An OCT fingerprint identification method based on a three-dimensional convolutional neural network

Abstract

The application discloses an OCT fingerprint recognition method based on a three-dimensional convolutional neural network, which comprises the following steps: obtaining OCT three-dimensional fingerprint data of a reference fingerprint and a fingerprint to be detected, and the OCT three-dimensional fingerprint data of the reference fingerprint and the fingerprint to be detected respectively contain one sweat gland and one sweat gland; the OCT fingerprint recognition method based on the three-dimensional convolutional neural network takes the three-dimensional data with the sweat gland as the center as the input of the recognition model, compared with the traditional two-dimensional epidermal pore recognition method, not only the feature information of the plane dimension is contained, but also the depth dimension information is increased, the three-dimensional morphological characteristics such as the spiral structure of the sweat gland are completely presented, these multi-dimensional three-dimensional data can provide more detailed and unique biological feature identification, and the distinguishing degree and the accuracy of the fingerprint recognition are effectively improved; meanwhile, the method can keep stable recognition performance under harsh conditions based on the three-dimensional data with the sweat gland as the center, and then the fingerprint anti-counterfeiting capability is significantly improved, and the false fingerprint attack can be effectively resisted.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fingerprint image recognition, and particularly relates to an OCT fingerprint recognition method based on a three-dimensional convolutional neural network. BACKGROUND

[0002] Among numerous biometric recognition technologies, fingerprint recognition has long dominated due to its uniqueness, permanence, universality and convenience. As a key means of individual identity authentication, fingerprint recognition technology is widely used in criminal investigation, entry and exit management, mobile device unlocking, financial payment security, access control and many other fields, and is of great importance to information security and personal privacy, and is one of the cornerstones of the modern identity authentication system.

[0003] Traditional fingerprint recognition mainly relies on primary and secondary features of fingerprints. However, these surface features have significant defects: they are easily affected by factors such as skin wear, stains, humidity, etc., leading to recognition failure; more seriously, they are relatively easy to copy and forge, posing a threat to system security. In contrast, tertiary features of fingerprints (such as the location, shape, density of sweat pores, and sweat glands (i.e., sweat gland structures)) are located deeper in the skin or are closely related to physiological processes, and have extremely high uniqueness and difficulty in being forged, providing a new direction for improving the anti-interference and anti-forgery capabilities of fingerprint recognition.

[0004] Previous research and application of tertiary features have mainly focused on epidermal sweat pores, which, as epidermal structures, are easily affected by significant changes in the physical state of the epidermis (such as wear, dryness, wetness, and scratches), and have limited stability and reliability, posing a security risk and the limitations of epidermal sweat pore features. SUMMARY

[0005] The purpose of the present application is to solve the problems raised in the background art, and to provide an OCT fingerprint recognition method based on a three-dimensional convolutional neural network.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0007] The OCT fingerprint recognition method based on a three-dimensional convolutional neural network proposed by the present application comprises:

[0008] OCT fingerprint three-dimensional data of a reference fingerprint and a fingerprint to be detected are obtained respectively, and the OCT fingerprint three-dimensional data of the reference fingerprint and the fingerprint to be detected respectively contain sweat glands and sweat glands;

[0009] Cutting a three-dimensional data of a preset size from the OCT fingerprint three-dimensional data corresponding to the fingerprint centering on each sweat gland, and inputting each three-dimensional data into a trained identification model, the identification model comprising two branches and a similarity module, each branch comprising an extraction module and an expression module connected in sequence, and a metric module, each extraction module outputting a low-level feature map, and each expression module outputting a sweat gland expression, the sweat gland expressions of the reference fingerprint constituting a first set, and the sweat gland expressions of the fingerprint to be detected constituting a second set;

[0010] The sweat gland expressions in the first set and the second set constitute corresponding nearest neighbor pairs and second nearest neighbor pairs, and the corresponding nearest neighbor pairs and second nearest neighbor pairs satisfying a preset condition are screened out, and the nearest neighbor pairs screened out are taken as candidate pairs;

[0011] The low-level feature maps corresponding to the two sweat gland expressions in each candidate pair are respectively input into the metric modules in the two branches to obtain two high-level feature maps, then the two high-level feature maps are subtracted and input into the similarity module to obtain a similarity score, and the similarity scores corresponding to all candidate pairs are accumulated, and when the accumulated result is greater than a first threshold, then the fingerprint to be detected and the reference fingerprint come from the same finger, otherwise, they come from different fingers.

