A small area fingerprint feature extraction method based on deep learning

Abstract

The present invention discloses a method for extracting small-area fingerprint features based on deep learning, and the method comprises: constructing a fingerprint feature extraction network, and collecting a first fingerprint image and a second fingerprint image; the second fingerprint image is a transformed image corresponding to the first fingerprint image; the output of the fingerprint feature extraction network is a fingerprint feature; the first fingerprint image is input into the fingerprint feature extraction network to obtain a first fingerprint feature; the second fingerprint image is input into the fingerprint feature extraction network to obtain a second fingerprint feature; the target loss function is calculated according to the first fingerprint feature and the second fingerprint feature; the fingerprint feature extraction network is trained according to the target loss function; the small-area fingerprint image to be extracted is collected, the small-area fingerprint image to be extracted is input into the trained fingerprint feature extraction network, and the third fingerprint feature is output. The fingerprint features extracted by the method have high discrimination and stability.

Description

Technical Field

[0001] The present invention relates to the technical field of fingerprint feature extraction, and in particular to a small-area fingerprint feature extraction method based on deep learning. Background Art

[0002] Since Apple integrated small-area fingerprint recognition modules into smartphones in 2013, small-area fingerprint recognition on mobile devices has begun to develop rapidly. Products such as fingerprint USB flash drives, fingerprint door locks, and fingerprint luggage have begun to appear in the market and grow rapidly.

[0003] Traditional fingerprint recognition algorithms require a fingerprint area that is large enough to extract a sufficient number of fingerprint minutiae points for effective matching and recognition. However, the number of minutiae points extracted from a small area fingerprint is generally small (usually less than 20), which cannot meet the requirements of the traditional algorithms based on minutiae points for the number of minutiae points. Among them, the combination of three-level features and minutiae points can greatly improve the performance of small area fingerprint recognition, but extracting reliable three-level features requires high-resolution fingerprint images (at least 1500DPI). The mainstream fingerprint sensors on the market are mainly 500DPI. High-resolution fingerprint sensors are expensive and lack competitiveness. Therefore, this type of method lacks competitiveness in terms of cost performance. The biggest problem with pixel-based methods is the large amount of calculation. Although there are techniques such as pruning search and fast Fourier transform to reduce the complexity of calculation, this type of method still has the problem of time-consuming calculation and cannot meet the real-time requirements in many applications. The disadvantage of methods based on SIFT-like features is that the memory overhead is large during calculation, and the extracted features occupy a large space; secondly, this type of feature has been proven to be unstable, and two fingerprint images of the same finger collected at a long interval may not be recognized. The main disadvantage of the ridge feature-based method is that it requires a relatively large amount of computation to extract a dense set of ridge points. Secondly, it has poor robustness under large elastic deformation conditions. Summary of the invention

[0004] Based on this, it is necessary to provide a small area fingerprint feature extraction method based on deep learning to address the existing problems.

[0005] In a first aspect, an embodiment of the present application provides a method for extracting small-area fingerprint features based on deep learning, the method comprising:

[0006] S1: construct a fingerprint feature extraction network, and collect a first fingerprint image and a second fingerprint image; the second fingerprint image is a transformed image corresponding to the first fingerprint image; the output of the fingerprint feature extraction network is a fingerprint feature;

[0007] S2: inputting the first fingerprint image into the fingerprint feature extraction network to obtain a first fingerprint feature; inputting the second fingerprint image into the fingerprint feature extraction network to obtain a second fingerprint feature; calculating a target loss function according to the first fingerprint feature and the second fingerprint feature; and training the fingerprint feature extraction network according to the target loss function;

[0008] S3: Collect a small-area fingerprint image to be extracted, input the small-area fingerprint image to be extracted into the trained fingerprint feature extraction network, and output a third fingerprint feature.

[0009] Preferably, the fingerprint feature extraction network includes a convolution layer, a feature vector layer and a feature score layer; the convolution layer includes seven layers; the seven convolution layers are used to transform a one-dimensional small-area fingerprint image into a 128-dimensional small-area fingerprint image; the feature vector layer is used to output a feature vector of the 128-dimensional small-area fingerprint image; and the feature score layer is used to output a reliable score of the 128-dimensional small-area fingerprint image.

