CN109902585A - A kind of three modality fusion recognition methods of finger based on graph model - Google Patents
A kind of three modality fusion recognition methods of finger based on graph model Download PDFInfo
- Publication number
- CN109902585A CN109902585A CN201910084199.3A CN201910084199A CN109902585A CN 109902585 A CN109902585 A CN 109902585A CN 201910084199 A CN201910084199 A CN 201910084199A CN 109902585 A CN109902585 A CN 109902585A
- Authority
- CN
- China
- Prior art keywords
- finger
- mode
- feature
- enhancing image
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
A kind of three modality fusion recognition methods of finger based on graph model.It enhances original three modality images of finger including the use of Gabor filter and obtains three mode of finger enhancing image;The graph model for establishing three mode of finger enhancing image, obtains figure feature;The graph structure feature of three mode of finger enhancing image is merged using fused in tandem or coding amalgamation mode, obtains multi-feature vector;Using transfiniting learning machine algorithm and classify to the multi-feature vector of three mode of finger enhancing image.Finger three modality fusion recognition methods provided by the invention based on graph model can not only give full expression to the feature of three modality images of finger, with good feature representation ability, and can effectively solve the problem that the problem of three modality images are restricted due to the inconsistent caused fusion of size, it is on homemade three modal data library of finger the experimental results showed that this method has certain feasibility.
Description
Technical field
The invention belongs to living things feature recognition fields, know more particularly to a kind of three modality fusion of finger based on graph model
Other method.
Background technique
With the development of computer vision technique, biometrics identification technology obtains extensive hair in terms of authentication
Exhibition and application.Due to customer acceptance degree with higher and convenience, the authentication techniques based on biological characteristic are
Gradually replace conventional identification techniques.The finger of people includes a variety of biological characteristic mode, wherein fingerprint refers to vein and phalangeal configurations three
A mode can be realized identity identification.However, finger single mode biometrics identification technology research with application the result shows that, only
It only relies on finger single mode and carries out identity identification, be unable to satisfy the requirement of accuracy of identification and stability.
It is identified relative to single mode, the multi-modal fusion identification based on biological characteristic shows in terms of generalization and safety
Outstanding performance out.Three mode physical characteristics collecting of finger is convenient, and present position is more compact, identifies for three modality fusion of finger
Establish basis.However the inconsistency of three modality images size of finger, it is the key of the multi-modal feature effective integration of influence
Problem.Therefore, the three modality images feature representation model of finger for establishing universality, builds effective fusion recognition frame, is hand
Refer to the basis of three modality fusions identification.
Summary of the invention
To solve the above-mentioned problems, the purpose of the present invention is to provide a kind of, and three modality fusion of finger based on graph model is known
Other method.
In order to achieve the above object, the finger three modality fusion recognition methods provided by the invention based on graph model includes pressing
The following steps that sequence carries out:
(1) original three modality images of finger are enhanced using Gabor filter and obtains three mode enhancing figure of finger
Picture;
(2) graph model for establishing above-mentioned three mode of finger enhancing image, obtains figure feature;
(3) the graph structure feature of above-mentioned three mode of finger enhancing image is carried out using fused in tandem or coding amalgamation mode
Fusion, obtains multi-feature vector;
(4) classified using learning machine algorithm is transfinited to the multi-feature vector of above-mentioned three mode of finger enhancing image.
In step (1), described enhances original three modality images of finger using Gabor filter and obtains hand
Referring to the method for three mode enhancing image is:
Gabor filter operator definitions are as follows:
Wherein,θ is the direction of Gabor filter, f0It is the centre frequency of Gabor filter, σ and γ divide
The length-width ratio of standard deviation and oval Gaussian envelope is not represented, and v is a kind of DC response factor, xθ=xcos θ+ysin θ and yθ
=-xsin θ+ycos θ is coordinate x, and y rotates the value after θ degree;
Three mode of finger enhances image R (x, y) and passes through original three modality images I (x, y) of finger and Gabor operator G (x, y)
Convolution obtains, and is shown below:
Wherein,Indicate two-dimensional convolution.
In step (2), the graph model for establishing above-mentioned three mode of finger enhancing image obtains the specific of figure feature
Steps are as follows:
1) piecemeal processing is carried out to above-mentioned three mode of finger enhancing image, using each segment as a top in weighted graph
Point;
By the way of sliding window, whole finger mode enhancing image is traversed according to this, forms several segments.If finger
It is M × N that mode, which enhances picture size, and the size of sliding window is h × w, is distinguished with the right and lower sliding step for glide direction
For r and d, then the calculation formula of segment number is as follows:
Wherein, round () is to seek whole operator;
2) it is based on triangulation, the side collection of weighted graph is chosen according to the vertex of above-mentioned weighted graph, thereby determines that out and adds
The structure of weight graph;
Finger mode enhancing image is characterized using weighted graph, i.e., only consider current vertex and it is right, under, bottom right neighbours push up
Incidence relation between point;
3) it is based on Steerable filter, the oriented energy distribution characteristics of different segments is obtained, calculates side in weighted graph
Weight function, finally determine three mode of finger enhancing image figure feature;
The weight function of use is defined as follows:
w(vi,vj|eij∈ E)=W (vi)×S(vi,vj)
Wherein, W (vi) represent vertex viCorresponding segment feature, S (vi,vj) characterization adjacent vertex vi,vjCorresponding figure
Block feature similarity.
