CN101551857B - High-precise palm-print identifying arithmetic based on single matching fractional layer combination - Google Patents

Abstract

The invention discloses a high-precise palm-print identifying method based on single matching fractional layer combination. firstly, the invention proposes a new palm-print image pre-processing methodbased on moving image method to positioning and partitioning the palm-print image for speeding up the pre-processing of the image; Then distantly filtering the palm-print image by the filter Gabor; d irectly filtering the sampling point by integrating the filtering step and the sampling step so as to greatly reduce the feature extracting time. During the matching identification, the vein information of the palm-prints towards the same direction is single operated Hamming matching, so the method solves the position excursion problem of the vein features towards different directions caused by different filters. The combination between the matching fractional layer and the average plan maximum remains and makes use of the phase and the direction information of the vein feature. the invention greatly speeds up the pre-processing of the palm-print image and the extraction of the vein feature and furthest enhances the ID identifying precision of the palm-print.

Description

High-precise palm-print identifying arithmetic based on the single matching fractional layer fusion

Technical field

What the present invention relates to is a kind of identity identifying method based on biological characteristic, a kind of specifically high-precise palm-print identifying arithmetic that merges based on single matching fractional layer.

Technical background

In daily life and a lot of occasions such as finance, the administration of justice, safety check, ecommerce all need identification accurately.The mode of human identification mainly contains three kinds: first kind of method that is based on knowledge, as access to your password, password etc.; Second kind of method that is based on article is as using key, ID card etc.; The third is based on the method for the biological characteristic of human body, as people's face, fingerprint, palmmprint, voice etc.Preceding two kinds of methods exist a lot of defectives, based on the method for article carry inconvenience and lose easily, damage, stolen or forge; Method based on knowledge passes into silence easily, cracks etc.Personal identification method based on human body biological characteristics has overcome above-mentioned defective, use biometric solution be based on human body intrinsic feature, can not lose or forget.

As an emerging biological identification technology, palmmprint identification is compared with the other biological feature and mainly contained following advantage: the area of palmmprint is bigger, the information that includes than one piece of rich fingerprint many, therefore, palmmprint has than the better property distinguished of fingerprint theoretically.Simultaneously, the principal character of palmmprint is several main lines and the fold line on the palm, and therefore the information that provides identification required is provided the feature of extracting under low-resolution image, and the feature of extracting is not subject to interference of noise.Compare with iris feature, palm-print image capture equipment is simple, and cost is far below the collecting device of iris image.Compare with the hand-type feature, palm print characteristics is stable, and uniqueness is stronger, be difficult for forging, and accuracy of identification is higher.The palmmprint obtain manner is with criminal related less with lawsuit, so user's acceptance level is higher.

At present, the researchist has carried out more deep research to the living things feature recognition method based on palmmprint, and has obtained certain achievement.People such as Zhang are at United States Patent (USP) [Zhang Dapeng David, KongWai-Kin Adams, Method of palmprint identification, patent publication No.: US 2004/0057604A1] and [Zhang Dapeng David, Kong Wai-Kin Adams, Palm print identification usingpalm line orientation, patent publication No.: US 2005/0281438 A1] and Zhang[D.Zhang, W.Kong, J.You and M.Wong, " Online palmprint identification ", IEEETrans.Pattern Anal.Machine Intell, vol.25, no.9, pp.1041-1050,2003] the Gabor wave filter is used to extract the textural characteristics of palmprint image, is called PalmCode.But this algorithm is to the information of a direction only having adopted palmprint image, and other directional information is lost, and its palmmprint accuracy of identification is low.Kong[A.Kong, D.Zhang and M.Kame; " Palmprint identification using feature-levelfusion " .Pattern Recognition; vol.39; no.3; pp.478-487; 2006.] the PalmCode method is improved; adopt the Gabor wave filter of four direction, use first filtering again the method for sampling extract the textural characteristics of the palmprint image of four direction respectively, then by the maximal value fusion criterion, the maximal value of four direction feature of each pixel of getting textural characteristics promptly merges coding (FusionCode) as the value that merges the textural characteristics respective pixel.The computation complexity that this algorithm characteristics is extracted the stage obviously increases, and the four direction Feature Fusion is a feature, has lost the useful palmmprint information of part, thereby has been difficult to make full use of the phase place and the directional information of palmmprint texture.Simultaneously, because the existence of DC component in the Gabor wave filter makes the feature of its extraction be subjected to the influence of illumination bigger, make the accuracy of identification of palmprint authentication low.