[0010] Preferably, the fingerprint feature includes a feature vector and a reliability score; the feature vector is a feature vector corresponding to a pixel point in the fingerprint image, and the feature vector is recorded as: The reliability score is the confidence that the pixel point in the fingerprint image is a feature point, and the reliability score is recorded as: S∈[0,1] H×W ;in, represents a real number set, H×W represents the size of the small area fingerprint image, and D represents the dimension of the feature vector.

[0011] Preferably, according to the first fingerprint feature and the second fingerprint feature, a final loss of a reliable score and a final loss of a feature vector are calculated; a weight is added to the final loss of the feature vector, and the final loss of the reliable score and the final loss of the feature vector with the added weight are added to obtain the target loss function; the expression of the target loss function is:

[0012]

[0013] in, represents the target loss function; represents the final loss of reliable score; λ d represents the first weight; represents the final loss of the feature vector; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image.

[0014] Preferably, the process of obtaining the final loss of the reliability score is:

[0015] According to the first fingerprint feature and the second fingerprint feature, a reliable score similarity loss function and a reliable score sparsity loss function are calculated; the reliable score similarity loss function and the reliable score sparsity loss function are weightedly combined to obtain a reliable score final loss; the expression of the reliable score final loss is:

[0016]

[0017] in, represents the final loss of reliable scores; represents the reliable score similarity loss function; λ p represents the second weight; represents a reliable score sparsity loss function of the first fingerprint image; represents the reliable fractional sparsity loss function of the second fingerprint image; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image.

[0018] Preferably, the process of obtaining the reliable score similarity loss function is:

[0019] The first fingerprint feature includes a first reliability score. The first reliability scores corresponding to a plurality of pixel points form a first reliability score map. The first reliability score map is denoted as: S 1 ;

[0020] The second fingerprint feature includes a second reliable score. The second reliable scores corresponding to a plurality of pixel points form a second reliable score map. The second reliable score map is denoted as: S 2 ;

[0021] The second reliable score map is transformed through the corresponding relationship to obtain a corresponding region score map, which is recorded as: Based on the first reliable score map and the corresponding region score map, a reliable score similarity loss function is calculated using cosine distance; the expression of the reliable score similarity loss function is:

[0022]

[0023] in, represents the reliable score similarity loss function; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the correspondence between the first fingerprint image and the second fingerprint image; P represents the first overlap area, which is the pixel point set on the first fingerprint image where the first fingerprint image and the second fingerprint image overlap; cosim() represents the cosine distance; S 1 [P] represents the reliable score distribution of the first overlap region on the first reliable score map; Represents the reliable score distribution of the first overlap area on the corresponding area score map.

[0024] Preferably, the process of obtaining the reliable score sparsity loss function of the first fingerprint image is:

[0025] Using the local maximum and the local mean, for each first pixel in the first overlap area, a first circular area with the first pixel as the center and a fixed threshold as the radius is used to calculate the reliable fractional sparsity, and obtain the reliable fractional sparsity loss function of the first fingerprint image; the expression of the reliable fractional sparsity loss function of the first fingerprint image is:

[0026]

[0027] in, represents the reliable score sparsity loss function of the first fingerprint image; P represents the first overlap area; p represents the first pixel point; represents the first circular area; represents the pixel point of the first overlapping area in the first fingerprint image; (i, j) represents the coordinates of the first pixel point;

[0028] The process of obtaining the reliable score sparsity loss function of the second fingerprint image is:

[0029] Using the local maximum and local mean, for each second pixel point in the second overlap area, a second circular area with the second pixel point as the center and a fixed threshold as the radius is used to calculate the reliable fractional sparsity loss function of the second fingerprint image; the second overlap area is a set of pixels on the second fingerprint image where the first fingerprint image and the second fingerprint image overlap; the expression of the reliable fractional sparsity loss function of the second fingerprint image is:

[0030]

[0031] in, represents the reliable score sparsity loss function of the second fingerprint image; P' represents the second overlap area; p' represents the second pixel point; represents the second circular area; represents the pixel point of the second overlapping area in the second fingerprint image; (i', j') represents the coordinates of the second pixel point.