Firstly, segment feature W (vi) calculation method it is as follows:
Segment feature W (vi) selection generally require and determined by testing, provide the following two kinds here and choose mode:
Wherein, AAD is mean absolute deviation operator, BiIt is the enhancing result of i-th of vertex correspondence segment;
Then, segment characteristic similarity S (vi,vj) calculation method it is as follows:
By the segment of segmentation and Steerable filter convolution, the oriented energy distribution characteristics of segment is obtained, direction is passed through
Energy-distributing feature carries out vertex similarity calculation;Used Steerable filter and oriented energy distribution characteristics meter
It is as follows to calculate formula:
OED=E (1), E (2) ..., E (θ) ..., E (360) | I }
Wherein, f (x, y) is the basis filters group that trigonometric function is constituted, φjIt is the direction of Steerable filter, N
For the number of basic filter group f (x, y), k (θ) is interpolating function, and θ is the direction of Steerable filter;E (θ) is direction
Energy balane formula, i.e. image I and Steerable filter hθEnergy response on a direction θ, X, Y are Steerable
The size of filter and image I;The oriented energy of different directions θ is formed into a vector, obtains oriented energy distribution characteristics.
Oriented energy distribution characteristics can indicate the characteristic information on vertex in weighted graph, be in vector form;It is pushed up based on two
The drift gage nomogram block feature similarity of the oriented energy distribution characteristics of point, formula are as follows:
Wherein fi, fjThe oriented energy distribution characteristics of respectively segment i, j, L are segment i, the oriented energy distribution characteristics of j
Length, σ be segment characteristic similarity mean value;
Figure feature is indicated that dimension is equal to number of vertices, the weight function on element side between respective vertices by adjacency matrix.
In step (3), the utilization fused in tandem or coding amalgamation mode enhance image to above-mentioned three mode of finger
Graph structure feature merged, the method for obtaining multi-feature vector is:
1) fused in tandem mode
Fused in tandem mode is that the figure feature vector of three mode of finger enhancing image is directly cascaded and obtains comprehensive
Feature vector is closed, i.e., adjacency matrix is stitched together;
2) amalgamation mode is encoded
Coding amalgamation mode is by the way of Competition coding by the figure feature of three mode of finger enhancing image, i.e., adjacent square
Battle array permeates multi-feature vector;Specific method is: by each pixel coder coding vector [c of one 7 byte1,c2,
c3,c4,c5,c6,c7] indicate, each coding vector element will be by three Mode Coding pixels of the position and itself and surrounding pixel
Between relationship be determined;Consider the encoded pixels fusion of position (i, j) in the graph structure of three mode of finger enhancing image, picture
Element value is respectively by FVi,j, FPi,jAnd FKPi,jIt indicates.The volume of position (i, j) in the graph structure of three mode of finger enhancing image
In code pixel fusion vector, (c3,c5) position pair is constituted, it is obtained by the figure feature calculation of finger vein mode enhancing image;(c2,
c6) position pair is constituted, it is obtained by the figure feature calculation of fingerprint mode enhancing image;(c1,c7) position pair is constituted, by phalangeal configurations mould
The figure feature calculation that state enhances image obtains;Calculation formula is as follows:
Coding vector c4In can retain three mode of finger enhancing image graph structure pixel comparison information, calculation formula is such as
Under:
In the coding fusion of figure feature, the encoded pixels fusion value of position (i, j) is obtained using weighting scheme, is calculated public
Formula is as follows:
In step (4), the use transfinites learning machine algorithm to the comprehensive special of above-mentioned three mode of finger enhancing image
The method that sign vector is classified is:
Given one group of training sample comprising N number of variable, S={ (xi,ti)|xi∈Rn,ti∈Rm, i=1,2 ..., N }, then
It hasThe neural networks with single hidden layer of a node can be expressed as:
Wherein, ajAnd bjIt is learning parameter, βjIt is the weight parameter between j-th of concealed nodes and output variable, xiAnd tiPoint
The feature and target output of i-th of variable, G (a are not indicatedj,bj,xi) it is a non-linear piecewise continuous function, meet transfinite
The omnipotent approximation capability of habit machine;
AboveA equation can be write as matrix form:
H β=T
Wherein,It is called hidden layer output matrix.
The output of learning machine algorithm of transfiniting can be acquired by least-squares algorithm:
Wherein,It indicatesUnit matrix is tieed up, C is non-zero regularization factors.
Finger three modality fusion recognition methods provided by the invention based on graph model can not only give full expression to finger three
The feature of modality images has good feature representation ability, and can effectively solve the problem that three modality images since size is different
The problem of fusion caused by causing restricts, it is on homemade three modal data library of finger the experimental results showed that this method is with certain
Feasibility.
Detailed description of the invention
Fig. 1 is original three modality images of finger and enhancing image, wherein (a)-(c) is respectively original finger vein image, refers to
Vein refined image and finger vein enhance image;(d)-(f) is respectively that original fingerprint image, fingerprint thinning image and fingerprint increase
Strong image;(g)-(i) is respectively original finger joint print image, phalangeal configurations refined image and phalangeal configurations enhancing image.
Fig. 2 is the segment division rule that three mode of finger enhances image.