[0012] Preferably, the output of the extraction module is taken as the input of the metric module, and the three-dimensional data of the reference fingerprint and the fingerprint to be detected are respectively taken as the input of the two branches;

[0013] The extraction module comprises five extraction structures connected in sequence, and each extraction structure comprises a first three-dimensional convolution, a three-dimensional batch normalization, a three-dimensional activation and an aggregation module connected in sequence;

[0014] The aggregation module comprises a three-dimensional channel attention submodule, a three-dimensional spatial attention submodule and a three-dimensional depth attention submodule connected in sequence.

[0015] Preferably, in the three-dimensional channel attention submodule, the input of the focusing module is input into parallel channel-based average pooling and maximum pooling, and after the outputs of the average pooling and the maximum pooling are spliced, they are input into a multi-layer perceptron to obtain a first feature map as the output of the three-dimensional channel attention submodule;

[0016] In the three-dimensional spatial attention submodule, the first feature map is input into parallel channel-based average pooling and maximum pooling, and after the outputs of the average pooling and the maximum pooling are spliced, they are input into a convolution layer and an activation layer in sequence to obtain a spatial attention feature map, and the spatial attention feature map is multiplied with the first feature map to obtain a second feature map, and the second feature map is taken as the output of the three-dimensional spatial attention submodule;

[0017] In the three-dimensional depth attention sub-module, the second feature map passes through parallel spatial-based global maximum pooling and global average pooling, and the outputs of the global maximum pooling and the global average pooling pass through convolution layers respectively, two results of the convolution layer outputs are spliced, then pass through an activation layer to obtain a depth attention feature map, and the depth attention feature map is multiplied with the second feature map to obtain a third feature map, and the third feature map is taken as an output of the aggregation module.

[0018] Preferably, the expression module comprises three extraction structures connected in sequence, a second three-dimensional convolution, a batch normalization layer and an L2 normalization operation, and the metric module comprises two extraction structures connected in sequence.

[0019] Preferably, the similarity module comprises an extraction structure, a three-dimensional average pooling and three fully connected layers connected in series.

[0020] Preferably, the sweat gland expressions in the first set and the second set constitute corresponding nearest neighbor pairs and second nearest neighbor pairs, and the method comprises:

[0021] For each sweat gland expression in the first set, the Euclidean distance between the current sweat gland expression and each sweat gland expression in the second set is calculated, the nearest neighbor and the second nearest neighbor are selected from the second set, and the corresponding first nearest neighbor pair and the second nearest neighbor pair are formed with the current sweat gland expression respectively;

[0022] For each sweat gland expression in the second set, the Euclidean distance between the current sweat gland expression and each sweat gland expression in the first set is calculated, the nearest neighbor and the second nearest neighbor are selected from the first set, and the corresponding second nearest neighbor pair and the second nearest neighbor pair are formed with the current sweat gland expression respectively.

[0023] Preferably, the corresponding nearest neighbor pairs and second nearest neighbor pairs satisfying the preset condition are screened out, and the screened nearest neighbor pairs are taken as candidate pairs, and the method comprises:

[0024] For each corresponding nearest neighbor pair and second nearest neighbor pair, the first Euclidean distance between the two sweat gland expressions in the nearest neighbor pair is calculated, the second Euclidean distance between the two sweat gland expressions in the second nearest neighbor pair is calculated, and then the ratio between the first Euclidean distance and the second Euclidean distance is compared with a second threshold value, when the ratio is less than the second threshold value, it is considered that the current corresponding nearest neighbor pair and second nearest neighbor pair satisfy the preset condition, and the nearest neighbor pair in the corresponding nearest neighbor pair and second nearest neighbor pair satisfying the preset condition is taken as a candidate pair.