Summary of the invention

The high-precise palm-print identifying arithmetic that the object of the present invention is to provide a kind of single matching fractional layer to merge, the computation complexity of this kind algorithm is low, and recognition speed is fast, the accuracy of identification height.

The object of the present invention is achieved like this: the high-precise palm-print identifying arithmetic that a kind of single matching fractional layer merges may further comprise the steps:

A, based on the palmprint image pre-service of mobile difference shadow method: the palmprint image that collects is cut out the palmprint image that comprises two angle points between middle finger and forefinger, the third finger and the little finger of toe, again the palmprint image that cuts out is carried out medium filtering and binary conversion treatment; With the palmprint image after the binaryzation respectively forward, upwards, 30 pixels of translation downwards, and be respectively 0 pixel to image rear portion, bottom, the top filling value of vacating after the translation respectively, obtain three images after the translation; With the image after three translations respectively with binaryzation after palmprint image subtract each other, the value of order smaller or equal to 0 pixel value be zero, the value of rest of pixels is 255, obtains three poor images; Three poor images are carried out and operation, obtain comprising between middle finger and forefinger, the third finger and the little finger of toe near the image in the space two angle points and two angle points, therefrom find out the coordinate of two angle points, again with the mid point of two angle points as initial point O, with the line of two angle points as ordinate T, set up new coordinate system O (Q, T); At new coordinate system O (Q, T) palmprint image to collecting in, along horizontal ordinate Q, in position (95 from initial point O 95 pixels, 0) be 128 * 128 for heartcut goes out size, and the horizontal edge rectangle palmprint image I parallel with horizontal ordinate (q, t), q=1,2,3...128, t=1,2,3...128;

B, based on the textural characteristics rapid extraction of interval filtering: adopt the direct current two-dimensional Gabor filter (q t) carries out 4 the filtering of being spaced apart of four direction, obtains the palmmprint textural characteristics of the four directions of 32 * 32 dimensions to 128 * 128 the rectangle palmprint image I in a step

$I_{\stackrel{\‾}{G}}(q^{\′},t^{\′},{\mathrm{\θ}}_k)=\underset{x=-n}{\overset{x=n}{\mathrm{\Σ}}}\underset{y=-n}{\overset{y=n}{\mathrm{\Σ}}}\stackrel{\‾}{G}(x,y,u,\mathrm{\σ},{\mathrm{\θ}}_k)\×I\left(4q^{\′}+x,{4t}^{\′}+y\right).$

Wherein q '=1,2,3...32, t '=1,2,3...32,4q '+x=q, 4t '+y=t; X, y are the coordinate of wave filter, and u is the frequency of filter function, and σ is the standard variance of direct current two-dimensional Gabor filter, θ _kBe the direction of direct current two-dimensional Gabor filter, k=1,2,3,4, n are the maximal value of absolute value of coordinate x, the y value of wave filter;

C, the palmprint authentication that merges based on independent Hamming coupling mark: carry out the operation in a～b step respectively for two palmprint images that collect, obtain the palmmprint textural characteristics of the four direction of two palmprint images

With

Palmmprint textural characteristics to two palmprint image equidirectionals calculates Hamming distance D ₀(θ _k):

$D_0\left({\mathrm{\&theta;}}_k\right)=\frac{\underset{q^{\&prime;}=1}{\overset{q^{\&prime;}=32}{\mathrm{\&Sigma;}}}\underset{t^{\&prime;}=1}{\overset{t^{\&prime;}=32}{\mathrm{\&Sigma;}}}\mathrm{<figure-callout\; id="1"\; label="I"\; filenames="H2009100592713E00022.png"\; state="\{\{state\}\}">I</figure-callout>}\frac{1}{\mathrm{GR}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)\&CircleTimes;I\frac{2}{\mathrm{GR}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)+\mathrm{<figure-callout\; id="1"\; label="I"\; filenames="H2009100592713E00022.png"\; state="\{\{state\}\}">I</figure-callout>}\frac{1}{\mathrm{GI}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)\&CircleTimes;I\frac{2}{\mathrm{GI}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)}{2\&times;32\&times;32}$