[0032] Preferably, the process of obtaining the final loss of the feature vector is:

[0033] The first fingerprint feature includes a first feature vector. The first feature vectors corresponding to a plurality of pixel points form a first feature vector map. The first feature vector map is recorded as: 1 ;

[0034] The second fingerprint feature includes a second feature vector. The second feature vectors corresponding to a plurality of pixel points form a second feature vector map. The second feature vector map is recorded as: 2 ;

[0035] For the first eigenvector map and the second eigenvector map, a metric learning function is used to learn on the first reliable score map and the corresponding region score map to obtain the final loss of the eigenvector; the expression of the final loss of the eigenvector is:

[0036]

[0037] in, represents the final loss of the feature vector; P represents the first overlapping area; p represents the first pixel point; S 1 [p] represents the reliable score distribution of the first pixel in the first overlap region on the first reliable score map; represents the reliable score distribution of the first pixel in the first overlapped area on the corresponding area score map; m(X 1 ,X 2 ,U,p) represents the metric learning function; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image; X 1 represents the first eigenvector graph; X 2 Represents the second eigenvector map; the smaller the value of the metric learning function, the more similar the eigenvectors of the first reliable score map and the corresponding region score map at the corresponding position are.

[0038] Preferably, the metric learning function adopts a triangular metric function; the process of obtaining the metric learning function is:

[0039] The feature vector of the first pixel is recorded as: 1 [p]; the feature vector of the second pixel is recorded as:

[0040] X 2 [p'];

[0041] Then the expression of the positive sample is:

[0042] Z(X 1 ,X 2 ,U,p)=||X 1 [p]-X 2 [p']|| 2

[0043] Among them, Z(X 1 ,X 2 ,U,p) represents the positive sample; X 1 represents the first eigenvector graph; X2 represents the second feature vector image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image; p represents the first pixel point; p' represents the second pixel point;

[0044] For the first pixel, a pixel that is most similar to the feature vector of the first pixel after removing an area centered on the second pixel in the second fingerprint image is selected, and the pixel is recorded as: n 1 ;

[0045]

[0046] Wherein, t represents any pixel remaining after dividing an area centered on the second pixel in the second fingerprint image; I' represents the second fingerprint image; X 2 [t] represents the feature vector of pixel t;

[0047] And the pixel point t needs to meet the conditions; the conditional expression is:

[0048] ||t-p'|| ∞ >K

[0049] Where K represents the radius threshold;

[0050] For the second pixel, a pixel that is most similar to the feature vector of the second pixel after removing an area centered on the first pixel in the first fingerprint image is selected, and the pixel is recorded as: n 2 ;

[0051]

[0052] Wherein, t' represents any pixel remaining after removing an area centered on the first pixel in the first fingerprint image; X 1 [t'] represents the feature vector of pixel t';

[0053] And the pixel point t' needs to meet the condition; the conditional expression is:

[0054] ||t'-p|| ∞ >K

[0055] Then the expression of negative samples is:

[0056] N(X 1 ,X 2 ,U,p)=min(||X 1 [p]-X 2 [n 1 ]|| 2, ||X 1 [n 2 ]-X 2 [p']||2 )

[0057] Among them, N(X 1 ,X 2 ,U,p) represents negative samples;

[0058] The spacing threshold is introduced to obtain the metric learning function; the expression of the metric learning function is:

[0059] m(X 1 ,X 2 ,U,p)=max(0,M+Z(X 1 ,X 2 ,U,p) 2 -N(X 1 ,X 2 ,U,p) 2 )

[0060] Wherein, M represents the spacing threshold.

[0061] Preferably, the fingerprint feature extraction network is converged by adjusting the first weight and the second weight; and the converged fingerprint feature extraction network is trained for multiple rounds, wherein a training batch in each round includes multiple pairs of fingerprint images; a pair of fingerprint images includes the first fingerprint image and the second fingerprint image

[0062] Beneficial effects: The fingerprint features extracted by this method have high distinctiveness and stability, and have the characteristics of small storage space occupation, fast matching speed and high recognition accuracy; it promotes the application and development of small-area fingerprint recognition technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] By referring to the following drawings, the exemplary embodiments of the present invention can be more completely understood. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present invention and do not constitute a limitation of the present invention. In the drawings, the same reference numerals generally represent the same parts or steps.

