CN103440480B - Non-contact palmprint recognition method based on palmprint image registration - Google Patents

Abstract

The present invention relates to a kind of Non-contact palmprint recognition method based on palmprint image registration, be divided into registration and identify two stages.At registration phase, the prototype figure picture that user provides is carried out ROI extraction, ROI image extract contention code feature and is stored in property data base, ROI image being extracted after pretreatment SIFT feature simultaneously and is stored in property data base.At cognitive phase, the query image that user provides carries out the operation that identical ROI extraction, contention code feature extraction and SIFT feature are extracted.The SIFT feature deposited in the SIFT feature extracted and data base is mated, obtains the SIFT feature point of coupling.Obtained the matching degree of SIFT feature by the SIFT feature point mated simultaneously, and two kinds of matching degrees are carried out fusion obtain final matching degree and be used for authentication or identification.Solve the deformation problems of noncontact palmprint image, improve accuracy of identification.

Description

Non-contact palmprint recognition method based on palmprint image registration

Technical field

The present invention relates to a kind of Non-contact palmprint recognition method based on palmprint image registration.

Background technology

Personal recognition is the most emerging a kind of biometrics identification technology.It has consumers' acceptable degree relatively Good, the advantage that accuracy of identification is higher.

It is the most right that existing palmprint feature extraction and matching process typically require two width palmprint images to be matched Together, and solve palmprint image alignment most popular method be design a kind of special palm-print image capture equipment, adopting During collection image, palm is placed on equipment by user, uses auxiliary locator to limit finger and the activity of palm, To ensure that the palmprint image gathered can preferably align every time.And the greatest drawback of this mode is to be substantially reduced The consumers' acceptable degree of personal recognition.

The effective means solving this problem is to use cordless to gather palmprint image.But at non-contact capture In mode, owing to having lacked auxiliary locator, the palmprint image collected often exist significantly rotation, Yardstick, translation etc. convert.These conversion are extracted even across area-of-interest (ROI) and also are difficult to eliminate.

Summary of the invention

Based on above weak point, the present invention proposes a kind of noncontact personal recognition side based on palmprint image registration Method.

The technology used in the present invention is as follows: a kind of Non-contact palmprint recognition method based on palmprint image registration, It is divided into registration and identifies two stages.At registration phase, the prototype palmprint image that user provides is carried out palmmprint ROI extracts, and obtains palmmprint ROI image, extracts palmmprint contention code feature and be stored in spy on palmmprint ROI image Levy data base, deposit extracting Scale invariant features transform (SIFT) feature after palmmprint ROI image pretreatment simultaneously In property data base；At cognitive phase, the inquiry palmprint image that user provides carries out identical palmmprint ROI Extraction, palmmprint contention code feature extraction, pretreatment and SIFT feature extract operation, the SIFT that will extract The SIFT feature deposited in feature and data base is mated, and obtains the SIFT feature point of coupling, utilizes coupling SIFT feature point to inquiry palmprint image registrate, and registration palmprint image on extract palmmprint competition Code feature, mates palmmprint contention code feature with the palmmprint contention code feature deposited in data base, obtains Degree of joining, is obtained the matching degree of SIFT feature simultaneously, and two kinds of matching degrees is carried out by the SIFT feature point mated Merge and obtain final matching degree for authentication or identification.

The present invention also has a following technical characteristic:

1, palmprint image is filtered at operation by the described circular Gabor filter that ROI image pretreatment used Reason, reaches the purpose strengthened,

Circular Gabor filter is shown below:

$G(x,y)=\exp (-\frac{(x^2+y^2)}{2{\mathrm{\σ}}^2})\·\exp \left(2\mathrm{\πiF}\left(\sqrt{x^2+y^2}\right)\right)$

Palmprint image to be pre-treated is I, and enhancing process carries out convolution behaviour by I with circular Gabor filter Realize, be shown below:

R=I*G

Wherein, G is circular Gabor filter, and * represents convolution operation.