Wherein,

With The real part of representing two palmmprint textural characteristics respectively,

With

The imaginary part of representing two palmmprint textural characteristics respectively; The strategy that re-uses " with average " merges in the coupling fractional layer, thus the final fusion coupling mark D of the system that obtains ₀:

D ₀＝(D ₀(θ ₁)+D ₀(θ ₂)+D ₀(θ ₃)+D ₀(θ ₄))/4，

If merge coupling mark D ₀Less than the threshold value D that sets _tJudge two palmprint image couplings, otherwise judge that two palmprint images do not match.

Compared with prior art, income effect of the present invention is:

1, cuts apart in the location that the palmmprint pretreatment stage has adopted mobile difference shadow method to be used for palmmprint, this method is only carried out three displacements to image, three subtractions, twice and operation, do not use the multiplication and the convolution algorithm of length consuming time, reduce computation complexity, accelerated the pre-treating speed of palmprint image.

2, when feature extraction, sampled point is directly carried out Gabor filtering, filtering and two steps of down-sampling have been combined into step of filtering at interval, significantly reduced calculated amount, make the feature extraction time only be first filtering again the method for sampling 1/16, thereby make that its recognition speed is fast; The palmmprint information that is used at interval filtering simultaneously and first filtering the palmmprint information of the method for sampling again are the same, so the or not reduction of palmmprint accuracy of identification.

3, at the coupling cognitive phase, the present invention has carried out independent Hamming coupling to the palmmprint texture information of each direction, has solved the position offset problem that a plurality of direction textural characteristics bring because of different wave filters.Merge in employing of coupling fractional layer and average strategy, effectively keep and used the phase place and the directional information of the textural characteristics of whole four directions, thereby obtained very high accuracy of identification.

In a word, image pre-service of the present invention and texture feature extraction speed are fast, cut apart the location accurately, have made full use of all information of textural characteristics, the accuracy of identification height.

Emulation experiment also illustrates the accuracy of identification height of the inventive method: when misclassification rate FAR equals 10-5%, the correct receptance GAR of this algorithm is 99.73%, be higher than CompetitiveCode algorithm (Competitivecoding scheme for palmprint verification, in:Proceedings of the 17th InternationalConference on Pattern Recognition, 520-523) 98.50%, RLOC algorithm (Palmprintverification based on robust line orientation code, Pattern Recognition, vol.41, (the Ordinal palmprint representation forpersonal identification of 97.29% and OrdinalCode algorithm 1504-1513), in:Proceedings of IEEE International Conference onComputer Vision and Pattern Recognition, 279-284) 98.15%, PPOC algorithm (Fusionof phase and orientation information for palmprint authentication, 2005 IEEEInternational Conference on Image Processing, 29-32) 94.85%.This shows that this algorithm has higher accuracy of identification.

Above-mentioned direct current parameters of two-dimensional Gabor filter value is respectively: θ=15 °, 60 °, 105 °, 150 °, u=0.0916, σ=5.1679, n=8.Experiment showed, that choosing these parameter value palm print identity authentication systems can obtain higher accuracy of identification.

The present invention is further detailed explanation below in conjunction with accompanying drawing and concrete embodiment.

Description of drawings

Fig. 1 is the palmprint image preprocessing process of embodiment of the invention method.

Fig. 2 is the ROC performance curve of embodiment of the invention method and existing P POC, CompetitiveCode, RLOC, OrdinalCode algorithm.Horizontal ordinate among the figure is false acceptance rate (FAR), and ordinate is correct receptance (GAR).

Fig. 3 is the FAR of embodiment of the invention method, the FRR performance curve.