[0064] Figure 1 The present invention provides a flowchart of a method according to an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0065] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0066] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by technicians in the field to which this application belongs.

[0067] In addition, the terms "first" and "second" etc. are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0068] The embodiment of the present application provides a small-area fingerprint feature extraction method based on deep learning, which is described below in conjunction with the accompanying drawings.

[0069] Please refer to Figure 1 , which shows a flowchart of a small area fingerprint feature extraction method based on deep learning provided by some embodiments of the present application. As shown in the figure, the method may include the following steps:

[0070] S1: construct a fingerprint feature extraction network, and collect a first fingerprint image and a second fingerprint image; the first fingerprint image is a small-area fingerprint image; the second fingerprint image is a transformed image corresponding to the first fingerprint image; the output of the fingerprint feature extraction network is a fingerprint feature;

[0071] Specifically,

[0072] The fingerprint feature extraction network includes convolution layer, feature vector layer and feature score layer.

[0073] Table 1 is a table of structural parameters of each layer of the fingerprint feature extraction network;

[0074] Network layer index Input size Operator Number of output channels The first convolutional layer 1×H×W Conv3×3 32 The second convolutional layer 32×H×W Conv3×3 32 The third convolutional layer 32×H×W Conv3×3 64 The fourth convolutional layer 64×H×W Conv3×3 64 The fifth convolutional layer 64×H×W Conv3×3 128 The sixth convolutional layer 128×H×W Conv3×3 128 The seventh convolutional layer 128×H×W Conv2×2 128 Feature Vector Layer 128×H×W L2 Normalization Operation 128 Feature score layer 128×H×W Pixel-by-pixel square + Conv1×1 1

[0075] As can be seen from Table 1, the parameters of each layer structure in the fingerprint feature extraction network are described. As shown in Table 1, the convolution layer includes seven layers; the seven convolution layers are used to transform a one-dimensional small-area fingerprint image into a 128-dimensional small-area fingerprint image; the feature vector layer outputs a feature vector of a 128-dimensional small-area fingerprint image through an L2 normalization operation; the feature score layer outputs a reliable score of a 128-dimensional small-area fingerprint image in a pixel-by-pixel square manner.

[0076] The fingerprint features include a feature vector and a reliability score; the feature vector is a feature vector corresponding to a pixel point in the fingerprint image, and the feature vector is recorded as: The reliability score is the confidence that the pixel point in the fingerprint image is a feature point, and the reliability score is recorded as: S∈[0,1] H×W ;in, represents a real number set, H×W represents the size of the small area fingerprint image, and D represents the dimension of the feature vector.

[0077] S2: inputting the first fingerprint image into the fingerprint feature extraction network to obtain a first fingerprint feature; inputting the second fingerprint image into the fingerprint feature extraction network to obtain a second fingerprint feature; calculating a target loss function according to the first fingerprint feature and the second fingerprint feature; and training the fingerprint feature extraction network according to the target loss function;

[0078] Specifically, according to the first fingerprint feature and the second fingerprint feature, a final loss of a reliable score and a final loss of a feature vector are calculated; a weight is added to the final loss of the feature vector, and the final loss of the reliable score and the final loss of the feature vector with the added weight are added to obtain the target loss function; the expression of the target loss function is:

[0079]

[0080] in, represents the target loss function; represents the final loss of reliable score; λ d represents the first weight; represents the final loss of the feature vector; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image.

[0081] The process of obtaining the final loss of the reliability score is:

[0082] According to the first fingerprint feature and the second fingerprint feature, a reliable score similarity loss function and a reliable score sparsity loss function are calculated; the reliable score similarity loss function and the reliable score sparsity loss function are weightedly combined to obtain a reliable score final loss; the expression of the reliable score final loss is:

[0083]

[0084] in, represents the final loss of reliable scores; represents the reliable score similarity loss function; λ p represents the second weight; represents a reliable score sparsity loss function of the first fingerprint image; represents the reliable fractional sparsity loss function of the second fingerprint image; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image.