2, described SIFT feature is extracted and is divided into four steps:

(1), by using the Gaussian filter of different scale to carry out convolution algorithm structure with palmprint image Gaussian metric space；

(2), by the image of yardstick adjacent in Gaussian metric space is carried out additive operation, formed Gaussian difference (DoG) space, characteristic point is by detecting the Local Extremum in DoG space Determine；

(3), around characteristic point, select an image block, and calculate the gradient orientation histogram of this image block (HOG), using the direction corresponding to the maximum of HOG as the principal direction of this feature point；

(4), calculate description of characteristic point, first calculate the HOG of image block around characteristic point, then utilize The principal direction of this point, by HOG travel direction normalization, has just obtained description of this point.

3, described the SIFT feature deposited in the SIFT feature extracted and data base is mated, two SIFT If characteristic point describes son meets following condition, then it is assumed that the two SIFT feature point is coupling:

d_ij＜ t min (d_ik), k=1,2 ..., N.k ≠ j

Wherein d_ij=||p_i-q_j||²It is two characteristic points p_iAnd q_jThe Euclidean distance described between son.

4, the described SIFT feature point utilizing coupling registrates the SIFT into using coupling to inquiry palmprint image Characteristic point calculates single answering, and for eliminating the impact that calculating is singly answered by Mismatching point, uses stochastic sampling consistent (RANSAC) single answering is calculated after algorithm gets rid of exterior point；If the two width palmprint image SIFT that the match is successful Feature point pairs is less than 4, then cannot calculate single answering, the most directly think that two width images are from different classifications.

5, the contention code feature extraction of described palmmprint, contention code algorithm first by one group of Gabor filter to the palm Stricture of vagina ROI image is filtered, and the wave filter used is:

$\mathrm{\ψ}(x,y,\mathrm{\ω},\mathrm{\θ})=\frac{\mathrm{\ω}}{\sqrt{2\mathrm{\π}}\mathrm{\κ}}\exp (-\frac{{\mathrm{\ω}}^2}{{8\mathrm{\κ}}^2}({4x}^{\′2}+y^{\′2}))(\exp \left({\mathrm{i\ωx}}^{\′}\right)-\exp (-\frac{{\mathrm{\κ}}^2}{2}))$

Wherein x '=(x-x₀)cosθ+(y-y₀) sin θ, y '=-(x-x₀)sinθ+(y-y₀) cos θ, (x₀, y₀) it is the center of wave filter；ω and θ determines frequency and the direction of wave filter, and $\mathrm{\κ}=\sqrt{2\ln 2}((2^{\mathrm{\δ}}+1)/(2^{\mathrm{\δ}}-1)),\mathrm{\ω}=\mathrm{\κ}/\mathrm{\σ}.$ Pixel in image (x, character representation y) is:

$W(x,y)=\underset{j}{\arg }\min (I(x,y)*{\mathrm{\ψ}}_R(x,y,\mathrm{\ω},{\mathrm{\θ}}_j))$

Wherein, ψ_RFor the real part of wave filter ψ, * is convolution algorithm；θ_j=j π/6, j=0,1 ..., 5,

In contention code feature, the feature of each pixel is to respond the index value of the strongest filter direction, for adding These index values are encoded by the speed of fast characteristic matching, and the matched rule of contention code is expressed as:

$D(P,Q)=1-\frac{1}{3\mathrm{MN}}\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}\underset{i=0}{\overset{3}{\mathrm{\Σ}}}{P}_b^i(x,y)\⊗Q_b^i(x,y)$

WhereinFor logic XOR, M, N are the size of image.

6, two kinds of described matching degrees carry out matching degree and the matching degree of SIFT feature being fused to merge contention code feature As the standard of decision-making, with the further precision improving and identifying.Fused matching degree is expressed as:

d=wd_SIFT+(1-w)d_CompCode

Wherein, d_SIFTFor the matching degree of SIFT feature, for the number of the characteristic point of coupling, and normalize to [0,1]； d_CompCodeMatching degree for contention code feature；W is the weights of both control contribution degree.

