CN101894256B - Iris identification method based on odd-symmetric 2D Log-Gabor filter - Google Patents

Abstract

The invention discloses an iris image feature extraction method based on an odd-symmetric 2D Log-Gabor filter, which comprises the following steps: (1) acquiring an iris image by using an iris acquisition device; (2) preprocessing the iris image, including iris positioning, iris image quality evaluation and normalization; (3) extracting the original iris features by using an improved algorithm based on the odd-symmetric 2D Log-Gabor filter: exporting the function of the 2D Log-Gabor filter, transforming the 2D Log-Gabor filter to rectangular coordinates from polar coordinates, resolving the 2D Log-Gabor filter into a two-dimensional odd-symmetric filter function, dividing iris subblocks and extracting the features; (4) fusing the features: extracting the amplitude value and phase of each subblock, and fusing the features; (5) and matching and classifying the features: encoding the features, encoding the Hamming distance, selecting the threshold of the Hamming distance and identifying the iris feature. The method breaks through the iris acquisition distance limit (2 meters) and enhances the accuracy and reliability for long-distance iris identification.

Description

Iris identification method based on odd symmetry 2D Log-Gabor wave filter

Technical field

The invention belongs to the biometrics identification technology field, relate to a kind of iris identification method based on odd symmetry 2D Log-Gabor wave filter.

Background technology

The quality of estimating a kind of biometrics identification technology has four aspects: antifalsification, ubiquity, stability and acceptable.In all biological characteristics, fingerprint is relatively stable but the admission fingerprint belongs to offensive.Shape of face characteristic has lot of advantages (like non-infringement property, user friendly), but the shape of face is with change of age, and camouflage easily.Sound characteristic has the advantage with shape of face feature similarity, but it is with factors vary such as age, health status and environment, and speech recognition system also easily by the recording cheat, forged easily.Iris feature identification solves these problems, and also has some advantages that above-mentioned other biological characteristic is had simultaneously.Iris recognition is the annular section texture that utilizes between pupil and the sclera, and just iris carries out the biosome recognition technology that identity is differentiated.The process of iris recognition forms iris feature after mainly comprising IMAQ, pre-service, feature extraction, coding, then with the iris templates database in feature templates compare coupling, draw recognition result.

Yet also there are a series of outstanding problems in existing iris identification method, as gathering distance less than 2 meters; The product environment for use is harsh; And there are problems such as crescent shape in the Gabor wave filter, causes the accuracy rate of iris recognition and matching efficiency lower, can't satisfy the actual demand in the application.

Summary of the invention

The invention provides a kind of iris identification method based on odd symmetry 2D Log-Gabor wave filter; Solve prior art and gathered the distance weak point; The product environment for use is harsh, and there is crescent shape in the Gabor wave filter, causes the accuracy rate and the lower problem of matching efficiency of iris recognition.

The technical scheme that the present invention adopted is, a kind of iris image feature extracting method based on odd symmetry 2D Log-Gabor wave filter, and this method is implemented according to following steps:

Step 1, collection iris image: the camera that adopts autozoom under the irradiation of halogen light modulation cooperates, collects iris image as iris capturing equipment;

Step 2, iris image pre-service:

2.1) Iris Location, through level and vertical Gray Projection to iris home position (x ₀, y ₀) carry out coarse positioning, use Gauss's gradient rounded edge to detect the location that operator carries out iris inward flange and outer boundary respectively then, suppose that (x is that coordinate is (x, the eye image intensity of y) locating, and the modelling center of circle at iris edge is at (x y) to I ₀, y ₀) locate, radius of circle is r, then tries to achieve central coordinate of circle and radius by formula (1):

Wherein * represents convolution algorithm,

Be with r ₀Be filter center, σ is the Gaussian function of standard deviation;

2.2) the iris image quality assessment;

2.3) iris normalization, being deployed on the polar coordinate system from the normalization of Descartes's rectangular coordinate system filtering the qualified iris image of assessment, conversion process is obtained by formula (2):

$\left\{\begin{array}{c}x(r,\mathrm{\θ})=(1-r)x_0\left(\mathrm{\θ}\right)+r{x}_1\left(\mathrm{\θ}\right)\\ y(r,\mathrm{\θ})=(1-r)y_0\left(\mathrm{\θ}\right)+r{y}_1\left(\mathrm{\θ}\right)\end{array}\right.---\left(2\right)$