[0085] The process of obtaining the reliable score similarity loss function is:

[0086] The first fingerprint feature includes a first reliability score. The first reliability scores corresponding to a plurality of pixel points form a first reliability score map. The first reliability score map is denoted as: S 1 ;

[0087] The second fingerprint feature includes a second reliable score. The second reliable scores corresponding to a plurality of pixel points form a second reliable score map. The second reliable score map is denoted as: S 2 ;

[0088] The second reliable score map is transformed through the corresponding relationship to obtain a corresponding region score map, which is recorded as: Based on the first reliable score map and the corresponding region score map, a reliable score similarity loss function is calculated using cosine distance; the expression of the reliable score similarity loss function is:

[0089]

[0090] in, represents the reliable score similarity loss function; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the correspondence between the first fingerprint image and the second fingerprint image; P represents the first overlap area, which is the pixel point set on the first fingerprint image where the first fingerprint image and the second fingerprint image overlap; cosim() represents the cosine distance; S 1 [P] represents the distribution of reliable scores of the first overlap region on the first reliable score map; Represents the reliable score distribution of the first overlap area on the corresponding area score map.

[0091] The process of obtaining the reliable score sparsity loss function of the first fingerprint image is:

[0092] Using the local maximum and the local mean, for each first pixel in the first overlap area, a first circular area with the first pixel as the center and a fixed threshold as the radius is used to calculate the reliable fractional sparsity, and obtain the reliable fractional sparsity loss function of the first fingerprint image; the expression of the reliable fractional sparsity loss function of the first fingerprint image is:

[0093]

[0094] in, represents the reliable score sparsity loss function of the first fingerprint image; P represents the first overlap area; p represents the first pixel point; represents the first circular area; represents the pixel point of the first overlapping area in the first fingerprint image; (i, j) represents the coordinates of the first pixel point;

[0095] The process of obtaining the reliable score sparsity loss function of the second fingerprint image is:

[0096] Using the local maximum and local mean, for each second pixel point in the second overlap area, a second circular area with the second pixel point as the center and a fixed threshold as the radius is used to calculate the reliable fractional sparsity loss function of the second fingerprint image; the second overlap area is a set of pixels on the second fingerprint image where the first fingerprint image and the second fingerprint image overlap; the expression of the reliable fractional sparsity loss function of the second fingerprint image is:

[0097]

[0098] in, represents the reliable score sparsity loss function of the second fingerprint image; P' represents the second overlap area; p' represents the second pixel point; represents the second circular area; represents the pixel point of the second overlapping area in the second fingerprint image; (i', j') represents the coordinates of the second pixel point.

[0099] In this embodiment, the reliable fractional sparsity is calculated in a circular area with the pixel as the center and the fixed threshold as the radius; in actual engineering, for the convenience of calculation, the reliable fractional sparsity can be calculated in a square area with the pixel as the center and the fixed threshold as half of the side length. It can be seen that the larger the fixed threshold, the greater the sparsity; if the fixed threshold is 0, it means that there is no requirement for sparsity; that is, selecting a suitable fixed threshold can make the features obtained by the final optimization solution meet the required sparsity requirements.

[0100] The process of obtaining the final loss of the feature vector is:

[0101] The first fingerprint feature includes a first feature vector. The first feature vectors corresponding to a plurality of pixel points form a first feature vector map. The first feature vector map is recorded as: 1 ;

[0102] The second fingerprint feature includes a second feature vector. The second feature vectors corresponding to a plurality of pixel points form a second feature vector map. The second feature vector map is recorded as: 2 ;

[0103] For the first eigenvector map and the second eigenvector map, a metric learning function is used to learn on the first reliable score map and the corresponding region score map to obtain the final loss of the eigenvector; the expression of the final loss of the eigenvector is:

[0104]

[0105] in, represents the final loss of the feature vector; P represents the first overlapping area; p represents the first pixel point; S 1 [p] represents the reliable score distribution of the first pixel in the first overlap region on the first reliable score map; represents the reliable score distribution of the first pixel in the first overlapped area on the corresponding area score map; m(X 1 ,X 2 ,U,p) represents the metric learning function; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image; X 1 represents the first eigenvector graph; X 2 Represents the second eigenvector map; the smaller the value of the metric learning function, the more similar the eigenvectors of the first reliable score map and the corresponding region score map at the corresponding position are.