The invention have the characteristics that

1. the circular Gabor filter of application strengthens palmprint image, to increase the number of the SIFT feature point extracted Amount.

2. utilize the SIFT feature point of coupling, theoretical according to Epipolar geometry, palmprint image is registrated, solves Contact by no means the deformation problems of palmprint image.

Carry out palmprint feature extraction and coupling on palmprint image the most after registration, improve accuracy of identification.

4. the matching degree of palm print characteristics and the matching degree of SIFT feature are merged, improve further and identify essence Degree.

Accompanying drawing explanation

Fig. 1 is the overall flow figure of the present invention

Fig. 2 is circular Gabor filter spatial domain response diagram；

Fig. 3 is to examine on the SIFT feature point and pretreated palmmprint ROI image detected on former palmmprint ROI image The SIFT feature point diagram measured；

Fig. 4 is palmprint image registration instance graph.

Detailed description of the invention

The present invention discloses a kind of Non-contact palmprint recognition method based on palmprint image registration, and application is the present invention propose Method carry out the process of personal recognition and can be divided into registration and identify two stages.At registration phase, user is carried The prototype figure picture of confession carries out ROI extraction, obtains ROI image.ROI image extracts contention code feature and deposits In property data base, ROI image is extracted after pretreatment SIFT feature simultaneously and is stored in property data base.? Cognitive phase, the query image that user provides is carried out identical ROI extract, contention code feature extraction and SIFT feature extractions etc. operate.The SIFT feature deposited in the SIFT feature extracted and data base is carried out Coupling, obtains the SIFT feature point of coupling.Query image is registrated by the SIFT feature point utilizing coupling, And on the image of registration, extract contention code feature.The contention code feature will deposited in contention code feature and data base Mate, obtain matching degree.Obtained the matching degree of SIFT feature by the SIFT feature point mated simultaneously, and Two kinds of matching degrees carry out fusion obtain final matching degree and be used for authentication or identification.

Embodiment 1

1. palmprint image pretreatment

The purpose of palmprint image pretreatment is to increase the quantity of the SIFT feature point extracted.SIFT feature is counted Measure the most, utilize its transformation model calculated the most accurate, be more conducive to palmprint image registration.The present invention uses Palmprint image is filtered operating the purpose reaching to strengthen by circular Gabor filter.Because circular Gabor Wave filter is a kind of isotropism wave filter, insensitive to direction change.Owing to noncontact palmprint image may be deposited Rotationally-varying, isotropism wave filter is therefore only used to can be only achieved the purpose of invariable rotary.

Circular Gabor filter is shown below:

$G(x,y)=\exp (-\frac{(x^2+y^2)}{2{\mathrm{\σ}}^2})\·\exp \left(2\mathrm{\πiF}\left(\sqrt{x^2+y^2}\right)\right)$

Its spatial domain responds as shown in Figure 2:

If palmprint image to be pre-treated is I, enhancing process carries out convolution by I with circular Gabor filter Operation realizes, and is shown below:

R=I*G

Wherein, G is circular Gabor filter, and * represents convolution operation.

2.SIFT feature extracting and matching

The extraction of SIFT feature can be divided into four steps:

(1) by using the Gaussian filter of different scale to carry out convolution algorithm structure with palmprint image Gaussian metric space.

(2) by the image of yardstick adjacent in Gaussian metric space is carried out additive operation, formed Gaussian difference (DoG) space.Characteristic point is determined by the Local Extremum in detection DoG space.

(3) a selected image block around characteristic point, and calculate the gradient orientation histogram of this image block (HOG).Using the direction corresponding to the maximum of HOG as the principal direction of this feature point.

(4) description of characteristic point is calculated.First calculate the HOG of image block around characteristic point, then utilize The principal direction changed the time, by HOG travel direction normalization, has just obtained description changed the time.