In the formula (2), r ∈ [0,1], θ ∈ [0,2 π], (x ₀(θ), y ₀(θ)), (x ₁(θ), y ₁(θ)) represent pupil and the marginal point at the inside and outside edge of sclera on the θ direction respectively;

Step 3, use are extracted original iris features based on the improvement algorithm of odd symmetry 2D Log-Gabor wave filter:

3.1) deriving 2D Log-Gabor filter function, 2D Log-Gabor is defined under the log polar coordinates, and function is expressed and is provided by formula (3):

$G_{(s,t)}(\mathrm{\ρ},\mathrm{\θ})=\exp \left(\frac{-{(\mathrm{\ρ}-{\mathrm{\ρ}}_s)}^2}{2{\mathrm{\σ}}_{\mathrm{\ρ}}^2}\right)\exp \left(\frac{-{(\mathrm{\θ}-{\mathrm{\θ}}_{(s,t)})}^2}{2{\mathrm{\σ}}_{\mathrm{\θ}}^2}\right)---\left(3\right)$

$\left\{\begin{array}{c}{\mathrm{\ρ}}_s={\log }_{2}n-s\\ {\mathrm{\θ}}_{(s,t)}=\left\{\begin{array}{cc}\frac{\mathrm{\π}}{n_t}t& \mathrm{if\; s\; is\; odd}\\ \frac{\mathrm{\π}}{n_t}(t+\frac{1}{2})& \mathrm{if\; s\; is\; even}\end{array}\right.\\ ({\mathrm{\σ}}_{\mathrm{\ρ}},{\mathrm{\σ}}_{\mathrm{\θ}})=0.996(\sqrt{\frac{2}{3}},\frac{\mathrm{\π}}{\sqrt{2}{n}_t})\end{array}\right.$

In the formula (3), (ρ is polar coordinates θ), is scale with Log2, wave filter in the multiple frequence scope, n _sThe=5th, the number of division of multiresolution, n _tBe direction numerical value, its scope between 3-20, s ∈ 1 ...., n _s, t ∈ 1 ...., n _t, be used for the frequency spectrum and the direction of mark wave filter, (ρ _s, θ _{(s, t)}) be the coordinate center of wave filter, (σ _ρ, σ _θ) be the bandwidth under ρ and θ respectively;

Convert 2D Log-Gabor wave filter to rectangular coordinate from polar coordinates again, the frequency domain form after conversion is:

$G(u,v)=\exp \left(\frac{-{\left(\ln (u_1/f_0)\right)}^2}{2{\left(\ln ({\mathrm{\σ}}_u/f_0)\right)}^2}\right)\×\exp \left(\frac{-v_1^2}{2{\mathrm{\σ}}_v^2}\right)---\left(4\right)$

Wherein,

f ₀Be centre frequency, σ _uControl μ ₁Bandwidth on the direction, σ _vControl v ₁Bandwidth on the direction, θ is the direction of wave filter, for different centre frequency f ₀, σ _u/ f ₀And σ _vRemain unchanged;

3.2) 2D Log-Gabor wave filter is decomposed into Two-Dimensional Odd balanced-filter function, try to achieve by formula (5):

$G^o(u,v)=\exp \left(\frac{-{\left(\ln (u_1/f_0)\right)}^2}{2{\left(\ln ({\mathrm{\σ}}_u/f_0)\right)}^2}\right)\×\exp \left(\frac{-v_1^2}{2{\mathrm{\σ}}_v^2}\right)\sin \left(2\mathrm{\π}{f}_0{u}_1\right);---\left(5\right)$

3.3) sub-piece division of iris and feature extraction, the iris texture image after normalization and the enhancing is divided into 32 * 8 sub-block by the wave filter size, and obtains the eigenwert F of each sub-piece _Mn, try to achieve by formula (6):

$F_{\mathrm{mn}}(i,j)=\underset{x=1}{\overset{N}{\mathrm{\Σ}}}\underset{y=1}{\overset{M}{\mathrm{\Σ}}}I(i+x-\frac{N}{2},j+y-\frac{M}{2}){G_{\mathrm{mn}}}^o(x,y)---\left(6\right)$