Fig. 3 is the weighted graph expression structure that the finger mode based on triangulation enhances image.
Fig. 4 is the graph structure expression example that finger mode enhances image, and (a) is segment division, (b) is the adjoining of graph structure
Matrix indicates.
Fig. 5 is fused in tandem method schematic diagram, is enhanced from top to bottom for finger vein, fingerprint and three mode of phalangeal configurations in (a)
Image enhances the graph structure of image (b) for corresponding three mode of finger, (c) enhances the figure of image for corresponding three mode of finger
The adjacency matrix of feature indicates, (d) enhances the figure feature vector of image for corresponding three mode of finger, (e) for after fused in tandem
Multi-feature vector.
Fig. 6 is coding fusion method schematic diagram.
Fig. 7 is the learning machine algorithm structure that transfinites.
Fig. 8 is that the fingerprint characteristic based on graph model identifies ROC curve.
Fig. 9 is that the finger vein pattern based on graph model identifies ROC curve.
Figure 10 is that the phalangeal configurations feature based on graph model identifies ROC curve.
Figure 11 is three modal characteristics graph structure settling time of finger.
Specific embodiment
Three modality fusion of finger identification to provided by the invention based on graph model in the following with reference to the drawings and specific embodiments
Method is described in detail.
Finger three modality fusion recognition methods provided by the invention based on graph model includes the following step carried out in order
It is rapid:
(1) original three modality images of finger are enhanced using Gabor filter and obtains three mode enhancing figure of finger
Picture;
Finger vein, fingerprint and three kinds of modality images of phalangeal configurations on finger have lines information abundant, but it is being acquired
In, easily it is affected by the external environment, it is lower so as to cause picture quality, and then accuracy of identification and recognition efficiency can be seriously affected.Figure
Image intensifying is the key that guarantee that finger biometric modal characteristics are accurately extracted and identified.In recent years, Gabor filter is widely used
In the fields such as image procossing, pattern-recognition and computer vision, frequency and direction indicate to be similar to human visual system
Frequency and direction expression.Gabor filter is sensitive for the image with textural characteristics, is suitable for finger biometric characteristic pattern
The enhancing of picture.
Gabor filter operator definitions are as follows:
Wherein,θ is the direction of Gabor filter, f0It is the centre frequency of Gabor filter, σ and γ divide
The length-width ratio of standard deviation (often referred to as scale) and oval Gaussian envelope is not represented, and v is a kind of DC response factor, xθ=
Xcos θ+ysin θ and yθ=-xsin θ+ycos θ is coordinate x, and y rotates the value after θ degree.Three mode of finger enhances image R (x, y)
It can be obtained, be shown below by original three modality images I (x, y) of finger and Gabor operator G (x, y) convolution:
Wherein,Indicate two-dimensional convolution.Fig. 1 is original three modality images of finger and enhancing image, wherein (a)-
(c) it is respectively original finger vein image, refers to vein refined image and refer to that vein enhances image;(d)-(f) is respectively original fingerprint
Image, fingerprint thinning image and fingerprint enhance image;(g)-(i) be respectively original finger joint print image, phalangeal configurations refined image and
Phalangeal configurations enhance image;
(2) graph model for establishing above-mentioned three mode of finger enhancing image, obtains figure feature;
Graph theory correlative study starts from 1736, is the important branch of science of mathematics research.Figure is commonly used to two objects of description
An or relationship between object, different parts.As a kind of effective Image Description Methods, figure can describe to scheme well
The association of privileged site as between.In the research of the past few decades, graph theory obtains successful application in area of pattern recognition.
The present invention characterizes the figure feature of three mode of finger enhancing image using weighted graph.Graph structure can solve well
Three mode of finger enhances graphical rule inconsistence problems, lays the foundation for fusion recognition.One weighted graph can be expressed as G (V,
E, W), wherein V indicates that vertex set, E indicate that side collection, W indicate the weight connection between side, also referred to as weight function.Vertex set
Selection it is most important, generally select vertex of the two kinds of characteristic point as figure: pixel and minutiae point.If selecting picture
The figure size on vertex of the vegetarian refreshments as figure, foundation is larger, and computation complexity is higher;Minutiae point refers to having in biological modality images
There is the point of mark meaning.For referring to vein image, lines bifurcation, endpoint can be used as minutiae point.The definition of minutiae point
Artificial dependence is stronger, and robustness is poor.
Specific step is as follows:
1) piecemeal processing is carried out to above-mentioned three mode of finger enhancing image, using each segment as a top in weighted graph
Point;
The graph structure that the present invention initially sets up three mode of finger enhancing image indicates, is determined and is added by the way of segment division
The vertex of weight graph.Fig. 2 illustrates the segment division rule of three mode of finger enhancing image, by the way of sliding window, according to this
Whole finger mode enhancing image is traversed, several segments are formed.If it is M × N, sliding window that finger mode, which enhances picture size,
Size be h × w, be respectively r and d with the right and lower sliding step for glide direction, then the calculation formula of segment number is as follows:
Wherein, round () is to seek whole operator.The present invention chooses each segment as a vertex in weighted graph,
Number of vertices in weighted graph can be not only reduced, moreover, can guarantee by adjusting the size and sliding step of sliding window
Three mode of finger enhances the segment that image obtains same number, i.e., is consistent in picture size, and then solve finger three
Mode enhances graphical rule inconsistence problems.