If two SIFT feature points describe the following condition of gestational edema foot, then it is assumed that the two SIFT feature point is Join:

d_ij<t min(d_ik), k=1,2 ..., N.k ≠ j

Wherein d_ij=||p_i-q_j||²It is two characteristic points p_iAnd q_jBetween Euclidean distance.Fig. 3 shows SIFT The testing result of characteristic point.The most left figure is direct testing result on original palmprint image, and right figure is in advance The testing result on palmprint image after process.

3. palmprint image registration

Theoretical according to Epipolar geometry, there are two planar objects of linear deformation, its conversion should be able to be come by single Describe.List should at least 4 pairs of corresponding point of two objects by deforming upon calculate.The present invention uses The SIFT feature point of coupling calculates single answering.For eliminating the impact that calculating is singly answered by Mismatching point, the present invention makes Single answering is calculated after getting rid of exterior point with RANSAC algorithm.

The homogeneous coordinates assuming certain pixel in piece image are [x, y, 1]^T, obtaining Dan Yinghou, its Corresponding point in another piece image can be calculated by following formula:

[x ', y ', c]=H. [x, y, 1]^T

Wherein, H is single answering.The homogeneous coordinates of result are converted to cartesian coordinate and just obtain the right of another piece image Answer point coordinates [x '/c, y '/c].

If the 2 width palmprint image SIFT feature that the match is successful points are to less than 4, then cannot calculate single answering, In this case, owing to the point of coupling is to less, the most directly think that two width images are from different classifications.Completely Palmprint image registration algorithm as follows:

Algorithm: palmprint image registration

Input: image I₁And I₂

Output: by I₂Result I after registration_r

(1). by I₁And I₂Carry out pretreatment

(2). at pretreated I₁And I₂Upper extraction SIFT feature is also mated

(3) if. coupling SIFT point right >=4

A () uses RANSAC algorithm to calculate and singly answers H

If b () H is not empty

I_r=H·I₂

C () otherwise

I_r=I₂

(4). otherwise

I_r=I₂

(5). return I_r。

Fig. 4 is the example of palmprint image registration.Wherein first it is classified as I₁, second is classified as I₂, the 3rd is classified as I₂ Result I after registration_r。

4. palmprint feature extraction with mate

Traditional palm print characteristics can be extracted for identifying on the palmprint image after being registered.The present invention extracts the palm The contention code feature of stricture of vagina.

Palmmprint ROI image is filtered by contention code algorithm first by one group of Gabor filter, is used Wave filter is:

$\mathrm{\&psi;}(x,y,\mathrm{\&omega;},\mathrm{\&theta;})=\frac{\mathrm{\&omega;}}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&kappa;}}\exp (-\frac{{\mathrm{\&omega;}}^2}{{8\mathrm{\&kappa;}}^2}({4x}^{\&prime;2}+y^{\&prime;2}))(\exp \left({\mathrm{i\&omega;x}}^{\&prime;}\right)-\exp (-\frac{{\mathrm{\&kappa;}}^2}{2}))$

Wherein x '=(x-x₀)cosθ+(y-y₀) sin θ, y '=-(x-x₀)sinθ+(y-y₀)cosθ。(x₀, y₀) it is The center of wave filter；ω and θ determines frequency and the direction of wave filter, and $\mathrm{\&kappa;}=\sqrt{2\ln 2}((2^{\mathrm{\&delta;}}+1)/(2^{\mathrm{\&delta;}}-1)),\mathrm{\&omega;}=\mathrm{\&kappa;}/\mathrm{\&sigma;}.$ Pixel in image (x, character representation y) is:

$W(x,y)=\underset{j}{\arg }\min (I(x,y)*{\mathrm{\&psi;}}_R(x,y,\mathrm{\&omega;},{\mathrm{\&theta;}}_j))$

Wherein, ψ_RFor the real part of wave filter ψ, * is convolution algorithm；θ_j=j π/6, j=0,1 ..., 5.