G wherein _Mn ^o(u, v) for yardstick is that m, direction are the odd symmetry 2D Log-Gabor wave filter of n, (i j) is sub-piece centre coordinate, and N, M are sub-block size;

Step 4, Feature Fusion:

4.1) extract the amplitude and the phase place of sub-piece, extract (x, y) the eigenwert F of sub-piece through step 3 _Mn(x, y) after, through type (7) and formula (8) are extracted the amplitude information M of corresponding sub-piece _Mn(x is y) with phase information P _Mn(x, y):

$M_{\mathrm{mn}}(x,y)=\sqrt{F_{\mathrm{mn}}(x,y)\×\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}}---\left(7\right)$

$P_{\mathrm{mn}}(x,y)=\arctan \left(\frac{i\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}-F_{\mathrm{mn}}(x,y)}{\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}+F_{\mathrm{mn}}(x,y)}\right)---\left(8\right)$

Wherein,

Expression F _Mn(x, conjugate function y);

4.2) fusion feature, find n/2 the maximum direction sequence number s of amplitude under the m yardstick by formula (9), try to achieve yardstick by formula (10) and encode h with the effective iris feature under these directions _r, h _iRepresent the 1bit coding that real part and imaginary part generate respectively:

$s=\arg \underset{n}{\max }\left(M_{\mathrm{mn}}(x,y)\right)---\left(9\right)$

$\left\{\begin{array}{cc}(h_r,h_i)=\left(\mathrm{1,1}\right),& 0\≤P_{\mathrm{ms}}(x,y)\≤\mathrm{\π}/2\\ (h_r,h_i)=\left(\mathrm{0,1}\right),& \mathrm{\π}/2\≤P_{\mathrm{ms}}(x,y)\≤\mathrm{\π}\\ (h_r,h_i)=\left(\mathrm{0,0}\right),& \mathrm{\π}\≤P_{\mathrm{ms}}(x,y)\≤3\mathrm{\π}/2\\ (h_r,h_i)=\left(\mathrm{1,0}\right),& 3\mathrm{\π}/2\≤P_{\mathrm{ms}}(x,y)\≤2\mathrm{\π}\end{array}\right.;---\left(10\right)$

Step 5, characteristic matching classification: adopt the weighting hamming distance that distributes based on texture to be used for the characteristic matching classification, concrete steps are following:

5.1) feature coding: at first, characteristic area is divided into 32 * 8 sub-block, every sub-block is chosen 5 yardsticks and 8 directions, obtains the 20480bit data; Secondly, through step 4 Feature Fusion, obtain the 10240bit characteristic sequence; Afterwards, this sequence according to the hexadecimal code mode, is once encoded, obtained the feature coding of 1280 bytes for per 8;

5.2) the hamming range coding, apart from as the similarity of weighing between two iris-encodings, the hamming distance between two iris-encodings is obtained by formula (11) with hamming:

$\mathrm{HD}=\frac{1}{L}\underset{j=1}{\overset{L}{\mathrm{\Σ}}}{A}_j\⊕B_j---\left(11\right)$

L representes the length of whole iris-encoding, and A, B represent the iris feature coded sequence, and j representes the j position of iris-encoding, and A, B do XOR and calculate, and work as A _jWith B _jBe 0 when identical, otherwise be 1;

Utilization is calculated based on the weighting hamming distance that texture distributes: at first, obtain the hamming distance h of each subband according to formula (11) ₁, h ₂..., h _n, n is the subband number, i is big more, the subband h of expression _iThe closer to the iris inner boundary, through type (12) is found the solution the weighted sum of subband sea prescribed distance again:

$\mathrm{WHD}=\frac{1}{n}(a_1{h}_1+a_2{h}_2+...+a_n{h}_n)---\left(12\right)$

Weighting factor satisfies in the formula: a ₁＜a ₂＜...＜a _nAnd a ₁+ a ₂+ ...+a _n=n;

5.3) select the hamming distance threshold, obtain the hamming range coding of iris through above-mentioned steps after, according to the experiment statistics result, confirm that 0.38 is the hamming distance threshold;

5.4) iris feature identification, the hamming distance threshold that the step is confirmed in the employing compares the feature coding of gathering sample in iris and the iris database, if its hamming distance is then thought same iris smaller or equal to threshold value; Otherwise, then think different irises.