Embodiment

Embodiment

A kind of embodiment of the present invention is that a kind of high-precise palm-print identifying arithmetic that merges based on single matching fractional layer may further comprise the steps:

A, based on the palmprint image pre-service of mobile difference shadow method: the palmprint image that collects is cut out the palmprint image that comprises two angle points between middle finger and forefinger, the third finger and the little finger of toe, again the palmprint image that cuts out is carried out medium filtering and binary conversion treatment; With the palmprint image after the binaryzation respectively forward, upwards, 30 pixels of translation downwards, and be respectively 0 pixel to image rear portion, bottom, the top filling value of vacating after the translation respectively, obtain three images after the translation; With the image after three translations respectively with binaryzation after palmprint image subtract each other, the value of order smaller or equal to 0 pixel value be zero, the value of rest of pixels is 255, obtains three poor images; Three poor images are carried out and operation, obtain comprising between middle finger and forefinger, the third finger and the little finger of toe near the image in the space two angle points and two angle points, therefrom find out the coordinate of two angle points, again with the mid point of two angle points as initial point O, with the line of two angle points as ordinate T, set up new coordinate system O (Q, T); At new coordinate system O (Q, T) palmprint image to collecting in, along horizontal ordinate Q, in position (95 from initial point O 95 pixels, 0) be 128 * 128 for heartcut goes out size, and the horizontal edge rectangle palmprint image I parallel with horizontal ordinate (q, t), q=1,2,3...128, t=1,2,3...128;

B, based on the textural characteristics rapid extraction of interval filtering: adopt the direct current two-dimensional Gabor filter

(q t) carries out 4 the filtering of being spaced apart of four direction, obtains the palmmprint textural characteristics of the four directions of 32 * 32 dimensions to 128 * 128 the rectangle palmprint image I in a step

$I_{\stackrel{\&OverBar;}{G}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)=\underset{x=-n}{\overset{x=n}{\mathrm{\&Sigma;}}}\underset{y=-n}{\overset{y=n}{\mathrm{\&Sigma;}}}\stackrel{\&OverBar;}{G}(x,y,u,\mathrm{\&sigma;},{\mathrm{\&theta;}}_k)\&times;I\left(4q^{\&prime;}+x,{4t}^{\&prime;}+y\right).$

Wherein q '=1,2,3...32, t '=1,2,3...32,4q '+x=q, 4t '+y=t; X, y are the coordinate of wave filter, and u is the frequency of filter function, and σ is the standard variance of direct current two-dimensional Gabor filter, θ _kBe the direction of direct current two-dimensional Gabor filter, k=1,2,3,4, n are the maximal value of absolute value of coordinate x, the y value of wave filter, and the Gabor wave filter of present embodiment is:

$G(x,y,u,\mathrm{\&sigma;},{\mathrm{\&theta;}}_k)=\frac{1}{{2\mathrm{\&pi;\&sigma;}}^2}\exp (-\frac{x^2+y^2}{{2\mathrm{\&sigma;}}^2})\exp \left(2\mathrm{\&pi;ui}(x\cos {\mathrm{\&theta;}}_k+y\sin {\mathrm{\&theta;}}_k)\right),$

Direct current Gabor wave filter

With Gabor wave filter G (x, y, u, σ, θ _k) transformational relation be:

$\stackrel{\&OverBar;}{G}(x,y,u,\mathrm{\&sigma;},{\mathrm{\&theta;}}_k)=G(x,y,u,\mathrm{\&sigma;},{\mathrm{\&theta;}}_k)-\frac{\underset{x=-n}{\overset{n}{\mathrm{\&Sigma;}}}\underset{y=-n}{\overset{n}{\mathrm{\&Sigma;}}}G(x,y,u,\mathrm{\&sigma;},{\mathrm{\&theta;}}_k)}{{(2n+1)}^2};$

C, the palmprint authentication that merges based on independent Hamming coupling mark: carry out the operation in a～b step respectively for two palmprint images that collect, obtain the palmmprint textural characteristics of the four direction of two palmprint images

With

Palmmprint textural characteristics to two palmprint image equidirectionals calculates Hamming distance D ₀(θ _k):