[0025] Preferably, the construction process of the training data set of the identification model comprises:

[0026] The OCT fingerprint three-dimensional data of different finger fingerprints is collected, and multiple OCT fingerprint three-dimensional data of each finger fingerprint is collected, and the OCT epidermal fingerprint image of each OCT fingerprint three-dimensional data is obtained, the OCT epidermal fingerprint images of the same finger fingerprint are formed into a group of sequences, the OCT epidermal fingerprint image with the largest effective area of fingerprint in each group of sequences is selected as a reference sample, and the OCT epidermal fingerprint images other than the reference sample in each group of sequences are referred to as ordinary samples, and each sweat gland of the OCT fingerprint three-dimensional data corresponding to each reference sample is labeled with a position tag;

[0027] For each OCT fingerprint three-dimensional data, the 3DSIFT feature of the sweat gland is extracted, and the center depth value of each sweat gland is extracted from the depth direction of the OCT fingerprint three-dimensional data as the depth feature of the sweat gland, and the depth feature and the 3DSIFT feature of the sweat gland are spliced in dimension to obtain the sweat gland descriptor;

[0028] Each sweat gland of the OCT fingerprint three-dimensional data corresponding to each ordinary sample is labeled with a position tag.

[0029] The OCT fingerprint three-dimensional data corresponding to each reference sample and ordinary sample is preprocessed, and the center coordinates of all sweat glands with position tags in the preprocessed OCT fingerprint three-dimensional data are taken as the center, and a three-dimensional data of a preset size is cropped from the preprocessed OCT fingerprint three-dimensional data, and all three-dimensional data constitute a training data set.

[0030] Preferably, the position tag of each sweat gland of the OCT fingerprint three-dimensional data corresponding to each ordinary sample is set, including: for each sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to each ordinary sample in each group of sequences, calculating the corresponding relationship value of each sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the reference sample, repeating iteration multiple times until the iteration number is reached, and among all the corresponding relationship values obtained by the last iteration, the sweat gland position tag of the OCT fingerprint three-dimensional data corresponding to the reference sample corresponding to the minimum corresponding relationship value is selected as the position tag of the current sweat gland of the OCT fingerprint three-dimensional data corresponding to the current ordinary sample.

[0031] The preprocessing of each OCT fingerprint three-dimensional data includes: extracting the epidermal contour curve of each b-scan image in the OCT fingerprint three-dimensional data, and using polynomial fitting to make the epidermal contour curve present as a horizontal straight line in the image.

[0032] Preferably, the calculation formula of the corresponding relationship value is as follows:

[0033] ;

[0034] Wherein, This represents the sweat gland descriptor of the OCT fingerprint 3D volumetric data corresponding to the baseline sampling in the first iteration. Sweat gland descriptors of OCT fingerprint 3D volumetric data corresponding to ordinary sampling The corresponding values ​​between them Sweat gland descriptor representing the OCT fingerprint 3D volumetric data corresponding to the baseline sampling. Sweat gland descriptors of OCT fingerprint 3D volumetric data corresponding to ordinary sampling The Euclidean distance between them Indicates the first Sweat gland descriptor of the OCT fingerprint 3D volume data corresponding to the benchmark sampling in the next iteration Sweat gland descriptors of OCT fingerprint 3D volumetric data corresponding to ordinary sampling The corresponding values ​​between them Indicates the first In the next iteration and The mean square error, Indicates the first Alignment transformation in the next iteration and All parameters are preset.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This OCT fingerprint recognition method based on a three-dimensional convolutional neural network uses three-dimensional data centered on sweat glands as input to the recognition model. Compared with traditional two-dimensional epidermal sweat pore recognition methods, it not only includes planar dimension feature information but also adds depth dimension information, fully presenting the three-dimensional morphological features such as the spiral structure of sweat glands. This multi-dimensional three-dimensional data can provide more detailed and unique biometric identifiers, effectively improving the discrimination and accuracy of fingerprint recognition. At the same time, in harsh external environments (such as when the finger surface is wet, stained, worn, or scratched), traditional two-dimensional epidermal sweat pore-based recognition methods are easily interfered with and their performance degrades. This method, based on three-dimensional data centered on sweat glands, can maintain stable recognition performance under harsh conditions because subcutaneous sweat glands are less affected by the external environment. This significantly improves the fingerprint anti-counterfeiting capability and effectively resists fake fingerprint attacks. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the OCT fingerprint recognition method based on a three-dimensional convolutional neural network according to the present invention.