2) it is based on triangulation, the side collection of weighted graph is chosen according to the vertex of above-mentioned weighted graph, thereby determines that out and adds
The structure of weight graph;
After the vertex of weighted graph determines, it is thus necessary to determine that Qi Bianji.In order to reduce computation complexity, need to simplify the knot of figure
Structure complexity.In finger mode enhancing image, local feature zonal relevancy adjacent thereto is stronger, and with the increasing of distance
Greatly, gradually weaken with remote neighbouring region relevance.In the structure of weighted graph, current vertex region vertex adjacent thereto is only considered
Relationship, rather than adjacent vertex relationship is ignored.In addition, weighted graph can preferably characterize vertex relationship relative to non-directed graph,
Therefore the present invention characterizes finger mode enhancing image using weighted graph.Fig. 3 illustrates the finger mode based on triangulation
Enhance the weighted graph expression structure of image, i.e., only consider current vertex and it is right, under, the incidence relation between the neighbours vertex of bottom right.
3) it is based on Steerable filter, the oriented energy distribution characteristics of different segments is obtained, calculates side in weighted graph
Weight function, finally determine three mode of finger enhancing image figure feature;
Weight function is used to describe the intension expression between vertex, i.e. relationship between vertex.Different weight functions and top
Point intension, corresponds to different types of image detail information, and the weight function that the present invention uses is defined as follows:
w(vi,vj|eij∈ E)=W (vi)×S(vi,vj)
Wherein, W (vi) represent vertex viCorresponding segment feature, S (vi,vj) characterization adjacent vertex vi,vjCorresponding figure
Block feature similarity.
Firstly, segment feature W (vi) calculation method it is as follows:
In the segment partition process of finger mode enhancing image, more blank segment may be generated or comprising a small amount of
The segment of texture.The segment characteristic similarity calculated based on blank segment and few texture segment does not have representativeness, is unfavorable for figure
Description of the model to image texture network.Segment feature W (v in weight functioni) introducing, can be used to screen out blank segment or
Segment comprising a small amount of texture, or weaken the influence of these segments.
Segment feature W (vi) selection generally require and determined by testing, the present invention provides the following two kinds and chooses mode:
Wherein, AAD is mean absolute deviation operator, BiIt is the enhancing result of i-th of vertex correspondence segment.
Then, segment characteristic similarity S (vi,vj) calculation method it is as follows:
The texture curvilinear structures that finger mode enhancing image is extracted using Steerable filter, carry out adjacent vertex
Segment characteristic similarity calculates.By the segment of segmentation and Steerable filter convolution, the oriented energy distribution for obtaining segment is special
Sign carries out vertex similarity calculation by oriented energy distribution characteristics.Steerable filter of the present invention and side
It is as follows to energy-distributing feature calculation formula:
OED=E (1), E (2) ..., E (θ) ..., E (360) | I }
Wherein, f (x, y) is the basis filters group that trigonometric function is constituted, φjIt is the direction of Steerable filter, N
For the number of basic filter group f (x, y), k (θ) is interpolating function, and θ is the direction of Steerable filter.E (θ) is direction
Energy balane formula, i.e. image I and Steerable filter hθEnergy response on a direction θ, X, Y are Steerable
The size of filter and image I.The oriented energy of different directions θ is formed into a vector, obtains oriented energy distribution characteristics.
Oriented energy distribution characteristics can indicate the characteristic information on vertex in weighted graph, be in vector form.The present invention is based on
The drift gage nomogram block feature similarity of the oriented energy distribution characteristics on two vertex, formula are as follows:
Wherein fi, fjThe oriented energy distribution characteristics of respectively segment i, j, L are segment i, the oriented energy distribution characteristics of j
Length, σ be segment characteristic similarity mean value.
By operating above, the graph Structure Representation of three mode of finger enhancing image can be obtained.
Figure feature can be indicated that dimension is equal to number of vertices, the weight on element side between respective vertices by adjacency matrix
Function.The graph structure that Fig. 4 illustrates finger mode enhancing image indicates example.In example, (a) is to enhance a finger mode
Image is divided into 3 × 3 segments, i.e. 9 vertex;Indicate there is association between two vertex;It (b) is the adjacency matrix of figure feature
It indicates, represents the graph structure feature of three mode of finger enhancing image.
(4) the graph structure feature of above-mentioned three mode of finger enhancing image is carried out using fused in tandem or coding amalgamation mode
Fusion, obtains multi-feature vector;
As described above, the figure feature of three mode of finger enhancing image can be indicated with adjacency matrix.Based on adjacency matrix, this hair
The bright figure Fusion Features mode of two kinds of three mode of finger enhancing images that proposes: fused in tandem mode and coding amalgamation mode.