$D_0\left({\mathrm{\&theta;}}_k\right)=\frac{\underset{q^{\&prime;}=1}{\overset{q^{\&prime;}=32}{\mathrm{\&Sigma;}}}\underset{t^{\&prime;}=1}{\overset{t^{\&prime;}=32}{\mathrm{\&Sigma;}}}\mathrm{<figure-callout\; id="1"\; label="I"\; filenames="H2009100592713E00022.png"\; state="\{\{state\}\}">I</figure-callout>}\frac{1}{\mathrm{GR}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)\&CircleTimes;I\frac{2}{\mathrm{GR}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)+\mathrm{<figure-callout\; id="1"\; label="I"\; filenames="H2009100592713E00022.png"\; state="\{\{state\}\}">I</figure-callout>}\frac{1}{\mathrm{GI}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)\&CircleTimes;I\frac{2}{\mathrm{GI}}(q^{\&prime;},t^{\&prime;},{\mathrm{\&theta;}}_k)}{2\&times;32\&times;32}$

Wherein,

With

The real part of representing two palmmprint textural characteristics respectively,

With

The imaginary part of representing two palmmprint textural characteristics respectively; The strategy that re-uses " with average " merges in the coupling fractional layer, thereby it is final to obtain system

D ₀＝(D ₀(θ ₁)+D ₀(θ ₂)+D ₀(θ ₃)+D ₀(θ ₄))/4，

If merge coupling mark D ₀Less than the threshold value D that sets _tJudge two palmprint image couplings, otherwise judge that two palmprint images do not match.

During enforcement, the threshold value D that the user can set the Hamming distance matching algorithm according to the actual requirement and the concrete condition of security of system _t

The computer simulation experiment of present embodiment method is as follows:

That use in the emulation experiment is the disclosed free palm print database PolyU PalmprintDatabase of The Hong Kong Polytechnic University (http://www4.comp.polyu.edu.hk/～biometrics/), this database comprises from 386 palms, and every palm is got about 20 images, totally 7752 palmprint images.These images are two phase acquisition of branch, and the average time interval of twice collection is 2 months, each palm are gathered about 10 images at every turn, and the size of image is 384 * 284 pixels.

In the emulation experiment of present embodiment, all samples in the database are mated identification (being that each sample standard deviation and other arbitrary sample mate identification) in twos.Two images from same palm are true coupling through what be identified as coupling, otherwise are false coupling.Carried out the inferior coupling in 30,042,876 (7752 * 7751/2) in the experiment altogether, wherein 74,086 times is true coupling, and remaining is false coupling.In order to compare the performance difference of this method and existing P POC algorithm, CompetitiveCode algorithm, RLOC algorithm and OrdinalCode algorithm, carried out of the in twos coupling identification of these algorithms simultaneously to all samples of this palm print database.Simulation result is as follows:

Usually (False Accept Rate, FAR) (Genuine Accept Rate GAR) weighs the performance of personal identification method with correct receptance by false acceptance rate.False acceptance rate (FAR) is meant system accepts the personator as validated user probability, and false acceptance rate (FAR) is low more good more; Correct acceptance rate (GAR) is meant the probability that system accepts validated user, and correct acceptance rate (GAR) is high more good more.These two performance index of FAR and GAR have reflected two different aspects of a biological recognition system.FAR is low more, and the received possibility of personator is low more, thereby the security of system is high more.GAR is high more, and the unaccepted possibility of validated user is low more, thereby the ease for use of system is good more.So the user should regulate FAR and GAR according to the different compromises that should be used for: for the security requirement higher system, such as some military system, safety is most important, therefore should reduce FAR, but reduce GAR simultaneously; Be not very high system to security requirement, such as a lot of civilian property system, ease for use is very important, at this moment should its ease for use be improved corresponding raising GAR, and FAR also can improve simultaneously, and security reduces.In order better to embody the relation between FAR and the GAR, and make things convenient for the mutual comparison between the algorithms of different, FAR under the common different threshold value and GAR form series of points in the two-dimensional coordinate system, and (FAR GAR), and is called the ROC curve with these curves that are drawn as in coordinate system.