[0038] Figure 2 This is a schematic diagram of the structure of the recognition model of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure extracted in this invention. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be apparently and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without any creative work fall within the protection scope of the present application.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0042] In one embodiment, as shown in FIG. 1, an OCT fingerprint recognition method based on a three-dimensional convolutional neural network is provided, which comprises the following steps: Figures 1-3

[0043] Step 1, respectively acquiring OCT fingerprint three-dimensional data of a reference fingerprint and a to-be-detected fingerprint (i.e., OCT fingerprint three-dimensional data of the reference fingerprint and OCT fingerprint three-dimensional data of the to-be-detected fingerprint), and the OCT fingerprint three-dimensional data of the reference fingerprint and the to-be-detected fingerprint respectively contain sweat glands and sweat glands (i.e., the OCT fingerprint three-dimensional data of the reference fingerprint contains sweat glands, and the OCT fingerprint three-dimensional data of the to-be-detected fingerprint contains sweat glands);

[0044] It should be noted that the OCT fingerprint three-dimensional data of the reference fingerprint can be data in a fingerprint library in a defense system. By comparing the OCT fingerprint three-dimensional data of the to-be-detected fingerprint with the reference fingerprint, it is determined whether the OCT fingerprint three-dimensional data of the to-be-detected fingerprint and the OCT fingerprint three-dimensional data of the reference fingerprint come from one finger, thereby effectively resisting attacks of fake fingerprints.

[0045] ​Step 2: Centering on each sweat gland (using the coordinates of the sweat gland's center point as the center), crop 3D data of a preset size from the OCT fingerprint 3D volume data of the corresponding fingerprint (i.e., the reference fingerprint and the fingerprint to be detected). The dimensions of the 3D data are 200 pixels for the x-axis, 200 pixels for the y-axis, and 40 pixels for the depth direction (i.e., the z-axis). Input each 3D data into the trained recognition model. The recognition model includes two branches and a similarity module. Each branch contains an extraction module and an expression module connected sequentially (from input to output), as well as a measurement module. The output of the extraction module serves as the input of the measurement module. The 3D data of the reference fingerprint and the fingerprint to be detected serve as the inputs of the two branches, respectively. The extraction module in each branch outputs a low-level feature map, and the expression module in each branch outputs a sweat gland expression (one 3D data corresponds to one sweat gland expression). The expression of individual sweat glands constitutes the first set (the first set) ), the fingerprint to be detected The expression of individual sweat glands constitutes the second set (the second set) );

[0046] It should be noted that the two branches of the recognition model adopt a twin structure (weight sharing).

[0047] The extraction module includes five sequentially connected extraction structures (from input to output), and each extraction structure contains a first 3D convolution (3×3×3), 3D batch normalization, 3D activation (ReLU activation function), and aggregation module (from input to output) connected sequentially. The 3D data with a size of 40×200×200 input to the extraction module passes through the five extraction structures, and the output sizes of the five extraction structures become 40×200×200, 40×200×200, 20×100×100, 20×100×100, and 10×50×50, respectively, and the number of channels becomes 32, 32, 64, 64, and 128, respectively. The extraction module outputs a low-level feature map of 128×10×50×50.

[0048] The aggregation module includes a 3D channel attention submodule, a 3D spatial attention submodule, and a 3D depth attention submodule connected sequentially (from input to output). In the 3D channel attention submodule, the input of the focusing module is subjected to parallel channel-based average pooling and max pooling, and the outputs of average pooling and max pooling are concatenated and then passed through a multilayer perceptron to obtain the first feature map as the output of the 3D channel attention submodule.

[0049] In the three-dimensional spatial attention sub-module, the first feature map is subjected to parallel channel-based average pooling and maximum pooling, the outputs of the average pooling and the maximum pooling are spliced, and then subjected to a convolution layer (7x7) and an activation layer (sigmoid activation function) in sequence to obtain a spatial attention feature map, and the spatial attention feature map is multiplied with the first feature map to obtain a second feature map (enhancing the feature response of the spatial region where the sweat gland is located), and the second feature map is taken as the output of the three-dimensional spatial attention sub-module;

[0050] In the three-dimensional depth attention sub-module, the second feature map is subjected to parallel spatial-based global maximum pooling and global average pooling, the outputs of the global maximum pooling and the global average pooling are respectively subjected to a convolution layer (7x7), the two results of the convolution layer outputs are spliced, and then subjected to an activation layer (sigmoid activation function) to obtain a depth attention feature map, and the depth attention feature map is multiplied with the second feature map to obtain a third feature map (focusing on the morphological features of the sweat gland in a specific depth layer), and the third feature map is taken as the output of the aggregation module.