In contention code feature, the feature of each pixel is the index value of filter direction, for accelerating characteristic matching Speed, these index values are encoded.The matched rule of contention code can be expressed as:

$D(P,Q)=1-\frac{1}{3\mathrm{MN}}\underset{x=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{y=0}{\overset{N-1}{\mathrm{\&Sigma;}}}\underset{i=0}{\overset{3}{\mathrm{\&Sigma;}}}{p}_b^i(x,y)\&CircleTimes;Q_b^i(x,y)$

WhereinFor logic XOR, M, N are the size of image.

5. matching degree merges

In addition to the matching degree of contention code feature can be as the standard of decision-making, the quantity of the SIFT feature point of coupling Can also be used for decision-making.Both is merged by the present invention, with the further precision improving and identifying.Fused Matching degree be expressed as:

d=wd_SIFT+(1-w)d_CompCode

Wherein, d_SIFTFor the matching degree of SIFT feature, for the number of the characteristic point of coupling, and normalize to [0,1]； d_CompCodeMatching degree for contention code feature；W is the weights of both control contribution degree.

Claims (7)

1. a Non-contact palmprint recognition method based on palmprint image registration, is divided into registration and identifies two stages, It is characterized in that: at registration phase, the palmmprint prototype figure picture that user provides is carried out palmmprint ROI extraction, Obtain palmmprint ROI image, palmmprint ROI image extracts palmmprint contention code feature and is stored in characteristic Storehouse, is stored in property data base by extracting SIFT feature after palmmprint ROI image pretreatment simultaneously；Identifying rank Section, the palmmprint query image that user provides is carried out identical palmmprint ROI extract, palmmprint contention code special Levy extraction and SIFT feature extracts operation, the SIFT that will deposit in the SIFT feature extracted and data base Feature is mated, and obtains the SIFT feature point of coupling, utilizes the SIFT feature point of coupling to the inquiry palm Print image registrates, and extracts palmmprint contention code feature on the palmprint image of registration, by palmmprint contention code Feature is mated with the palmmprint contention code feature deposited in data base, obtains matching degree, simultaneously by mating SIFT feature point obtains the matching degree of SIFT feature, and two kinds of matching degrees are carried out fusion is finally mated Degree is for authentication or identification.

A kind of Non-contact palmprint recognition method based on palmprint image registration the most according to claim 1, it is special Levy and be: palmprint image is filtered by the described circular Gabor filter that ROI image pretreatment used Ripple operation processes, and reaches the purpose strengthened,

Circular Gabor filter is shown below:

$G(x,y)=\exp (-\frac{(x^2+y^2)}{2{\mathrm{\&sigma;}}^2})\&CenterDot;\exp \left(2\mathrm{\&pi;}iF\right(\sqrt{x^2+y^2}\left)\right)$

Wherein x, y represent image pixel coordinates, and i is imaginary unit, and σ is Gaussian function standard deviation, and F is frequency Parameter,

Palmprint image to be pre-treated is I, and enhancing process carries out convolution behaviour by I with circular Gabor filter Realize, be shown below:

R=I*G

Wherein, G is circular Gabor filter, and * represents convolution operation.

A kind of Non-contact palmprint recognition method based on palmprint image registration the most according to claim 1, it is special Levy and be: described SIFT feature is extracted and is divided into four steps:

(1), by using the Gaussian filter of different scale to carry out convolution algorithm structure with palmprint image Gaussian metric space；

(2), by the image of yardstick adjacent in Gaussian metric space is carried out additive operation, formed Gaussian difference DoG space, characteristic point is come by the Local Extremum in detection DoG space Determine；

(3), around characteristic point, select an image block, and calculate the gradient orientation histogram of this image block HOG, using the direction corresponding to the maximum of HOG as the principal direction of this feature point；

(4), calculate description of characteristic point, first calculate the HOG of image block around characteristic point, then profit By the principal direction of this point by HOG travel direction normalization, just obtain description changed the time.