[0106] The metric learning function adopts a triangular metric function; the process of obtaining the metric learning function is:

[0107] The feature vector of the first pixel is recorded as: 1 [p]; the feature vector of the second pixel is recorded as:

[0108] X 2 [p'];

[0109] Then the expression of the positive sample is:

[0110] Z(X 1 ,X 2 ,U,p)=||X 1 [p]-X 2 [p']|| 2

[0111] Among them, Z(X 1 ,X 2 ,U,p) represents the positive sample; X 1 represents the first eigenvector graph; X 2 represents the second feature vector image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image; p represents the first pixel point; p' represents the second pixel point;

[0112] For the first pixel, a pixel that is most similar to the feature vector of the first pixel after removing an area centered on the second pixel in the second fingerprint image is selected, and the pixel is recorded as: n 1 ;

[0113]

[0114] Wherein, t represents any pixel remaining after dividing an area centered on the second pixel in the second fingerprint image; I' represents the second fingerprint image; X 2 [t] represents the feature vector of pixel t;

[0115] And the pixel point t needs to meet the conditions; the conditional expression is:

[0116] ||t-p'|| ∞ >K

[0117] Wherein, K represents the radius threshold, and in this embodiment, the radius threshold is set to 5;

[0118] For the second pixel, a pixel that is most similar to the feature vector of the second pixel after removing an area centered on the first pixel in the first fingerprint image is selected, and the pixel is recorded as: n 2 ;

[0119]

[0120] Wherein, t' represents any pixel remaining after removing an area centered on the first pixel in the first fingerprint image; X 1 [t'] represents the feature vector of pixel t';

[0121] And the pixel point t' needs to meet the condition; the conditional expression is:

[0122] ||t'-p|| ∞ >K

[0123] Then the expression of negative samples is:

[0124] N(X 1 ,X 2 ,U,p)=min(‖X 1 [p]-X 2 [n 1 ]|| 2 ,||X 1 [n 2 ]-X 2 [p']|| 2 )

[0125] Among them, N(X 1 ,X 2,U,p) represents negative samples;

[0126] The spacing threshold is introduced to obtain the metric learning function; the expression of the metric learning function is:

[0127] m(X 1 ,X 2 ,U,p)=max(0,M+Z(X 1 ,X 2 ,U,p) 2 -N(X 1 ,X 2 ,U,p) 2 )

[0128] Wherein, M represents the spacing threshold.

[0129] The fingerprint feature extraction network is converged by adjusting the first weight and the second weight; and the converged fingerprint feature extraction network is trained for multiple rounds, wherein the training batch in each round contains multiple pairs of fingerprint images; a pair of fingerprint images includes the first fingerprint image and the second fingerprint image. In this embodiment, the weight decay is 0.0005; a total of 20 rounds of training are performed, and each training batch contains 64 pairs of 80×64 fingerprint images; the initial learning rate is 0.01, and the learning rate is multiplied by 0.1 after every 6 rounds.

[0130] S3: Collect a small-area fingerprint image to be extracted, input the small-area fingerprint image to be extracted into the trained fingerprint feature extraction network, and output a third fingerprint feature.

[0131] The fingerprint features extracted by this method can be used in smart door locks, smart padlocks, smart USB flash drives, electronic cigarettes and other products and can show excellent results.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and they should all be included in the scope of the claims and specification of the present application.