1) fused in tandem mode
Fused in tandem mode is that the figure feature vector of three mode of finger enhancing image is directly cascaded and obtains comprehensive
Feature vector is closed, i.e., adjacency matrix is stitched together.In view of the foundation rule of figure, there is 0 value element in adjacency matrix,
Therefore adjacency matrix is sparse matrix, can be indicated by vector and (only retain non-zero element in vector), be called figure feature vector.Fig. 5
The implementation block diagram of fused in tandem is illustrated, from top to bottom to refer to that vein, fingerprint and three mode of phalangeal configurations enhance image in (a), (b)
Enhance the graph structure of image for corresponding three mode of finger, (c) enhances the neighbour of the figure feature of image for corresponding three mode of finger
Matrix expression is connect, (d) enhances the figure feature vector of image for corresponding three mode of finger, (e) is comprehensive special after fused in tandem
Levy vector.Although fused in tandem mode is simple, easy to implement, the splicing of figure feature vector will lead to multi-feature vector dimension
Become larger, to increase storage and calculate consumption.
2) amalgamation mode is encoded
Coding amalgamation mode is by the way of Competition coding by the figure feature of three mode of finger enhancing image, i.e., adjacent square
Battle array permeates multi-feature vector.Fig. 6 illustrates the implementation block diagram of coding amalgamation mode.Specific method is: by each picture
Coding vector [the c of element one 7 byte of coding1,c2,c3,c4,c5,c6,c7] indicate, each coding vector element will be by the position
The three Mode Coding pixels and its relationship between surrounding pixel set are determined.Consider the figure of three mode of finger enhancing image
The encoded pixels fusion of position (i, j) in structure, pixel value is respectively by FVi,j, FPi,jAnd FKPi,jIt indicates.In three mould of finger
State enhances in the encoded pixels fusion vector of position (i, j) in the graph structure of image, (c3,c5) position pair is constituted, by finger vein mould
The figure feature calculation that state enhances image obtains;(c2,c6) position pair is constituted, it is obtained by the figure feature calculation of fingerprint mode enhancing image
?;(c1,c7) position pair is constituted, it is obtained by the figure feature calculation of phalangeal configurations mode enhancing image.Calculation formula is as follows:
Coding vector c4In can retain three mode of finger enhancing image graph structure pixel comparison information, calculation formula is such as
Under:
In the coding fusion of figure feature, the encoded pixels fusion value of position (i, j) is obtained using weighting scheme, is calculated public
Formula is as follows:
(4) classified using learning machine algorithm is transfinited to the multi-feature vector of above-mentioned three mode of finger enhancing image;
After fused in tandem or coding fusion, three mode of finger enhances the figure feature integration of image as comprehensive characteristics change
Amount, therefore can connect any classifier and classify.In order to improve recognition efficiency and accuracy of identification, the present invention, which uses, to transfinite
Learning machine algorithm is classified.
The learning machine algorithm that transfinites belongs to the training method of neural networks with single hidden layer, it is characterized in that hidden layer weight is random
It chooses, and exports weight and acquired using least-squares algorithm, Fig. 7 illustrates the learning machine algorithm structure that transfinites.
Given one group of training sample comprising N number of variable, S={ (xi,ti)|xi∈Rn,ti∈Rm, i=1,2 ..., N }, then
It hasThe neural networks with single hidden layer of a node can be expressed as:
Wherein, ajAnd bjIt is learning parameter, βjIt is the weight parameter between j-th of concealed nodes and output variable, xiAnd tiPoint
The feature and target output of i-th of variable, G (a are not indicatedj,bj,xi) it is a non-linear piecewise continuous function, meet transfinite
The omnipotent approximation capability of habit machine.
AboveA equation can be write as matrix form:
H β=T
Wherein,It is called hidden layer output matrix.
The output of learning machine algorithm of transfiniting can be acquired by least-squares algorithm:
Wherein,It indicatesUnit matrix is tieed up, C is non-zero regularization factors.
In order to verify the effect of the method for the present invention, the present invention has carried out following experiment:
It has made three modal data library of finger by oneself, has acquired 585 finger-images altogether, three mode of each finger collection, often
A mode acquires ten images.
The figure feature of finger mode enhancing has outstanding ability to express, and the biological characteristic that can be individually used for the mode is known
Not.Fig. 8,9,10 are respectively shown under different segment sizes, fingerprint, the single modes feature identification such as refer to vein, phalangeal configurations
ROC curve.There it can be seen that segment size is smaller, the feature recognition capability of extraction is higher, waits the rate of mistake lower.Figure 11 is shown
Different segment brings calculate the times, and segment size is smaller, computation complexity is higher, and so as to cause calculating, the time is longer.
In order to verify the stability and robustness of this fusion identification method, mean value is increased for 20, side for test data
The white noise that difference is 10.Table 1 shows Experimental comparison results, there it can be seen that identifying compared to finger single mode, the present invention
The fusion identification method precision of offer is higher;Coding fusion is higher relative to fused in tandem discrimination;This fusion identification method is known
The other time is longer, but within an acceptable range.
1 finger single mode of table and multi-modal fusion recognition result compare
Claims (5)
1. a kind of three modality fusion recognition methods of finger based on graph model, it is characterised in that: the method includes in order
The following steps of progress:
(1) original three modality images of finger are enhanced using Gabor filter and obtains three mode of finger enhancing image;
(2) graph model for establishing above-mentioned three mode of finger enhancing image, obtains figure feature;
(3) the graph structure feature of above-mentioned three mode of finger enhancing image is melted using fused in tandem or coding amalgamation mode
It closes, obtains multi-feature vector;
(4) classified using learning machine algorithm is transfinited to the multi-feature vector of above-mentioned three mode of finger enhancing image.