Fig. 2 is the ROC curve and PPOC algorithm commonly used at present, CompetitiveCode algorithm, the ROC curve of RLOC algorithm and OrdinalCode algorithm at present embodiment method simulation result.Algorithm as can be seen from Figure 2 of the present invention is much better than additive method, and when misclassification rate FAR was identical, the correct receptance GAR of the inventive method all was higher than existing several algorithms commonly used.As when misclassification rate FAR equals 10-5%, the correct receptance of the inventive method is the highest, is 99.73%; CompetitiveCode algorithm commonly used is 98.50%, and the PPOC algorithm is 94.85%, the RLOC algorithm be 97.29% and the OrdinalCode algorithm be 98.15%.This shows that this algorithm has higher accuracy of identification.

Fig. 3 is the FAR of method simulation result of the present invention and the performance chart of FRR.The Hamming distance threshold value D of the inventive method as can be seen from Figure 3 ₀In the time of between 0.38-0.41, can obtain extremely low wrong acceptance rate (10 simultaneously ^-5% is following) and false rejection rate (below 0.27%).

Claims (2)

1. high-precise palm-print identifying arithmetic that merges based on single matching fractional layer may further comprise the steps:

A, based on the palmprint image pre-service of mobile difference shadow method: the palmprint image that collects is cut out the palmprint image that comprises two angle points between middle finger and forefinger, the third finger and the little finger of toe, again the palmprint image that cuts out is carried out medium filtering and binary conversion treatment; With the palmprint image after the binaryzation respectively forward, upwards, 30 pixels of translation downwards, and be respectively 0 pixel to image rear portion, bottom, the top filling value of vacating after the translation respectively, obtain three images after the translation; With the image after three translations respectively with binaryzation after palmprint image subtract each other, the value of order smaller or equal to 0 pixel value be zero, the value of rest of pixels is 255, obtains three poor images; Three poor images are carried out and operation, obtain comprising between middle finger and forefinger, the third finger and the little finger of toe near the image in the space two angle points and two angle points, therefrom find out the coordinate of two angle points, again with the mid point of two angle points as initial point O, with the line of two angle points as ordinate T, set up new coordinate system O (Q, T); At new coordinate system O (Q, T) palmprint image to collecting in, along horizontal ordinate Q, in position (95 from initial point O 95 pixels, 0) be 128 * 128 for heartcut goes out size, and the horizontal edge rectangle palmprint image I parallel with horizontal ordinate (q, t), q=1,2,3...128, t=1,2,3...128;

B, based on the textural characteristics rapid extraction of interval filtering: adopt the direct current two-dimensional Gabor filter

(q t) carries out 4 the filtering of being spaced apart of four direction, obtains the palmmprint textural characteristics of the four directions of 32 * 32 dimensions to 128 * 128 the rectangle palmprint image I in a step

Wherein q '=1,2,3...32, t '=1,2,3...32,4q '+x=q, 4t '+y=t; X, y are the coordinate of wave filter, and u is the frequency of filter function, and σ is the standard variance of direct current two-dimensional Gabor filter, θ _kBe the direction of direct current two-dimensional Gabor filter, k=1,2,3,4, n are the maximal value of absolute value of coordinate x, the y value of wave filter;

C, the palmprint authentication that merges based on independent Hamming coupling mark: carry out the operation in a～b step respectively for two palmprint images that collect, obtain the palmmprint textural characteristics of the four direction of two palmprint images

With

Palmmprint textural characteristics to two palmprint image equidirectionals calculates Hamming distance D ₀(θ _k):

Wherein,

With

The real part of representing two palmmprint textural characteristics respectively,

With

The imaginary part of representing two palmmprint textural characteristics respectively; The strategy that re-uses " with average " merges in the coupling fractional layer, thus the final fusion coupling mark D of the system that obtains ₀:

D ₀＝(D ₀(θ ₁)+D ₀(θ ₂)+D ₀(θ ₃)+D ₀(θ ₄))/4，

If merge coupling mark D ₀Less than the threshold value D that sets _tJudge two palmprint image couplings, otherwise judge that two palmprint images do not match.

2. the high-precise palm-print identifying arithmetic that merges based on single matching fractional layer according to claim 1 is characterized in that: described direct current two-dimensional Gabor filter G (x, y, u, σ, θ _k) parameter value be respectively: u=0.0916, σ=5.1679, n=8, θ=15 °, 60 °, 105 °, 150 °.

Images