Claims

1. A small area fingerprint feature extraction method based on deep learning, characterized in that: include: S1: construct a fingerprint feature extraction network, and collect a first fingerprint image and a second fingerprint image; the second fingerprint image is a transformed image corresponding to the first fingerprint image; the output of the fingerprint feature extraction network is a fingerprint feature; S2: inputting the first fingerprint image into the fingerprint feature extraction network to obtain a first fingerprint feature; inputting the second fingerprint image into the fingerprint feature extraction network to obtain a second fingerprint feature; calculating a target loss function according to the first fingerprint feature and the second fingerprint feature; and training the fingerprint feature extraction network according to the target loss function; Calculate a final loss of a reliable score and a final loss of a feature vector according to the first fingerprint feature and the second fingerprint feature; A weight is added to the final loss of the feature vector, and the final loss of the reliable score is added to the final loss of the feature vector with the weight added to obtain the target loss function; the expression of the target loss function is: in, represents the target loss function; represents the final loss of reliable score; λ d represents the first weight; represents the final loss of the feature vector; I represents the first fingerprint image; I' represents the second fingerprint image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image; The process of obtaining the final loss of the reliability score is: According to the first fingerprint feature and the second fingerprint feature, a reliable score similarity loss function and a reliable score sparsity loss function are calculated; the reliable score similarity loss function and the reliable score sparsity loss function are weightedly combined to obtain a reliable score final loss; the expression of the reliable score final loss is: in, represents the reliable score similarity loss function; λ p represents the second weight; represents a reliable score sparsity loss function of the first fingerprint image; A reliable score sparsity loss function representing the second fingerprint image; The process of obtaining the reliable score similarity loss function is: The first fingerprint feature includes a first reliability score. The first reliability scores corresponding to a plurality of pixel points form a first reliability score map. The first reliability score map is denoted as: S 1 ; The second fingerprint feature includes a second reliable score. The second reliable scores corresponding to a plurality of pixel points form a second reliable score map. The second reliable score map is denoted as: S 2 ; The second reliable score map is transformed through the corresponding relationship to obtain a corresponding region score map, which is recorded as: Based on the first reliable score map and the corresponding region score map, a reliable score similarity loss function is calculated using cosine distance; the expression of the reliable score similarity loss function is: Wherein, P represents the first overlapping area, which is a set of pixel points on the first fingerprint image where the first fingerprint image and the second fingerprint image overlap; cosim() represents the cosine distance; S 1 [P] represents the distribution of reliable scores of the first overlap region on the first reliable score map; represents the reliable score distribution of the first overlap region on the corresponding region score map; The process of obtaining the reliable score sparsity loss function of the first fingerprint image is: Using the local maximum and the local mean, for each first pixel in the first overlap area, a first circular area with the first pixel as the center and a fixed threshold as the radius is used to calculate the reliable fractional sparsity, and obtain the reliable fractional sparsity loss function of the first fingerprint image; the expression of the reliable fractional sparsity loss function of the first fingerprint image is: Wherein, p represents the first pixel; represents the first circular area; represents the pixel point of the first overlapping area in the first fingerprint image; (i, j) represents the coordinates of the first pixel point; The process of obtaining the reliable score sparsity loss function of the second fingerprint image is: Using the local maximum and local mean, for each second pixel point in the second overlap area, a second circular area with the second pixel point as the center and a fixed threshold as the radius is used to calculate the reliable fractional sparsity loss function of the second fingerprint image; the second overlap area is a set of pixels on the second fingerprint image where the first fingerprint image and the second fingerprint image overlap; the expression of the reliable fractional sparsity loss function of the second fingerprint image is: Wherein, P' represents the second overlapping area; p' represents the second pixel point; represents the second circular area; represents the pixel point of the second overlapping area in the second fingerprint image; (i', j') represents the coordinates of the second pixel point; The process of obtaining the final loss of the feature vector is: The first fingerprint feature includes a first feature vector. The first feature vectors corresponding to a plurality of pixel points form a first feature vector map. The first feature vector map is recorded as: 1 ; The second fingerprint feature includes a second feature vector. The second feature vectors corresponding to a plurality of pixel points form a second feature vector map. The second feature vector map is recorded as: 2 ; For the first eigenvector map and the second eigenvector map, a metric learning function is used to learn on the first reliable score map and the corresponding region score map to obtain the final loss of the eigenvector; the expression of the final loss of the eigenvector is: Among them, S 1 [p] represents the reliable score distribution of the first pixel in the first overlap region on the first reliable score map; represents the reliable score distribution of the first pixel in the first overlapped area on the corresponding area score map; m(X 1 ,X 2 ,U,p) represents the metric learning function; X 1 represents the first eigenvector graph; X 2 represents the second eigenvector map; the smaller the value of the metric learning function, the more similar the eigenvectors of the first reliable score map and the corresponding region score map at the corresponding position are; S3: Collect a small-area fingerprint image to be extracted, input the small-area fingerprint image to be extracted into the trained fingerprint feature extraction network, and output a third fingerprint feature.