[0051] Among them, the expression module includes (from input to output direction) three extraction structures connected in sequence (the output sizes of the three extraction structures in the expression module are 10x50x50, 5x25x25, 5x25x25 in sequence, and the channel numbers are changed to 128, 256, 256 respectively), a second three-dimensional convolution (5x25x25), a batch normalization layer and an L2 normalization operation, and the measurement module includes two extraction structures connected in sequence (the output sizes of the two extraction structures in the measurement module are 10x50x50, 5x25x25 in sequence, and the channel numbers are changed to 128, 256 respectively).

[0052] Among them, the similarity module includes (from input to output direction) an extraction structure connected in sequence, a three-dimensional average pooling (compressed to 1x1x1, and the channel number is 256), and three fully connected layers connected in series.

[0053] Step 3, the sweat gland expressions in the first set and the second set constitute corresponding nearest neighbor pairs and second nearest neighbor pairs, including:

[0054] For each sweat gland expression in the first set, the Euclidean distance between the current sweat gland expression and each sweat gland expression in the second set is calculated, the nearest neighbor and the second nearest neighbor are selected from the second set, and the corresponding first nearest neighbor pair and the second nearest neighbor pair are formed with the current sweat gland expression respectively;

[0055] For each sweat gland expression in the second set, the Euclidean distance between the current sweat gland expression and each sweat gland expression in the first set is calculated, the nearest neighbor and the second nearest neighbor are selected from the first set, and the corresponding second nearest neighbor pair and the second nearest neighbor pair are formed with the current sweat gland expression respectively.

[0056] Step 4, screen out the corresponding nearest neighbor pair and the second nearest neighbor pair meeting the preset condition, and take the nearest neighbor pair as a candidate pair, comprising:

[0057] For each corresponding nearest neighbor pair and second nearest neighbor pair, calculate the first Euclidean distance between the two sweat gland expressions in the nearest neighbor pair, and calculate the second Euclidean distance between the two sweat gland expressions in the second nearest neighbor pair, and then compare the ratio between the first Euclidean distance and the second Euclidean distance with the second threshold value. When the ratio is less than the second threshold value, it is considered that the current corresponding nearest neighbor pair and second nearest neighbor pair meet the preset condition, and the nearest neighbor pair in the corresponding nearest neighbor pair and second nearest neighbor pair meeting the preset condition is taken as a candidate pair.

[0058] Step 5, the low-level feature map corresponding to the two sweat gland expressions in each candidate pair (the low-level feature map is the output of the extraction module and is the intermediate feature in the process of generating sweat gland expression) is input into the metric module in two branches to obtain two high-level feature maps. Then, the two high-level feature maps are subtracted and input into the similarity module to obtain a similarity score. The similarity scores corresponding to all candidate pairs are accumulated. When the accumulated result is greater than the first threshold value, the to-be-detected fingerprint and the reference fingerprint come from the same finger, otherwise, they come from different fingers.

[0059] The construction process of the training data set of the identification model comprises:

[0060] OCT fingerprint three-dimensional data of different finger fingerprints are collected, and multiple OCT fingerprint three-dimensional data of each finger fingerprint are collected. An OCT epidermal fingerprint image (the OCT epidermal fingerprint image is a plane composed of x and y directions, which is obtained by mapping the OCT fingerprint three-dimensional data to the fingerprint surface) of each OCT fingerprint three-dimensional data is obtained. The OCT epidermal fingerprint images of the same finger fingerprint are formed into a group of sequences. The OCT epidermal fingerprint image with the largest effective area of fingerprint in each group of sequences is selected as a reference sample, and the OCT epidermal fingerprint images other than the reference sample in each group of sequences are referred to as normal samples. Each sweat gland of the OCT fingerprint three-dimensional data corresponding to each reference sample is labeled with a position label.

[0061] For each OCT fingerprint three-dimensional data, the 3DSIFT algorithm is used to extract the 3DSIFT features of the sweat glands. The center depth value of each sweat gland is extracted from the depth direction of the OCT fingerprint three-dimensional data as the depth feature of the sweat gland. The depth feature and the 3DSIFT feature of the sweat gland are dimensionally spliced to obtain a sweat gland descriptor.