2. the finger three modality fusion recognition methods according to claim 1 based on graph model, it is characterised in that: in step
(1) in, described enhances original three modality images of finger using Gabor filter and obtains three mode enhancing figure of finger
The method of picture is:
Gabor filter operator definitions are as follows:
Wherein,θ is the direction of Gabor filter, f0It is the centre frequency of Gabor filter, σ and γ are respectively represented
The length-width ratio of standard deviation and oval Gaussian envelope, v is a kind of DC response factor, xθ=xcos θ+ysin θ and yθ=-xsin θ
+ ycos θ is coordinate x, and y rotates the value after θ degree;
Three mode of finger enhances image R (x, y) and passes through original three modality images I (x, y) of finger and Gabor operator G (x, y) convolution
It obtains, is shown below:
Wherein,Indicate two-dimensional convolution.
3. the finger three modality fusion recognition methods according to claim 1 based on graph model, it is characterised in that: in step
(2) in, the graph model for establishing above-mentioned three mode of finger enhancing image, obtaining figure feature, specific step is as follows:
1) piecemeal processing is carried out to above-mentioned three mode of finger enhancing image, using each segment as a vertex in weighted graph;
By the way of sliding window, whole finger mode enhancing image is traversed according to this, forms several segments;If finger mode
Enhancing picture size is M × N, and the size of sliding window is h × w, with the right and lower sliding step for glide direction be respectively r and
D, then the calculation formula of segment number is as follows:
Wherein, round () is to seek whole operator;
2) it is based on triangulation, the side collection of weighted graph is chosen according to the vertex of above-mentioned weighted graph, thereby determines that out weighted graph
Structure;
Finger mode enhancing image is characterized using weighted graph, i.e., only consider current vertex and it is right, under, between the neighbours vertex of bottom right
Incidence relation;
3) it is based on Steerable filter, the oriented energy distribution characteristics of different segments is obtained, calculates the power on side in weighted graph
Value function finally determines the figure feature of three mode of finger enhancing image;
The weight function of use is defined as follows:
w(vi,vj|eij∈ E)=W (vi)×S(vi,vj)
Wherein, W (vi) represent vertex viCorresponding segment feature, S (vi,vj) characterization adjacent vertex vi,vjCorresponding segment is special
Levy similarity;
Firstly, segment feature W (vi) calculation method it is as follows:
Segment feature W (vi) selection generally require and determined by testing, provide the following two kinds here and choose mode:
Wherein, AAD is mean absolute deviation operator, BiIt is the enhancing result of i-th of vertex correspondence segment;
Then, segment characteristic similarity S (vi,vj) calculation method it is as follows:
By the segment of segmentation and Steerable filter convolution, the oriented energy distribution characteristics of segment is obtained, oriented energy is passed through
Distribution characteristics carries out vertex similarity calculation;Used Steerable filter and oriented energy distribution characteristics calculate public
Formula is as follows:
OED=E (1), E (2) ..., E (θ) ..., E (360) | I }
Wherein, f (x, y) is the basis filters group that trigonometric function is constituted, φjIt is the direction of Steerable filter, based on N
The number of filter group f (x, y), k (θ) are interpolating functions, and θ is the direction of Steerable filter;E (θ) is oriented energy meter
Calculate formula, i.e. image I and Steerable filter hθEnergy response on a direction θ, X, Y are Steerable filters
With the size of image I;The oriented energy of different directions θ is formed into a vector, obtains oriented energy distribution characteristics;
Oriented energy distribution characteristics can indicate the characteristic information on vertex in weighted graph, be in vector form;Based on two vertex
The drift gage nomogram block feature similarity of oriented energy distribution characteristics, formula are as follows:
Wherein fi, fjThe oriented energy distribution characteristics of respectively segment i, j, L are segment i, the length of the oriented energy distribution characteristics of j
Degree, σ are the mean value of segment characteristic similarity;
Figure feature is indicated that dimension is equal to number of vertices, the weight function on element side between respective vertices by adjacency matrix.
