined as:

[0023]

[0024] where N is the number of samples; S i,j is the class flag bit of the i-th and j-th images. When the two hash codes are of the same class, S i,j = 1, otherwise it is 0; m is the boundary threshold, which is set to the length k of the hash code in the instance;

[0025] Before training the subsequent models (the 2nd to the nth models), evaluate the superiority and inferiority of the hash codes generated by the previous model (model t - 1), and the evaluation process is as follows:

[0026] Input the training set T of the t-th model t into the trained (t - 1)-th model to obtain a series of hash codes, calculate the similarity score of each hash code, and thus evaluate the superiority and inferiority of each hash code;

[0027] Evaluate the similarity preservation ability of the hash codes by calculating the distance between samples within the class of the hash codes and performing normalization processing. The normalized similarity score calculation formula for the i-th hash code b i is:

[0028]

[0029] where d H (b i , b j ) is the Hamming distance between hash code b i and b j ; Ν is the total number of samples; N s is the number of samples in the same class as b i ; μ is the normalization parameter, set to the number of bits k of the hash code;

[0030] The t-th model is enhanced trained, and the training process is as follows:

[0031] A series of scores are obtained through the evaluation of the quality of the hash code, and the weight term is calculated according to the scores. The calculation formula of the weight term is:

[0032]

[0033] The attention of the model t to the inferior code is increased by multiplying the weight term in the complementary loss function. The complementary loss function is defined as:

[0034]

[0035] When the t-th model finishes training, judge whether t is equal to n. If not, continue to train the (t + 1)-th model in an enhanced manner; if so, end the training and obtain the n trained models, that is, n hash functions (Μ1, Μ2,..., Μ n ).

[0036] Furthermore, in the step of the multi-table code-level weighted finger vein image retrieval, the registration of the multi-table finger vein database includes: generating the multi-table finger vein library graph hash code and calculating the hash code matching weight, where:

[0037] The generating of the multi-table finger vein library graph hash code includes:

[0038] Taking the finger vein library graph set as the input, generating the multi-table finger vein library graph hash code Β ∈ {0, 1} n×N×k through n hash functions; for any finger vein library graph x i , the binary hash code generated by feature extraction and quantization coding through the t-th hash function Μ t is defined as b i t ∈ B;

[0039] The calculating of the hash code matching weight includes:

[0040] For the generated multi-table finger vein library graph hash code Β ∈ {0, 1}n×N×k Evaluate the advantages and disadvantages of each hash code in it, that is, evaluate n×N hash codes to obtain a score set G∈[0,1] n×N , where the score of the i-th hash code in the t-th hash table is defined as Calculate its matching weight according to the score of the advantages and disadvantages of each hash code. The matching weight of the i-th hash code in the t-th hash table The calculation formula of the matching weight is defined as:

[0041]

[0042] where g p is the score threshold. For abnormal hash codes with scores lower than the score threshold g p , they are rejected from participating in the matching;

[0043] Calculate the matching weight of each hash code in turn to obtain a weight set M∈[0,1] n×N ; Store the multi-table hash code database Β and the matching weight set M offline in the finger vein database to complete the registration of the multi-table finger vein database.

[0044] Furthermore, in the step of finger vein image retrieval based on multi-table code-level weighting, the finger vein image query based on multi-table code-level weighting includes: generating query graph encoding and multi-table code-level weighted matching, where:

[0045] The generation of the query graph encoding includes:

[0046] Input the finger vein query image q into n hash functions (Μ1,Μ2,...,Μ n ) to generate n segments of query graph encoding; where the query graph hash code generated by the t-th hash function is defined as

[0047] The multi-table code-level weighted matching includes:

[0048] Calculate the Hamming distance between the query graph encoding and the library graph encoding in each hash table. The query graph encoding generated by the t-th hash function and the encoding of the i-th library graph x i is The Hamming distance between them is represented by ; Perform multi-table code-level weighting according to the matching weights of the library graph encodings in multiple hash tables to obtain the weighted Hamming distance d between the query graph q and the i-th library graph w :

[0049]

[0050] Sort according to the weighted Hamming distance between the finger vein query graph and the library graph, and return the top l finger vein library graphs with the smallest distance as the retrieval results.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. The finger vein hash index method based on positive and negative balance round optimization and multi-table complementarity provided by the present invention proposes a data loading method based on positive and negative balance round optimization, enabling the network to have a sufficient number of similar pairs for training, and improving the aggregation ability of the finger vein index network for the same type of hash codes.

[0053] 2. The finger vein hash index method based on positive and negative balance round optimization and multi-table complementarity provided by the present invention proposes an evaluation algorithm for the quality of hash codes, trains multiple models in an enhanced manner, integrates multiple hash tables for multi-table code-level weighted matching, effectively reduces the influence caused by abnormal inferior codes, and improves the retrieval performance.

[0054] For the above reasons, the present invention can be widely promoted in the fields such as finger vein recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0056] Figure 1 It is a training flow chart of the multi-model finger vein index network of the present invention.

[0057] Figure 2 It is a structural diagram of the multi-model finger vein index network of the present invention.

[0058] Figure 3 It is a finger vein image retrieval flow chart based on multi-table code-level weighting of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail the present invention.

[0060] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. The description of at least one exemplary embodiment is actually only illustrative and in no way restrictive of the present invention and its application or use. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0061] It should be noted that the terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly dictates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they specify the presence of the stated features, steps, operations, devices, components, and / or combinations thereof.

[0062] Unless otherwise specifically stated, the relative arrangements, numerical expressions, and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be clear that, for the convenience of description, the dimensions of the various parts shown in the drawings are not drawn in actual proportional relationships. Technologies, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the said technologies, methods, and devices should be regarded as part of the authorization specification. In all the examples shown and discussed here, any specific values should be construed as merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that like reference numerals and letters denote like items in the following drawings, and thus, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0063] In the description of the present invention, it should be understood that the orientation terms such as "front, rear, upper, lower, left, right", "lateral, vertical, perpendicular, horizontal", and "top, bottom", etc. generally indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description. Without contrary instructions, these orientation terms do not indicate and imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of the present invention. The orientation terms "inside, outside" refer to the inside and outside relative to the contour of each component itself.

[0064] For ease of description, spatial relative terms such as "above", "over", "on the upper surface", "upper", etc. can be used here to describe the spatial positional relationship between a device or feature shown in the figure and other devices or features. It should be understood that the spatial relative terms are intended to encompass different orientations in use or operation in addition to the orientation described in the figure of the device. For example, if the device in the attached drawing is inverted, the device described as "above other devices or structures" or "over other devices or structures" will then be positioned as "below other devices or structures" or "under other devices or structures". Thus, the exemplary term "above" can include both the orientations of "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and corresponding interpretations are made for the spatial relative descriptions used here.

[0065] In addition, it should be noted that the use of words such as "first", "second", etc. to limit components is only for the convenience of differentiating the corresponding components. Without additional statements, the above words have no special meanings, and thus should not be construed as limiting the protection scope of the present invention.

[0066] The present invention provides a finger vein hashing index method based on positive and negative balanced round optimization and multi-table complementarity, including: a training step of a multi-model finger vein index network and a finger vein image retrieval step based on multi-table code-level weighting; wherein:

[0067] The training step of the multi-model finger vein index network, as Figure 2 shown, includes constructing a multi-model training set, setting a multi-model network structure, data loading based on positive and negative balanced round optimization, and multi-model complementary training;

[0068] The finger vein image retrieval step based on multi-table code-level weighting includes registering a multi-table finger vein database and querying finger vein images based on multi-table code-level weighting.

[0069] Specifically, as a preferred embodiment of the present invention, in the training step of the multi-model finger vein index network, the construction of the multi-model training set specifically includes:

[0070] Randomly extracting the finger vein image training set I n times, each time extracting 90% of the training set, to form n training subsets (T1, T2,..., T n ), which respectively correspond to the training sets of n models.

[0071] Specifically, as a preferred embodiment of the present invention, in the training step of the multi-model finger vein index network, the structural diagram of the multi-model finger vein index network is as Figure 2As shown in the figure. The multi-model finger vein indexing network structure is set up, specifically including:

[0072] The network structure of each sub-model uses AlexNet as the CNN to extract features of finger vein images, and the original features are mapped into finger vein hash codes by setting a fully connected layer;

[0073] The fully connected layer part includes a fully connected layer and a hash layer. The number of output units of the fully connected layer is 1000, relu is used as the activation function, and a dropout layer with a connection loss rate of 0.5 is connected. The hash layer is set with k units to output a hash code with a length of k;

[0074] Set n models with the same network structure to form a multi-model network structure.

[0075] Specifically in implementation, as a preferred implementation manner of the present invention, in the training step of the multi-model finger vein indexing network, the data loading based on positive and negative balanced rounds of optimization specifically includes:

[0076] Data is loaded for each training subset in two ways, namely the randomly shuffled data loading method and the positive sample-preserving data loading method; among them, the randomly shuffled data loading method includes: randomly shuffling the training data and dividing the shuffled data into multiple batches; the positive sample-preserving data loading method includes: loading the training data in the order of the categories of finger vein images and then performing batching processing;

[0077] Set every c rounds as a training cycle, and determine the data loading method according to whether the number of training rounds e is a multiple of c. If it is, the positive sample-preserving data loading method is performed, otherwise the randomly shuffled data loading method is performed.

[0078] Specifically in implementation, as a preferred implementation manner of the present invention, in the training step of the multi-model finger vein indexing network, the multi-model complementary training includes: sequentially inputting the loaded training subsets into the corresponding models for multi-model complementary training, specifically including:

[0079] Train the first basic model, and the training process is as follows:

[0080] The 1st model is trained using the basic contrast loss function, and the basic contrast loss function is defined as:

[0081]

[0082] where N is the number of samples; S i,j is the class flag bit of the i-th and j-th images. When the two hash codes are of the same class, S i,j= 1, otherwise 0; m is the boundary threshold, set to the length k of the hash code in the example;

[0083] Before training the subsequent models (the 2nd to the nth models), evaluate the quality of the hash codes generated by the previous model (model t - 1). The evaluation process is as follows:

[0084] Input the training set T of the t-th model t into the trained (t - 1)-th model to obtain a series of hash codes, calculate the similarity score of each hash code, and thus evaluate the quality of each hash code;

[0085] Evaluate the similarity preservation ability of the hash codes by calculating the distance between samples within the same class of the hash codes and performing normalization processing. The normalized similarity score formula for the i-th hash code b i is:

[0086]

[0087] where d H (b i , b j ) is the Hamming distance between hash code b i and b j ; Ν is the total number of samples; N s is the number of samples in the same class as b i ; μ is the normalization parameter, set to the number of bits k of the hash code;

[0088] Perform enhanced training on the t-th model. The training process is as follows:

[0089] After evaluating the quality of the hash codes, obtain a series of scores, calculate the weight term according to the scores, and the formula for the weight term is:

[0090]

[0091] Increase the attention of model t to poor codes by multiplying the weight term in the complementary loss function. The complementary loss function is defined as:

[0092]

[0093] When the t-th model finishes training, determine whether t is equal to n. If not, continue to train the (t + 1)-th model in an enhanced manner; if so, end the training to obtain the trained n models, that is, n hash functions (Μ1, Μ2,..., Μ n ).

[0094] Specifically, in a preferred embodiment of the present invention, such as Figure 3As shown, in the steps of the multi-table code-level weighted finger vein image retrieval, the registration of the multi-table finger vein database includes: generating the multi-table finger vein library graph hash code and calculating the hash code matching weight, where:

[0095] The generation of the multi-table finger vein library graph hash code includes:

[0096] Taking the finger vein library graph set as the input, generating the multi-table finger vein library graph hash code Β ∈ {0, 1} n×N×k through n hash functions; where for any finger vein library graph x i , the binary hash code generated by feature extraction and quantization encoding through the t-th hash function Μ t is defined as

[0097] The calculation of the hash code matching weight includes:

[0098] Evaluating the superiority and inferiority of each hash code in the generated multi-table finger vein library graph hash code Β ∈ {0, 1} n×N×k , that is, evaluating n × N hash codes to obtain the scoring set G ∈ [0, 1] n×N , where the score of the i-th hash code in the t-th hash table is defined as Calculating the matching weight according to the superiority and inferiority scores of each hash code, and the matching weight calculation formula of the i-th hash code in the t-th hash table is defined as:

[0099]

[0100] where g p is the score threshold, which is set to 0.6 in the example, and abnormal hash codes with scores lower than 0.6 are rejected from participating in the matching;

[0101] Calculating the matching weight of each hash code in turn to obtain the weight set M ∈ [0, 1] n×N ; Storing the multi-table hash code database Β and the matching weight set M offline in the finger vein database together to complete the registration of the multi-table finger vein database.

[0102] In specific implementation, as a preferred implementation manner of the present invention, continue to refer to Figure 3 , in the steps of the multi-table code-level weighted finger vein image retrieval, the multi-table code-level weighted finger vein image query includes: generating the query graph encoding and multi-table code-level weighted matching, where:

[0103] The generation of the query graph encoding includes:

[0104] Input the finger vein query image q into n hash functions (Μ1, Μ2,..., Μ n ) to generate n segments of query graph codes; the query graph hash code generated by the t-th hash function is defined as

[0105] The multi-table code-level weighted matching includes:

[0106] Calculate the Hamming distance between the query graph code and the gallery graph code in each hash table. The Hamming distance between the query graph code generated by the t-th hash function and the code i of the i-th gallery graph x is represented by ; perform multi-table code-level weighting according to the matching weights of the gallery graph codes in multiple hash tables to obtain the weighted Hamming distance d w between the query graph q and the i-th gallery graph:

[0107]

[0108] Sort according to the weighted Hamming distance between the finger vein query graph and the gallery graph, and return the top l finger vein gallery graphs with the smallest distance as the retrieval results.

[0109] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A finger vein hashing index method based on positive and negative balance round optimization and multi-table complementation, characterized in that Including: The training steps of the multi-model finger vein indexing network and the finger vein image retrieval steps based on multi-table code-level weighting; where: The training steps of the multi-model finger vein indexing network include constructing a multi-model training set, setting the multi-model network structure, data loading based on positive-negative balanced rounds optimization, and multi-model complementary training; The data loading based on positive-negative balanced rounds optimization specifically includes: Performing data loading in two ways for each training subset, namely randomly shuffled data loading method and positive sample-preserved data loading method; among them, the randomly shuffled data loading method includes: randomly shuffling the training data and dividing the shuffled data into multiple batches; the positive sample-preserved data loading method includes: loading the training data in the category order of finger vein images and then performing batching processing; Setting every c rounds as a training cycle, and determining the data loading method according to whether the number of training rounds e is a multiple of c. If so, perform the positive sample-preserved data loading method, otherwise perform the randomly shuffled data loading method; The finger vein image retrieval steps based on multi-table code-level weighting include registering a multi-table finger vein database and finger vein image query based on multi-table code-level weighting; The registering of the multi-table finger vein database includes: generating multi-table finger vein library graph hash codes and calculating hash code matching weights, where: The generating of the multi-table finger vein library graph hash codes includes: Finger Vein Database Atlas As input, multiple - table finger vein database map hash codes Β∈{0,1} are generated through n hash functions n×N×k ; where for any finger vein database map x i , the binary hash code generated by feature extraction and quantization coding through the t - th hash function Μ t is defined as The calculating of the hash code matching weights includes: Evaluate the quality of each generated multi-modal finger vein database graph hash code Β ∈ {0, 1} n×N×k in the set, that is, evaluate n × N hash codes to obtain a score set G ∈ [0, 1] n×N , where the score of the i-th hash code in the t-th hash table is defined as Calculate the matching weight according to the quality score of each hash code. The matching weight of the i-th hash code in the t-th hash table is defined by the following calculation formula: Among them, g p is the fractional threshold. For abnormal hash codes with scores lower than the fractional threshold g p are rejected from participating in the matching; Calculate the matching weights of each hash code in sequence to obtain a weight set \(M\in[0,1]\). n×N Store the multi-table hash code database \(\varBeta\) and the matching weight set \(M\) offline in the finger vein database to complete the registration of the multi-table finger vein database.