[0062] Each sweat gland of the OCT fingerprint three-dimensional data corresponding to each normal sample is set with a position label, comprising:

[0063] For each sweat gland descriptor in the OCT fingerprint 3D volumetric data corresponding to each normal sample in each sequence group, the correspondence value between the sweat gland descriptor and the OCT fingerprint 3D volumetric data corresponding to the baseline sample is calculated, and the formula for calculating the correspondence value is as follows:

[0064] ;

[0065] in, This represents the sweat gland descriptor of the OCT fingerprint 3D volumetric data corresponding to the baseline sampling in the first iteration. Sweat gland descriptors of OCT fingerprint 3D volumetric data corresponding to ordinary sampling The corresponding values ​​between them Sweat gland descriptor representing the OCT fingerprint 3D volumetric data corresponding to the baseline sampling. Sweat gland descriptors of OCT fingerprint 3D volumetric data corresponding to ordinary sampling The Euclidean distance between them Indicates the first Sweat gland descriptor of the OCT fingerprint 3D volume data corresponding to the benchmark sampling in the next iteration Sweat gland descriptors of OCT fingerprint 3D volumetric data corresponding to ordinary sampling The corresponding values ​​between them Indicates the first All in the next iteration and The mean square error, Indicates the first Alignment transformation in the next iteration and All are preset parameters ( Used to control the ratio, such as 500. Used for numerical stability, to avoid denominators in formulas ( If the value is close to zero, such as taking... ), Represents matrix operations;

[0066] Repeat the iteration multiple times (i.e., repeatedly calculate the correspondence value) until the iteration number is reached. Among all the correspondence values ​​obtained in the last iteration, select the sweat gland location label of the OCT fingerprint 3D volume data corresponding to the minimum correspondence value and assign it as the current sweat gland location label of the OCT fingerprint 3D volume data corresponding to the current ordinary sample (that is, set the current sweat gland location label of the OCT fingerprint 3D volume data corresponding to the current ordinary sample to be the same as the sweat gland location label of the OCT fingerprint 3D volume data corresponding to the minimum correspondence value).

[0067] The OCT fingerprint three-dimensional data corresponding to each normal sampling is preprocessed (the preprocessing process is: extracting the epidermis contour curve of each b-scan image in the OCT fingerprint three-dimensional data, using polynomial fitting to make the epidermis contour curve in the image present as a horizontal straight line, eliminating the distortion of the image due to the curvature of the finger, wherein the b-scan image is a plane composed of xz direction), and the center coordinates of all the sweat glands with assigned labels in the preprocessed OCT fingerprint three-dimensional data are taken as the center. The preprocessed OCT fingerprint three-dimensional data is cut into three-dimensional data (40x200x200) of a predetermined size, and all the three-dimensional data constitute a training data set (and the label of the three-dimensional data is the label of the corresponding assigned sweat gland); the training data in the training data set is used to train the recognition model until the recognition model converges, wherein the loss function of the expression module is a three-tuple loss based on the relative distance between samples, and the measurement module and the similarity module as a whole use a binary cross-entropy loss containing a Sigmoid activation function.

[0068] Compared with the traditional two-dimensional epidermal pore recognition method, the OCT fingerprint recognition method based on a three-dimensional convolutional neural network takes the three-dimensional data centered on the sweat gland as the input of the recognition model, not only contains the feature information of the plane dimension, but also increases the depth dimension information, and completely presents the three-dimensional morphological features such as the spiral structure of the sweat gland. These multi-dimensional three-dimensional data can provide more detailed and unique biometric features, effectively improving the discrimination and accuracy of fingerprint recognition. At the same time, in a harsh external environment (such as a wet finger surface, contaminated with stains, with wear or scratches, etc.), the traditional two-dimensional epidermal pore-based recognition method is easily disturbed and causes performance degradation. The method based on three-dimensional data centered on the sweat gland for recognition can maintain stable recognition performance in harsh conditions because the subcutaneous sweat gland is less affected by the external environment, thereby significantly improving the fingerprint anti-counterfeiting capability and effectively resisting false fingerprint attacks.

[0069] It should be understood that, although Figure 1 The steps in the flowchart of the method are displayed in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the method can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0070] Any combination of the technical features in the above-described embodiments can be made, and for the sake of brevity, not all possible combinations are described, however, as long as the combination of the technical features does not exist in contradiction, it shall be considered within the scope of the present disclosure.