4. the finger three modality fusion recognition methods according to claim 1 based on graph model, it is characterised in that: in step
(3) in, it is described using fused in tandem or coding amalgamation mode to above-mentioned three mode of finger enhancing image graph structure feature into
Row fusion, the method for obtaining multi-feature vector is:
1) fused in tandem mode
Fused in tandem mode is that the figure feature vector of three mode of finger enhancing image is directly cascaded and obtains comprehensive spy
Vector is levied, i.e., adjacency matrix is stitched together;
2) amalgamation mode is encoded
Coding amalgamation mode is by the figure feature of three mode of finger enhancing image by the way of Competition coding, i.e. adjacency matrix melts
It is combined into a multi-feature vector;Specific method is: by each pixel coder coding vector [c of one 7 byte1,c2,c3,
c4,c5,c6,c7] indicate, each coding vector element by by three Mode Coding pixels of the position and its between surrounding pixel
Relationship be determined;Consider the encoded pixels fusion of position (i, j) in the graph structure of three mode of finger enhancing image, pixel
Value is respectively by FVi,j, FPi,jAnd FKPi,jIt indicates;The coding of position (i, j) in the graph structure of three mode of finger enhancing image
In pixel fusion vector, (c3,c5) position pair is constituted, it is obtained by the figure feature calculation of finger vein mode enhancing image;(c2,c6)
Position pair is constituted, is obtained by the figure feature calculation of fingerprint mode enhancing image;(c1,c7) position pair is constituted, increased by phalangeal configurations mode
The figure feature calculation of strong image obtains;Calculation formula is as follows:
Coding vector c4In can retain the graph structure pixel comparison information of three mode of finger enhancing image, calculation formula is as follows:
In the coding fusion of figure feature, the encoded pixels fusion value of position (i, j) is obtained using weighting scheme, and calculation formula is such as
Under:
5. the finger three modality fusion recognition methods according to claim 1 based on graph model, it is characterised in that: in step
(4) in, described classifies to the multi-feature vector of above-mentioned three mode of finger enhancing image using learning machine algorithm is transfinited
Method be:
Given one group of training sample comprising N number of variable, S={ (xi,ti)|xi∈Rn,ti∈Rm, i=1,2 ..., N }, then it hasThe neural networks with single hidden layer of a node can be expressed as:
Wherein, ajAnd bjIt is learning parameter, βjIt is the weight parameter between j-th of concealed nodes and output variable, xiAnd tiTable respectively
Show the feature and target output of i-th of variable, G (aj,bj,xi) it is a non-linear piecewise continuous function, meet the learning machine that transfinites
Omnipotent approximation capability;
AboveA equation can be write as matrix form:
H β=T
Wherein,It is called hidden layer output matrix;
The output of learning machine algorithm of transfiniting can be acquired by least-squares algorithm:
Wherein,It indicatesUnit matrix is tieed up, C is non-zero regularization factors.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910084199.3A CN109902585B (en) | 2019-01-29 | 2019-01-29 | Finger three-mode fusion recognition method based on graph model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910084199.3A CN109902585B (en) | 2019-01-29 | 2019-01-29 | Finger three-mode fusion recognition method based on graph model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109902585A true CN109902585A (en) | 2019-06-18 |
CN109902585B CN109902585B (en) | 2023-04-07 |
Family
ID=66944348
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910084199.3A Active CN109902585B (en) | 2019-01-29 | 2019-01-29 | Finger three-mode fusion recognition method based on graph model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109902585B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263726A (en) * | 2019-06-24 | 2019-09-20 | 山东浪潮人工智能研究院有限公司 | A kind of finger vein identification method and device based on depth correlation feature learning |
CN110516595A (en) * | 2019-08-27 | 2019-11-29 | 中国民航大学 | Finger multi-modal fusion recognition methods based on convolutional neural networks |
CN111967405A (en) * | 2020-08-19 | 2020-11-20 | 深圳职业技术学院 | Finger three-mode fusion identification method and device based on crystal graph structure |
CN112084938A (en) * | 2020-09-08 | 2020-12-15 | 哈尔滨工业大学(深圳) | Method and device for improving representation stability of plane target based on graph structure |
CN112381147A (en) * | 2020-11-16 | 2021-02-19 | 虎博网络技术(上海)有限公司 | Dynamic picture similarity model establishing method and device and similarity calculating method and device |
CN113591629A (en) * | 2021-07-16 | 2021-11-02 | 深圳职业技术学院 | Finger three-mode fusion recognition method, system, device and storage medium |
CN113963387A (en) * | 2021-10-12 | 2022-01-21 | 华南农业大学 | Finger multi-modal feature extraction and fusion method based on optimal coding bit |
CN114360007A (en) * | 2021-12-22 | 2022-04-15 | 浙江大华技术股份有限公司 | Face recognition model training method, face recognition device, face recognition equipment and medium |
CN113963387B (en) * | 2021-10-12 | 2024-04-19 | 华南农业大学 | Finger multi-mode feature extraction and fusion method based on optimal coding bits |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105095880A (en) * | 2015-08-20 | 2015-11-25 | 中国民航大学 | LGBP encoding-based finger multi-modal feature fusion method |
CN105138974A (en) * | 2015-08-12 | 2015-12-09 | 中国民航大学 | Gabor coding based finger multimodal feature fusion method |
CN105160305A (en) * | 2015-08-10 | 2015-12-16 | 中国民航大学 | Finger multi-mode characteristic fusion method |
CN107463952A (en) * | 2017-07-21 | 2017-12-12 | 清华大学 | A kind of object material sorting technique based on multi-modal fusion deep learning |
CN108596126A (en) * | 2018-04-28 | 2018-09-28 | 中国民航大学 | A kind of finger venous image recognition methods based on improved LGS weighted codings |