A kind of Non-contact palmprint recognition method based on palmprint image registration the most according to claim 1, it is special Levy and be: described the SIFT feature deposited in the SIFT feature extracted and data base is carried out Joining, two SIFT feature points describe the following condition of gestational edema foot, it is believed that the two SIFT feature point is Join:

d_ij＜ t min (d_ik), k=1,2 ..., N；k≠j

Wherein d_ij=| | p_i-q_j||²It is two characteristic points p_iAnd q_jThe Euclidean distance described between son.

A kind of Non-contact palmprint recognition method based on palmprint image registration the most according to claim 1, it is special Levy and be: inquiry palmprint image is registrated as using coupling by the described SIFT feature point utilizing coupling SIFT feature point calculate single should, for eliminating the impact that calculating is singly answered by Mismatching point, use RANSAC algorithm calculates single answering after getting rid of exterior point；If the two width palmprint image SIFT that the match is successful are special Levy a little to less than 4, then cannot calculate single answering, the most directly think that two width images are from different classifications.

A kind of Non-contact palmprint recognition method based on palmprint image registration the most according to claim 1, it is special Levying and be: the contention code feature extraction of described palmmprint, contention code algorithm is filtered first by one group of Gabor Palmmprint ROI image is filtered by ripple device, and the wave filter used is:

$\mathrm{\&psi;}(x,y,\mathrm{\&omega;},\mathrm{\&theta;})=\frac{\mathrm{\&omega;}}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&kappa;}}\exp (-\frac{{\mathrm{\&omega;}}^2}{8{\mathrm{\&kappa;}}^2}(4x^{\&prime;2}+y^{\&prime;2}\left)\right)\left(\exp \right({\mathrm{i\&omega;x}}^{\&prime;})-\exp (-\frac{{\mathrm{\&kappa;}}^2}{2}\left)\right)$

Wherein x '=(x-x₀)cosθ+(y-y₀) sin θ, y '=-(x-x₀)sinθ+(y-y₀) cos θ, (x₀, y₀) it is the center of wave filter；ω and θ determines frequency and the direction of wave filter, andδ is frequency response parameter, ω=κ/σ, pixel in image (x, character representation y) is:

$W(x,y)=\underset{j}{\arg }min\left(I\right(x,y)*{\mathrm{\&psi;}}_R(x,y,\mathrm{\&omega;},{\mathrm{\&theta;}}_j\left)\right)$

Wherein, ψ_RFor the real part of wave filter ψ, * is convolution algorithm；θ_j=j π/6, j=0,1 ..., 5, In contention code feature, the feature of each pixel is the index value of filter direction, for accelerating characteristic matching Speed, these index values are encoded, the matched rule of contention code is expressed as:

$D(P,Q)=1-\frac{1}{3MN}\underset{x=0}{\overset{M-1}{\&Sigma;}}\underset{y=0}{\overset{N-1}{\&Sigma;}}\underset{i=0}{\overset{3}{\&Sigma;}}{p}_b^i(x,y)\&CircleTimes;Q_b^i(x,y)$

WhereinFor logic XOR, M, N are the size of image, and P, Q represent the two width palms to be matched respectively The characteristic pattern of print image.

A kind of Non-contact palmprint recognition method based on palmprint image registration the most according to claim 1, it is special Levy and be: two kinds of described matching degrees carry out the SIFT of matching degree and the coupling being fused to contention code feature The quantity of characteristic point is as the standard of decision-making, with the further precision improving and identifying, fused matching degree It is expressed as:

D=wd_SIFT+(1-w)d_CompCode

Wherein, d_SIFTFor the matching degree of SIFT feature, for the number of the characteristic point of coupling, and normalize to [0,1]；d_CompCodeMatching degree for contention code feature；W is the weights of both control contribution degree.