[0071] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it shall not be understood as a limitation on the scope of the present application. It shall be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these shall be within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. An OCT fingerprint recognition method based on a three-dimensional convolutional neural network, characterized in that: The method comprises: The OCT fingerprint three-dimensional data of the reference fingerprint and the to-be-detected fingerprint are respectively acquired, and the OCT fingerprint three-dimensional data of the reference fingerprint and the to-be-detected fingerprint respectively include pilosebaceous units and pilosebaceous units; Cutting a three-dimensional data of a preset size from the OCT fingerprint three-dimensional data corresponding to the fingerprint centering on each sweat gland, and inputting each three-dimensional data into a trained identification model, the identification model comprising two branches and a similarity module, each branch comprising an extraction module and an expression module connected in sequence, and a measurement module, each extraction module outputting a low-level feature map, each expression module outputting a sweat gland expression, and the sweat gland expressions of the reference fingerprint constituting a first set, and the sweat gland expressions of the to-be-detected fingerprint constituting a second set. ​​ The sweat gland expressions in the first set and the second set form corresponding nearest neighbor pairs and second nearest neighbor pairs, the corresponding nearest neighbor pairs and second nearest neighbor pairs satisfying preset conditions are screened out, and the screened nearest neighbor pairs are taken as candidate pairs; The low-level feature maps corresponding to the two sweat gland expressions in each candidate pair are respectively input into the metric modules in the two branches to obtain two high-level feature maps, the two high-level feature maps are subtracted and input into the similarity module to obtain a similarity score, and the similarity scores corresponding to all candidate pairs are accumulated; when the accumulated result is greater than a first threshold, the to-be-detected fingerprint and the reference fingerprint come from the same finger, otherwise, they come from different fingers. 2.The OCT fingerprint identification method based on a three-dimensional convolutional neural network according to claim 1, wherein: The output of the extraction module is taken as the input of the metric module, and the three-dimensional data of the reference fingerprint and the to-be-detected fingerprint are respectively taken as the inputs of the two branches; The extraction module comprises five extraction structures connected in sequence, and each extraction structure comprises a first three-dimensional convolution, a three-dimensional batch normalization, a three-dimensional activation and an aggregation module connected in sequence; The aggregation module comprises a three-dimensional channel attention submodule, a three-dimensional spatial attention submodule and a three-dimensional depth attention submodule connected in sequence. 3.The OCT fingerprint identification method based on three-dimensional convolutional neural network of claim 2, wherein: In the three-dimensional channel attention submodule, the input of the focusing module is input into parallel channel-based average pooling and maximum pooling, the outputs of the average pooling and the maximum pooling are spliced, and then input into a multi-layer perceptron to obtain a first feature map as the output of the three-dimensional channel attention submodule; In the three-dimensional spatial attention submodule, the first feature map is input into parallel channel-based average pooling and maximum pooling, the outputs of the average pooling and the maximum pooling are spliced, and then input into a convolution layer and an activation layer in sequence to obtain a spatial attention feature map, the spatial attention feature map is multiplied with the first feature map to obtain a second feature map, and the second feature map is taken as the output of the three-dimensional spatial attention submodule; In the three-dimensional depth attention submodule, the second feature map is input into parallel spatial-based global maximum pooling and global average pooling, the outputs of the global maximum pooling and the global average pooling are respectively input into a convolution layer, the two outputs of the convolution layer are spliced, and then input into an activation layer to obtain a depth attention feature map, the depth attention feature map is multiplied with the second feature map to obtain a third feature map, and the third feature map is taken as the output of the aggregation module. 4.The OCT fingerprint identification method based on three-dimensional convolutional neural network of claim 2, wherein: The expression module comprises three extraction structures connected in sequence, a second three-dimensional convolution, a batch normalization layer and an L2 normalization operation, and the metric module comprises two extraction structures connected in sequence. 5.The OCT fingerprint identification method based on three-dimensional convolutional neural network of claim 2, wherein: The similarity module comprises an extraction structure, a three-dimensional average pooling and three serially connected fully connected layers. 6.The OCT fingerprint identification method based on three-dimensional convolutional neural network of claim 1, wherein: The sweat gland expressions in the first set and the second set form corresponding nearest neighbor pairs and second nearest neighbor pairs, comprising: For each sweat gland expression in the first set, the Euclidean distances between the current sweat gland expression and each sweat gland expression in the second set are calculated, the nearest neighbor and the second nearest neighbor are selected from the second set, and the corresponding first nearest neighbor pair and the first second nearest neighbor pair are formed with the current sweat gland expression respectively; For each sweat gland expression in the second set, a Euclidean distance between the current sweat gland expression and each sweat gland expression in the first set is calculated, the nearest neighbor and the second nearest neighbor are selected from the first set, and a corresponding second nearest neighbor pair and a second second nearest neighbor pair are formed with the current sweat gland expression, respectively.