2. According to the method for extracting small-area fingerprint features based on deep learning in claim 1, it is characterized in that: The fingerprint feature extraction network includes a convolution layer, a feature vector layer and a feature score layer; the convolution layer includes seven layers; the seven convolution layers are used to transform a one-dimensional small-area fingerprint image into a 128-dimensional small-area fingerprint image; the feature vector layer is used to output a feature vector of the 128-dimensional small-area fingerprint image; the feature score layer is used to output a reliable score of the 128-dimensional small-area fingerprint image.

3. The method for extracting small-area fingerprint features based on deep learning according to claim 2, characterized in that: The fingerprint features include a feature vector and a reliability score; the feature vector is a feature vector corresponding to a pixel point in the fingerprint image, and the feature vector is recorded as: The reliability score is the confidence that the pixel point in the fingerprint image is a feature point, and the reliability score is recorded as: S∈[0,1] H×W ;in, represents a real number set, H×W represents the size of the small area fingerprint image, and D represents the dimension of the feature vector.

4. The method for extracting small-area fingerprint features based on deep learning according to claim 1, characterized in that: The metric learning function adopts a triangular metric function; the process of obtaining the metric learning function is: The feature vector of the first pixel is recorded as: 1 [p]; the feature vector of the second pixel is recorded as: X 2 [p']; Then the expression of the positive sample is: Z(X 1 ,X 2 ,U,p)=||X 1 [p]-X 2 [p']||2 Among them, Z(X 1 ,X 2 ,U,p) represents the positive sample; X 1 represents the first eigenvector graph; X 2 represents the second feature vector image; U represents the corresponding relationship between the first fingerprint image and the second fingerprint image; p represents the first pixel point; p' represents the second pixel point; For the first pixel, a pixel which is most similar to the feature vector of the first pixel after removing an area centered on the second pixel in the second fingerprint image is selected, and the pixel is recorded as: n1; Wherein, t represents any pixel remaining after removing an area centered on the second pixel in the second fingerprint image; I' represents the second fingerprint image; X 2 [t] represents the feature vector of pixel t; And the pixel point t needs to meet the conditions; the conditional expression is: ||t-p'|| ∞ >K Where K represents the radius threshold; For the second pixel, a pixel which is most similar to the feature vector of the second pixel after removing an area centered on the first pixel in the first fingerprint image is selected, and the pixel is recorded as: n2; Wherein, t' represents any pixel remaining after removing an area centered on the first pixel in the first fingerprint image; X 1 [t'] represents the feature vector of pixel t'; And the pixel point t' needs to meet the condition; the conditional expression is: ||t'-p|| ∞ >K Then the expression of negative samples is: N(X 1 ,X 2 ,U,p)=min(||X 1 [p]-X 2 [n1]||2,||X 1 [n2]-X 2 [p']||2) Among them, N(X 1 ,X 2 ,U,p) represents negative samples; The spacing threshold is introduced to obtain the metric learning function; the expression of the metric learning function is: m(X 1 ,X 2 ,U,p)=max(0,M+Z(X 1 ,X 2 ,U,p) 2 -N(X 1 ,X 2 ,U,p) 2 ) Wherein, M represents the spacing threshold.

5. The method for extracting small-area fingerprint features based on deep learning according to claim 4, characterized in that: The fingerprint feature extraction network is converged by adjusting the first weight and the second weight; and the converged fingerprint feature extraction network is trained for multiple rounds, wherein a training batch in each round includes multiple pairs of fingerprint images; a pair of fingerprint images includes the first fingerprint image and the second fingerprint image.

Images