2. The finger vein hashing index method based on positive and negative balance round optimization and multi-table complementarity according to claim 1, wherein In the training steps of the multi-model finger vein indexing network, the constructing of the multi-model training set specifically includes: Randomly extract the finger vein image training set I n times, each time extracting 90% of the training set to form n training subsets (T1, T2,..., T n ), which respectively correspond to the training sets of n models.

3. The finger vein hashing index method based on positive and negative balance round optimization and multi-table complementarity according to claim 1, wherein In the training steps of the multi-model finger vein indexing network, setting the multi-model finger vein indexing network structure specifically includes: The network structure of each sub-model uses AlexNet as the CNN for feature extraction of finger vein images, and maps the original features into finger vein hash codes by setting a fully connected layer; The fully connected layer part includes a fully connected layer and a hash layer. The number of output units of the fully connected layer is 1000, using relu as the activation function, connecting a dropout layer with a connection loss rate of 0.5, and the hash layer sets the number of units to k to output a hash code with a length of k; Setting n models with the same network structure to form a multi-model network structure.

4. The finger vein hashing index method based on positive and negative balance round optimization and multi-table complementarity according to claim 1, characterized in that In the training steps of the multi-model finger vein indexing network, the multi-model complementary training includes: sequentially inputting the loaded training subsets into the corresponding models for multi-model complementary training, specifically including: Training the first basic model, and the training process is as follows: The 1st model is trained using the basic contrast loss function, and the basic contrast loss function is defined as: where N is the number of samples; S i,j is the class flag bit of the i-th and j-th images. When the two hash codes are of the same class, S i,j = 1, otherwise 0; m is the boundary threshold, which is set to the length k of the hash code in the example; Before training the 2nd model to the nth model, evaluating the superiority and inferiority of the hash codes generated by the (t - 1)th model, and the evaluation process is as follows: Input the training set T of the t-th model t into the trained (t - 1)-th model to obtain a series of hash codes, calculate the similarity scores of each hash code, and thus evaluate the quality of each hash code; The similarity preservation ability of the hash code is evaluated by calculating the distance between samples within the class of the hash code and performing normalization processing. The normalized similarity score calculation formula for the i-th hash code b i is as follows: where d H (b i , b j ) is the Hamming distance between hash code b i and b j ; Ν is the total number of samples; N s is the number of samples in the same class as b i ; μ is the normalization parameter, set to the number of bits k of the hash code; Performing enhanced training on the tth model, and the training process is as follows: Obtaining a series of scores through the evaluation of the superiority and inferiority of the hash codes, calculating the weight term according to the scores, and the calculation formula of the weight term is: Increasing the attention of model t to inferior codes by multiplying the weight term in the complementary loss function, and the complementary loss function is defined as: When the t-th model finishes training, determine whether t is equal to n. If not, continue to train the (t + 1)-th model in an enhanced manner; if so, end the training to obtain n trained models, that is, n hash functions (Μ1, Μ2,..., Μ n ).

5. The finger vein hashing index method based on positive and negative balance round optimization and multi-table complementarity according to claim 1, characterized in that In the steps of the finger vein image retrieval based on multi-table code-level weighting, the finger vein image query based on multi-table code-level weighting includes: generating a query graph code and multi-table code-level weighted matching, where: The generating of the query graph code includes: Input the finger vein query image q into n hash functions (Μ1, Μ2,..., Μ n ) to generate n segments of query graph codes; the query graph hash code generated by the t-th hash function is defined as The multi-table code-level weighted matching includes: Calculate the Hamming distance between the query graph encoding and the library graph encoding within each hash table, where the query graph encoding generated by the t-th hash function and the encoding of the i-th library graph x i is denoted using ; perform multi-table code-level weighting based on the matching weights of the library graph encodings in multiple hash tables to obtain the weighted Hamming distance d between the query graph q and the i-th library graph w : Sort according to the weighted Hamming distance between the finger vein query graph and the library graph, and return the top l finger vein library graphs with the smallest distance as the retrieval result.

Images