7. The OCT fingerprinting method based on three-dimensional convolutional neural network of claim 1, wherein: The corresponding nearest neighbor pair and the second nearest neighbor pair satisfying the preset condition are screened out, and the screened nearest neighbor pair is taken as the candidate pair, which comprises: For each corresponding nearest neighbor pair and second nearest neighbor pair, a first Euclidean distance between the two sweat gland expressions in the nearest neighbor pair is calculated, a second Euclidean distance between the two sweat gland expressions in the second nearest neighbor pair is calculated, and a ratio between the first Euclidean distance and the second Euclidean distance is compared with a second threshold value, when the ratio is less than the second threshold value, it is considered that the current corresponding nearest neighbor pair and the second nearest neighbor pair satisfy the preset condition, and the nearest neighbor pair in the corresponding nearest neighbor pair and the second nearest neighbor pair satisfying the preset condition is taken as the candidate pair. 8.The OCT fingerprint identification method based on three-dimensional convolutional neural network of claim 1, wherein: The construction process of the training data set of the identification model comprises: OCT fingerprint three-dimensional data of different finger fingerprints are collected, and multiple OCT fingerprint three-dimensional data of each finger fingerprint are collected, and an OCT epidermal fingerprint image of each OCT fingerprint three-dimensional data is obtained, OCT epidermal fingerprint images of the same finger fingerprint are formed into a group of sequences, an OCT epidermal fingerprint image with the largest effective area of fingerprint is selected from each group of sequences as a reference sample, and OCT epidermal fingerprint images other than the reference sample in each group of sequences are referred to as ordinary samples, and a position label of each sweat gland of the OCT fingerprint three-dimensional data corresponding to each reference sample is set; For each OCT fingerprint three-dimensional data, a 3DSIFT feature of the sweat gland is extracted, and a center depth value of each sweat gland is extracted from a depth direction of the OCT fingerprint three-dimensional data as a depth feature of the sweat gland, and the depth feature and the 3DSIFT feature of the sweat gland are spliced in dimensions to obtain a sweat gland descriptor; A position label of each sweat gland of the OCT fingerprint three-dimensional data corresponding to each ordinary sample is set; The OCT fingerprint three-dimensional data corresponding to each reference sample and ordinary sample is preprocessed, and a center coordinate of each sweat gland with a position label in the preprocessed OCT fingerprint three-dimensional data is taken as a center, a three-dimensional data of a preset size is cropped from the preprocessed OCT fingerprint three-dimensional data, and all the three-dimensional data constitute a training data set. 9.The OCT fingerprint identification method based on three-dimensional convolutional neural network of claim 8, wherein: The position label of each sweat gland of the OCT fingerprint three-dimensional data corresponding to each ordinary sample is set, which comprises: for each sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to each ordinary sample in each group of sequences, a corresponding relationship value of each sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the reference sample is calculated, and the iteration is repeated for multiple times until the iteration number is reached, and among all the corresponding relationship values obtained by the last iteration, a sweat gland position label of the OCT fingerprint three-dimensional data corresponding to the reference sample corresponding to the minimum corresponding relationship value is taken as a position label of a current sweat gland of the OCT fingerprint three-dimensional data corresponding to the current ordinary sample. The OCT fingerprint three-dimensional data is pre-processed, including: extracting the epidermis profile curve of each b-scan image in the OCT fingerprint three-dimensional data, and adopting polynomial fitting to make the epidermis profile curve present as a horizontal straight line in the image.

10. The OCT fingerprinting method based on a three-dimensional convolutional neural network of claim 9, wherein: The calculation formula of the corresponding relationship value is as follows: ； wherein, a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the reference sample in the first iteration a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the normal sample a corresponding relationship value between, a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the reference sample a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the normal sample a Euclidean distance between, a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the reference sample in the first iteration a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the normal sample in the first iteration a corresponding relationship value between, a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the reference sample in the first iteration a sweat gland descriptor of the OCT fingerprint three-dimensional data corresponding to the normal sample in the first iteration a mean square error of, a mean square error of, an alignment transformation in the first iteration, an alignment transformation in the first iteration, an alignment transformation in the first iteration, an alignment transformation in the first iteration, an alignment transformation in the first iteration,

Images