-
2019
- 2019-01-29 CN CN201910084199.3A patent/CN109902585B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105160305A (en) * | 2015-08-10 | 2015-12-16 | 中国民航大学 | Finger multi-mode characteristic fusion method |
CN105138974A (en) * | 2015-08-12 | 2015-12-09 | 中国民航大学 | Gabor coding based finger multimodal feature fusion method |
CN105095880A (en) * | 2015-08-20 | 2015-11-25 | 中国民航大学 | LGBP encoding-based finger multi-modal feature fusion method |
CN107463952A (en) * | 2017-07-21 | 2017-12-12 | 清华大学 | A kind of object material sorting technique based on multi-modal fusion deep learning |
CN108596126A (en) * | 2018-04-28 | 2018-09-28 | 中国民航大学 | A kind of finger venous image recognition methods based on improved LGS weighted codings |
Non-Patent Citations (3)
Title |
---|
SHUYI LI 等: "Finger Vein Recognition Based on Weighted Graph Structural Feature Encoding", 《SPRING》 * |
杨金锋 等: "嵌入式手指多模态特征同步采集系统设计", 《中国民航大学学报》 * |
胡锦丽 等: "基于判别相关分析法的手指多模态融合身份识别方法", 《桂林电子科技大学学报》 * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110263726B (en) * | 2019-06-24 | 2021-02-02 | 浪潮集团有限公司 | Finger vein identification method and device based on deep correlation feature learning |
CN110263726A (en) * | 2019-06-24 | 2019-09-20 | 山东浪潮人工智能研究院有限公司 | A kind of finger vein identification method and device based on depth correlation feature learning |
CN110516595A (en) * | 2019-08-27 | 2019-11-29 | 中国民航大学 | Finger multi-modal fusion recognition methods based on convolutional neural networks |
CN110516595B (en) * | 2019-08-27 | 2023-04-07 | 中国民航大学 | Finger multi-mode feature fusion recognition method based on convolutional neural network |
CN111967405B (en) * | 2020-08-19 | 2023-06-27 | 深圳职业技术学院 | Finger tri-modal fusion recognition method and device based on crystal diagram structure |
CN111967405A (en) * | 2020-08-19 | 2020-11-20 | 深圳职业技术学院 | Finger three-mode fusion identification method and device based on crystal graph structure |
CN112084938A (en) * | 2020-09-08 | 2020-12-15 | 哈尔滨工业大学(深圳) | Method and device for improving representation stability of plane target based on graph structure |
CN112084938B (en) * | 2020-09-08 | 2023-07-28 | 哈尔滨工业大学(深圳) | Method and device for improving stability of plane target representation based on graph structure |
CN112381147A (en) * | 2020-11-16 | 2021-02-19 | 虎博网络技术(上海)有限公司 | Dynamic picture similarity model establishing method and device and similarity calculating method and device |
CN112381147B (en) * | 2020-11-16 | 2024-04-26 | 虎博网络技术(上海)有限公司 | Dynamic picture similarity model establishment and similarity calculation method and device |
CN113591629A (en) * | 2021-07-16 | 2021-11-02 | 深圳职业技术学院 | Finger three-mode fusion recognition method, system, device and storage medium |
CN113591629B (en) * | 2021-07-16 | 2023-06-27 | 深圳职业技术学院 | Finger tri-modal fusion recognition method, system, device and storage medium |
CN113963387A (en) * | 2021-10-12 | 2022-01-21 | 华南农业大学 | Finger multi-modal feature extraction and fusion method based on optimal coding bit |
CN113963387B (en) * | 2021-10-12 | 2024-04-19 | 华南农业大学 | Finger multi-mode feature extraction and fusion method based on optimal coding bits |
CN114360007A (en) * | 2021-12-22 | 2022-04-15 | 浙江大华技术股份有限公司 | Face recognition model training method, face recognition device, face recognition equipment and medium |
CN114360007B (en) * | 2021-12-22 | 2023-02-07 | 浙江大华技术股份有限公司 | Face recognition model training method, face recognition device, face recognition equipment and medium |
Also Published As
Publication number | Publication date |
---|---|
CN109902585B (en) | 2023-04-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109902585A (en) | A kind of three modality fusion recognition methods of finger based on graph model | |
CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
CN105869173B (en) | A kind of stereoscopic vision conspicuousness detection method | |
Tavallali et al. | Robust cascaded skin detector based on AdaBoost | |
CN103927511B (en) | image identification method based on difference feature description | |
CN106096561A (en) | Infrared pedestrian detection method based on image block degree of depth learning characteristic | |
CN105956560A (en) | Vehicle model identification method based on pooling multi-scale depth convolution characteristics | |
CN108681737B (en) | Method for extracting image features under complex illumination | |
CN104850825A (en) | Facial image face score calculating method based on convolutional neural network | |
CN111291604A (en) | Face attribute identification method, device, storage medium and processor | |
CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
CN106023257A (en) | Target tracking method based on rotor UAV platform | |
CN103514456A (en) | Image classification method and device based on compressed sensing multi-core learning | |
CN108520204A (en) | A kind of face identification method | |
CN112906550B (en) | Static gesture recognition method based on watershed transformation | |
CN106407978B (en) | Method for detecting salient object in unconstrained video by combining similarity degree | |
CN110516533A (en) | A kind of pedestrian based on depth measure discrimination method again | |
Zang et al. | Traffic lane detection using fully convolutional neural network | |
CN106611158A (en) | Method and equipment for obtaining human body 3D characteristic information | |
Hu et al. | RGB-D image multi-target detection method based on 3D DSF R-CNN | |
CN110909778B (en) | Image semantic feature matching method based on geometric consistency | |
CN114419732A (en) | HRNet human body posture identification method based on attention mechanism optimization | |
CN108090460B (en) | Weber multidirectional descriptor-based facial expression recognition feature extraction method | |
Pathak et al. | Multimodal eye biometric system based on contour based E-CNN and multi algorithmic feature extraction using SVBF matching | |
CN109165551B (en) | Expression recognition method for adaptively weighting and fusing significance structure tensor and LBP characteristics |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |