Specific embodiment
To keep the technical problem to be solved in the present invention, technical solution and advantage clearer, below in conjunction with attached drawing and tool
Body embodiment is described in detail.
On the one hand, the present invention provides the method that a kind of pair of iris image is positioned, as shown in Figure 1, comprising:
Step S101: pupil coarse positioning is carried out to iris image.
In this step, a variety of methods can be used, coarse positioning is carried out to pupil, the position of pupil is positioned, then by pupil position
Appropriate expansion is set to get the rough position for arriving iris, there is employed herein radial symmetry transform methods.
Step S102: the coordinate of multiple key points is initialized on the basis of coarse positioning.
The initial position of key point is determined according to the result (center of circle and radius including pupil) of coarse positioning, we can lead to
It crosses and training after hand labeled key point is carried out to the iris image sample that shoots in advance obtains the initial relative position of key point, one
As take the average value of multiple samples, in order to describe and convenience of calculation, these coordinate values are generally described as 2n dimensional vector
Form, i.e.,N is the number of key point,For the real vector space of 2n dimension.
Step S103: the SIFT feature vector of key point is extracted.
SIFT, i.e. Scale invariant features transform (Scale-invariant feature transform, SIFT) are to use
In a kind of description of field of image processing.This description has scale invariability, can detect key point in the picture, be one
Kind local feature description;In this step, the SIFT feature vector of key point can be extracted by various methods.
Step S104: using SDM algorithm by the SIFT feature DUAL PROBLEMS OF VECTOR MAPPING of key point at increment of coordinate;Using SDM algorithm
Training obtains relevant parameter, by the SIFT feature DUAL PROBLEMS OF VECTOR MAPPING of key point at increment of coordinate.For convenience of each coordinate points of calculating
Increment of coordinate can be indicated with the column vector that 2n is tieed up, if the shape vector of prediction is expressed asReally
Shape vector is expressed asThen shape incremental representation is
The extraction of SIFT feature vector finishes, and needs that maps feature vectors to coordinate are calculated using suitable algorithm
Shape increment uses SDM algorithm here.Gradually unlike gradient descent method, Newton method, quasi-Newton method, simulated annealing, EM scheduling algorithm
More new strategy, SDM adhere to that a step updates in place, directly obtain optimal vector of the current point away from target point by supervised learning, calculate
Method complexity is low, and speed is fast.
Step S105: increment of coordinate is used, the coordinate value of key point is updated.
Step S106: carrying out curve fitting to key point using robust regression method, obtains upper eyelid parabola, palpebra inferior
The parameter of parabola and iris outer circle, iris profile include the parabola and iris outer circle in upper eyelid and palpebra inferior, certainly, when
When eyes are opened bigger, although upper eyelid or palpebra inferior may not intersect with iris outer circle, as shown in c and d in Fig. 3, upper eye
Each point of eyelid or palpebra inferior be overlapped (actually may each point suffer close, be extremely difficult to be completely coincident), but we according to
It so can use a parabola to be fitted to obtain effective iris region.
In the method positioned to iris image of the invention, pupil coarse positioning is carried out to iris image first, thick
The coordinate for initializing multiple key points after positioning on the basis of result on iris image (is chosen multiple key points, and is recorded
The coordinate value of key point);Then the SIFT feature vector of key point is extracted, and uses SDM algorithm by SIFT feature DUAL PROBLEMS OF VECTOR MAPPING
The coordinate value of key point is updated (i.e. by the coordinate value of increment of coordinate and key point using increment of coordinate at increment of coordinate
It is added), obtain the new coordinate value of key point;It is finally carried out curve fitting using robust regression method to key point, obtains eye
The parameter of eyelid parabola, palpebra inferior parabola and iris outer circle completes the positioning to iris image.
Compared with prior art, the method for the invention positioned to iris image is by the upper and lower eye in iris recognition
The parameter determination process of eyelid parabola and iris outer circle, regard as upper and lower eyelid and iris edge key point positioning and
Curve fit problem, clear thinking is simple and convenient, and speed is fast.The positioning of key point uses SIFT feature, fixed using SIFT feature
Position obtains several key points of upper and lower eyelid and iris edge, and locating speed is fast, precision is high.Using SDM algorithm by key point
SIFT feature DUAL PROBLEMS OF VECTOR MAPPING is at increment of coordinate and uses increment of coordinate, is updated to the coordinate value of key point, algorithm complexity
Low, speed is fast.Using the parameter of robust regression parabola of fit and circle, completes upper and lower eyelid boundary and iris is outer peripheral fixed
Position process, robustness are good.And the parameter of this method is independently stablized, strong antijamming capability steady to conditions such as illumination, and extensive
Property it is high, for example, traditional localization method is directed to different image libraries, need to adjust suitable parameter, and method of the invention is not
Adjusting parameter is needed, parameter is obtained by training, all equally applicable for different image libraries.
Therefore the method for the invention positioned to iris image is simple and convenient, speed is fast;Strong antijamming capability, precision
It is high;Generalization is high;Robustness is good.
In the present invention, can also the coordinate value to key point repeatedly updated (successive ignition), at this point, step S105 it
Afterwards, before step S106 further include:
Step S1051: judging whether to meet termination condition, if it is not, going to the SIFT feature vector (step for extracting key point
S103);
The present invention calculates the coordinate value of key point using iterative method, and step S103 to step S105 claims an iteration, passes through
Repeatedly (setting value) iteration can be avoided an iteration and fall into local better solution, obtain the coordinate value of stable key point;At this
In by fixed number of iterations as termination condition.The present invention further improves the accuracy of positioning.
As a kind of improvement of the method for the invention positioned to iris image, extract the SIFT feature of key point to
Measuring (step S103) includes:
Step S1031: the scale space of the iris image after construction pupil coarse positioning;
Convolution is carried out to the iris image after pupil coarse positioning using Gaussian kernel, obtains scale space;
Step S1032: in the case where existing scale space and key point coordinate, the main gradient direction of key point is determined;
In the case where existing scale space and key point coordinate, in crucial neighborhood of a point (adjacent domain), for ruler
The each pixel of degree spatially, calculates amplitude and the direction of its gradient;Then centered on key point, using in certain area
Gradient direction constitute the direction histogram of this key point, the range of direction histogram is 0~360 degree, if per mass dryness fraction being one
Column, the corresponding direction of every column are the gradient direction of pixel, the length of the column gradient magnitude after the adduction of direction thus.Choose part
Gradient orientation histogram peak-peak direction is the principal direction of key point, peak value place side of 0.8 times greater than peak-peak
To also as the direction of key point, Gauss weighting, the principal direction obtained in this way can be carried out to the gradient magnitude of crucial vertex neighborhood
It is relatively stable.
Step S1033: carrying out rotation to reference axis keeps the direction of reference axis identical as main gradient direction;Guarantee to send out in image
In the case where raw rotation, feature vector still has invariance;
Step S1034: under new reference axis, centered on key point, adjacent domain appropriate is selected, and be divided into
Several sub-regions extract the SIFT feature vector of key point.
The above method ensure that the rotational invariance of SIFT feature vector in the present invention, and method is simple and convenient.
Further, under new reference axis, centered on key point, adjacent domain appropriate is selected, and be divided into
Several sub-regions, after the SIFT feature vector (step S1034) for extracting key point further include:
Step S1035: SIFT feature vector is normalized;The gradient of every bit is subtracted each other by adjacent pixel in image
It obtains, so above-mentioned condition can be improved by being normalized if the case where whole drift occurs in the gray value of image;
Step S1036: whether the SIFT feature vector after judging normalization is greater than threshold value, if so, feature vector is set
It is set to threshold value, and is normalized again;Because of in some cases, the change of nonlinear illumination condition and camera saturation degree
Change will lead to some directions gradient value it is larger, and direction almost do not have it is impacted.At this point, after vector is normalized, Ke Yishe
A threshold value is set, biggish gradient value is truncated using this threshold value, then the vector after truncation is once normalized again
Processing, can so be further reduced interference, improve the distinctive of SIFT feature.
Therefore the invention avoids the influences that the gray value of iris image integrally drifts about, and further reduce interference, improve
The distinctive of SIFT feature.
There are many ways to carrying out pupil coarse positioning to iris image, the present invention can use radial symmetry transform to iris
Image carries out pupil coarse positioning, comprising:
Step S1011: sobel operator convolution iris image is used, gradient magnitude and the gradient side of each pixel are obtained
To;
Step S1012: being voted on gradient direction with different radii r length, and accumulative corresponding position gradient magnitude,
Gradient direction mapping graph and gradient magnitude mapping graph are obtained, in conjunction with the gradient direction mapping graph and gradient magnitude mapping graph, is taken
Being worth maximum position is the center of circle, and corresponding radius r is pupil radium;It can position to obtain the position of pupil with the center of circle and pupil radium
It sets.
Step S1011 to step S1012 is the method for radial symmetry transform, can quickly position pupil position.
Corresponding with the above-mentioned method positioned to iris image, this law is bright also to be provided a kind of pair of iris image and determines
The device of position, as shown in Figure 2, comprising:
Locating module 11, for carrying out pupil coarse positioning to iris image;
Initial Value module 12, for initializing the coordinate of multiple key points on the basis of coarse positioning;
Extraction module 13, for extracting the SIFT feature vector of key point;
Mapping block 14, for using SDM algorithm by the SIFT feature DUAL PROBLEMS OF VECTOR MAPPING of key point at increment of coordinate;
Update module 15 is updated the coordinate value of key point for using increment of coordinate;
Fitting module 16, for being carried out curve fitting using robust regression method to key point, obtain upper eyelid parabola,
The parameter of palpebra inferior parabola and iris outer circle.
The device positioned to iris image of the invention is simple and convenient, and speed is fast;Strong antijamming capability, precision are high;
Generalization is high;Robustness is good.
In the present invention, can also the coordinate value to key point repeatedly updated (successive ignition), at this point, update module it
Afterwards, before fitting module further include:
Judgment module meets termination condition for judging whether, if it is not, going to extraction module;
The present invention calculates the coordinate value of key point using iterative method, and extraction module to update module claims an iteration, passes through
Repeatedly (setting value) iteration can be avoided an iteration and fall into local better solution, obtain the coordinate value of stable key point.This hair
The bright accuracy for further improving positioning.
As a kind of improvement of the device of the invention positioned to iris image, extraction module includes:
Structural unit, for constructing the scale space of the iris image after pupil coarse positioning;
Statistic unit, for determining the main gradient side of key point in the case where existing scale space and key point coordinate
To;
Rotary unit keeps the direction of reference axis identical as main gradient direction for rotating to reference axis;
Extraction unit centered on key point, selects adjacent domain appropriate under new reference axis, and by its
It is divided into several sub-regions, extracts the SIFT feature vector of key point.
The device ensure that the rotational invariance of SIFT feature vector in the present invention, and method is simple and convenient.
Further, after extraction unit further include:
Normalization unit, for SIFT feature vector to be normalized;
Unit is truncated, for judging whether the SIFT feature vector after normalization is greater than threshold value, if so, by feature vector
It is set as threshold value, and is normalized again.
Therefore the invention avoids the influences that the gray value of iris image integrally drifts about, and further reduce interference, improve
The distinctive of SIFT feature.
There are many ways to carrying out pupil coarse positioning to iris image, the present invention can use radial symmetry transform to iris
Image is into row pupil coarse positioning, at this point, locating module includes:
Convolution unit obtains gradient magnitude and the side of each pixel for using sobel operator convolution iris image
To;
Ballot unit, is voted on gradient direction with different radii r length, and accumulative corresponding position gradient magnitude,
Gradient direction mapping graph and gradient magnitude mapping graph are obtained, in conjunction with the gradient direction mapping graph and gradient magnitude mapping graph, is taken
Being worth maximum position is the center of circle, and corresponding radius r is pupil radium;It can position to obtain the position of pupil with the center of circle and pupil radium
It sets.
The method that the present invention utilizes radial symmetry transform, quickly positions pupil position.
Below with specific one embodiment for example, the process of the present embodiment is as shown in Figure 9:
(1) characteristic extraction procedure:
(a) shape vector assigns initial value
Shape vector can be made of the transverse and longitudinal coordinate of 32 points, be 64 dimensional vectors, be each 13 of upper palpebra inferior respectively
Point, totally 26 points;Each 3 points of iris outer circle or so, totally 6 points.In Fig. 3 shown in (a), (b) shows main point in Fig. 3
Label, (c) and (d) lists iris outer circle and only one intersection point of eyelid and when without intersection point respectively in Fig. 3, the setting of key point
(it should be pointed out that in (c) palpebra inferior each key point be overlapped, (d) in upper eyelid each key point be overlapped, this is
Ideally, it is in fact extremely difficult to, and this refers to the final result after subsequent calculating, rather than initial value, initial value are real
Being still on border is 32 points, specific as shown in Figure 4).
Fast rough is carried out to pupil using radial symmetry transform (Radial Symmetric Transform, RST),
Using the pupil rough position of positioning, using displacement information, initial value is assigned for shape vector, coarse positioning guarantees shape vector position base
This is reliable, and it is too far not deviate by true value.Such as Fig. 4.
(b) the main gradient direction of key point is chosen
Firstly, the scale space of construction iris image.Verified Gaussian kernel is uniquely to can produce more rulers to Lindeberg
The core in space is spent, for a width two-dimensional image I (x, y), scale space L (x, y, σ) is defined as:
Wherein, G (x, y, σ) is the Gaussian function of changeable scale, and σ is scale size, and (x, y) is space coordinate.The size of σ
Determine the fog-level of image, σ value is bigger, and image is fuzzyyer, i.e., resolution ratio is lower, and σ value is smaller, can more retain the details of image
Information.σ=1.6 are taken at this time, it, can be first before establishing scale space in order to avoid Gaussian Blur loses excessive high-frequency information
Image is enlarged into original 2 times to retain the information of original image.
In the case where the scale space and key point coordinate of existing image, in crucial neighborhood of a point, surrounding picture is counted
The gradient direction distribution of element is that each key point specifies principal direction on specific Gaussian image L with this, descriptor is made to have rotation
Turn invariance, specific as follows:
For each point (x, y) on L, the amplitude m (x, y) and direction θ (x, y) of its gradient are calculated:
Then, centered on key point, the direction histogram of this key point is constituted using the gradient direction in certain area.
The Statistical Radius of histogram is generally 3*1.5* σ, and the range of direction histogram is 0~360 degree, and every 10 degree are a column, shares
36 columns, the corresponding direction of every column are the gradient direction of pixel, the length of the column gradient magnitude after the adduction of direction thus.It chooses
Local gradient direction histogram peak-peak direction is the principal direction of key point, peak value institute of 0.8 times greater than peak-peak
The direction of key point is also used as in direction.Fig. 5 has carried out Gauss weighting to the gradient magnitude of crucial vertex neighborhood, the master obtained in this way
Direction is relatively stable, and (in Fig. 5, it (b) is direction that (a), which is the gradient magnitude of each pixel and direction in crucial neighborhood of a point,
Histogram).
(c) the SIFT descriptor of key point is extracted
I. image-region needed for calculating descriptor is determined
Key point adjacent domain is divided into the subregion of d × d (d=4), each region is every as a seed point
The rectangular area that sub-regions distribute 3 σ is sampled, and is asked unlike principal direction, at this point, each subregion gradient direction is straight
Side schemes have 8 directions, and 0,π,In view of needing bilinear interpolation when practical calculate, thus it is required
The side length of image-region is 3 σ * (d+1), the radius of image-region are as follows:
Wherein σ is the corresponding scale of Gaussian image where key point.
II. area coordinate axis rotates
In order to guarantee that feature vector still has invariance in the case where image rotates, the seat with key point is needed
It is designated as center, adjacent domain is obtained after reference axis is rotated the angle θ, that is, rotates in principal direction, as shown in Figure 6.
Rotate the new coordinate of sampled point in rear region are as follows:
III. sampling area histogram of gradients is calculated
Postrotational sample point coordinate is assigned to d × d sub-regions in the circle that radius is radius, then obtains every
The gradient magnitude of a sampled point and direction are weighted each gradient magnitude with the standard gaussian function that parameter is 3 σ d/2, this
Sample can obtain the histogram of gradients with 8 directions, such as Fig. 7 at each seed point.
The normalization of IV.SIFT feature vector
The adjacent domain of key point is divided into the subregion of d × d (d=4), i.e. 16 seed points, and each seed point
These gradient informations are combined into the vector of 4*4*8=128 dimension, are exactly a key by the gradient information comprising 8 directions
SIFT feature around point.
After combination obtains SIFT feature, by the way that feature vector is normalized, illumination variation can be mitigated and brought
Influence.Because the gradient of every bit is subtracted each other by adjacent pixel in image, if the gray value of image occurs
The case where whole drift, above-mentioned condition can also be improved by being normalized.If obtained Descriptor vector is H=(h1,h2,...,
h128), the feature vector after normalization is L=(l1,l2,...,l128), then:
In addition, in some cases, the variation of nonlinear illumination condition and camera saturation degree will lead to some directions
Gradient value is larger, and direction almost do not have it is impacted.At this point, a threshold value can be set, generally take after vector is normalized
0.2, biggish gradient value is truncated using this threshold value, then carries out a normalized again to the vector after truncation, such as
The distinctive of SIFT feature can be improved in this.
(d) validity of SIFT feature
Traditional upper and lower eyelid localization method, is highly susceptible to the interference of eyelashes.In order to illustrate SIFT feature to eyelashes
Interference have certain robustness, different type image is counted here, such as Fig. 8, respectively count ciliation interference and
The true eyelid SIFT feature and non-genuine eyelid SIFT feature of no eyelashes interference, right side are real features, and left side is non-genuine
Feature, it is seen then that comparison in difference is obvious between the two, meanwhile, for true feature, ciliation interference and the spy without eyelashes interference
Although being had a certain difference between sign, whole similarity or relatively high.SIFT feature is to iris, eyelid and non-phase
The point of pass has good identification and robustness.
(2) SDM is returned
Feature extraction finishes, and needs that shape is calculated in Feature Mapping to coordinate using suitable regression iterative algorithm
Increment uses SDM algorithm here.Unlike gradient descent method, Newton method, quasi-Newton method, simulated annealing, EM scheduling algorithm progressive updating
Strategy, SDM adhere to that a step updates in place.SDM method is published in CVPR2013, is proposed by Xuehan Xiong et al., mesh
Be directly through the supervised learning optimal vector away from target point that obtains current point.
If the target of given piece image I, SDM positioning: it minimizes formula (6), obtains optimal shape increment Delta S,
Wherein, h (I (S)) indicates to extract the SIFT feature of shape vector S.The target of SDM parametric solution such as formula (7),
Wherein,Indicate the shape increment of i-th of sample,It indicates in current shape vectorIt extracts at place
SIFT feature, R0And b0It is the regression parameter to be learnt,Indicate the original shape of i-th of sample.In general, for algorithm
Robustness, when training, piece image will assign multiple original shapes.But since the characteristic variations of iris picture itself are little,
And when shooting, human eye has stringent control away from the distance of iris capture apparatus, so, here, sample original shape is assigned entirely to
Average shape (mean values of all training sample shape vectors), formula (7) becomes,
SDM in view of the loss function f in formula (6) about the concavity and convexity of Δ S be it is indefinite, single-step iteration is fallen in order to prevent
Entering local better solution, SDM is returned using multistage, such as formula (8),
Formula (7) is typical Linear least squares minimization problem, we are using gradient descent method to parameter Rk, bkIt exercises supervision
Study, has,
Wherein,
After parameter obtains, the shape of sample is updated to be updated using following form,
It is returned by multistage, generally 3-4 times, the shape vector estimated, that is, the key point coordinate predicted.
(3) robust regression
After obtaining key point coordinate, before segmentation obtains iris region, need to obtain up and down by curve-fitting method
Eyelid parabola and the outer Circle Parameters of iris.
There are many curve-fitting methods, and the most commonly used is least square regressions, error sum of squares minimum is pursued, just because of such as
This, haves the defects that fatal a, processing of the shortage to exceptional value.In order to solve this problem, individual point locations are reduced to miss
The curve matching of the larger generation of difference deviates risk, we use robust regression method, and the processing to exceptional value is added, and enhances curve
The robustness of fitting.
Robust regression method has very much, here, we using M estimate (M-estimation) method, be by Huber in
The Maximum-likelihood estimation of the broad sense proposed in 1964.By taking the circle for being fitted exterior iris boundary as an example, round equation is,
x2+y2- Dx-Ey-F=0 (11)
Wherein, x, y respectively indicate round cross, ordinate, and D, E, F are respectively the parameter of circle normal equation, we are write
At the form for being suitable for robust regression,
Specific to the fitting of each point, error model can be obtained,
Yi=Xib+εi (13)
Here,Xi=(1xi yi), xi、yiIndicate the key point positioned through SDM cross, ordinate.Its
In, εiThe error term obtained for i-th of point estimation parameter.We select Bisquare function as M estimator (M-
Estimator), i.e., the weights omega of each pointi, it is as follows to obtain cost function (cost function),
Here, ωiControl the i-th point of influence degree to fitting, and ωiIt is εiFunction, i.e. ωi=ω (εi), according to
Rely the error term ε at i-th pointi(ωi=1A(εi< b) (1- (εi/b)2)2, here, b is bandwidth), and εiDependent on i-th point
The determination of the parameter b, b of estimation again rely on the weight term ω in cost functioni, this is a typical iterative problem.Here,
We are solved using weighted iteration least square (Iterative reweighted least-square) method.
In this way, just having obtained the parameter of upper and lower eyelid parabola and iris outer circle.
Technical solution of embodiment of the present invention bring the utility model has the advantages that
The present invention regards the upper and lower eyelid parabola in iris recognition and the parameter determination process of iris outer circle as
The positioning and curve fit problem of upper and lower eyelid and iris edge key point.It is firstly introduced SIFT feature, using efficient
SDM algorithm, which quickly positions, obtains several key points of upper and lower eyelid and iris edge, utilizes robust regression parabola of fit and circle
Parameter, complete upper and lower eyelid boundary and the outer peripheral position fixing process of iris.This method, compared to calculus detective operators etc.
Conventional mapping methods have many advantages, such as that locating speed is fast, positioning accuracy is high, robustness is good, generalization is high, for example, traditional determines
Position method is directed to different image libraries, needs to adjust suitable parameter, and our method does not need adjusting parameter, regression parameter
It is obtained by training, it is all equally applicable for different image libraries.
The above is a preferred embodiment of the present invention, it is noted that for those skilled in the art
For, without departing from the principles of the present invention, it can also make several improvements and retouch, these improvements and modifications
It should be regarded as protection scope of the present invention.