The invention has the beneficial effects as follows; Avoid of the influence of inferior quality iris image through the iris method for evaluating quality to recognition result; Use improved feature extracting method to remedy the crescent shape defective that Log-Gabor wave filter and 2D Log-Gabor wave filter cause owing to the conversion of angle direction bandwidth cycle property based on odd symmetry 2D Log-Gabor wave filter; Break through 2 meters of iris capturing distances with interior restriction, improved the accuracy and the reliability of long distance iris identification; And, farthest obtain the iris texture detailed information through the wavelet filtering technology, and bottom line is reduced in noise effect from the adverse effect that algorithm weakening image capture device brings.

Description of drawings

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 is the iris image after the inventive method is launched;

Fig. 3 is the lattice-shaped iris feature image that crescent shape waveform that the inventive method is directed against causes;

Fig. 4 is the iris feature image that the inventive method is extracted;

Fig. 5 is that the Hamming distance of the inventive method and even symmetry 2D Log-Gabor wave filter separates the cloth comparison diagram, and o1 and e1 curve are the range distribution that compares between two kinds of method classes, and o2 and e2 curve are the range distribution that compare in two kinds of method classes;

Fig. 6 is the ROC curve map of the inventive method to existing C ASIA iris storehouse;

Fig. 7 is the ROC curve map of the inventive method to existing UBIRIS iris storehouse.

Embodiment

Below in conjunction with accompanying drawing and embodiment the present invention is elaborated.

With reference to Fig. 1, the iris image feature extracting method based on odd symmetry 2D Log-Gabor wave filter of the present invention, implement according to following steps:

Step 1, long distance are gathered iris image.

Iris capturing is one of guardian technique of iris authentication system.Iris itself is small-sized, how under the situation of not invading human body, obtains the usefulness that high-quality iris image supplies identification, and this is to realize one of key problem of growing distance iris identification.But the present invention has adopted the high-resolution camera that has autozoom as iris capturing equipment, and camera lens is driven by micro-step motor, can focus automatically according to the feedback information of iris quality evaluation model, need not user's active assistance.Lighting source selects photography with halogen light modulation (power of selecting for use is 50W, and colour temperature is 5100K), can on iris, not stay apparent in view hot spot, forms noise pollution.

Step 2, iris image pre-service.

A series of iris images that step 1 is obtained carry out pre-service, from the eyes image of gathering, to obtain meeting the normalization iris image of identification requirement.The iris image pre-service comprises steps such as Iris Location, iris image quality assessment, normalization.

2.1) Iris Location.Through level and vertical Gray Projection to iris home position (x ₀, y ₀) carry out coarse positioning, use Gauss's gradient rounded edge to detect the location that operator carries out iris inward flange (pupil edge) and outer boundary (sclera border) respectively then.Suppose that (x is that coordinate is (x, the eye image intensity of y) locating, and the modelling center of circle at iris edge is at (x y) to I ₀, y ₀) locate, radius of circle is r, then through type (1) is tried to achieve central coordinate of circle and radius.

Wherein * represents convolution algorithm,

Be with r ₀Be filter center, σ is the Gaussian function of standard deviation.

2.2) the iris image quality assessment.

The purpose of iris image quality assessment is to filter underproof iris image, and exports qualified image in real time to follow-up identifying.When gathering, follow motions such as the blinking of eyes, rotation, possibly cause iris portion or all blocked or, thereby before identification, need estimate picture quality because focal length is forbidden to cause imaging unintelligible by eyelid, eyelash.

The iris image quality assessment is carried out in two steps:

2.21) feedback with evaluation.For let harvester can focusing, each parameter such as illumination carries out from main regulation, at first the brightness and the fog-level of images acquired are assessed.Make Q _Light, Q _BlurRepresent brightness, fuzzy two assessment results, then the assessment result Q of feedback respectively _FeedbackFor:

Q _feedback＝Q _Light+Q _Blur (2)

2.22) filter and assess.Through step 2.21) harvester is adjusted to optimum condition after, the assessment result Q of use formula (3) _IrisIris image to gathering filters, and with the image filtering that does not meet quality requirements, remaining iris image gets into next step processing.

Q _iris＝E _Bright×Q _Bright+E _Eyelid×Q _Eyelid+E _Eyelash×Q _Eyelash+E _Loc×Q _Loc (3)

Wherein, Q _Bright, Q _Eyelid, Q _EyelashAnd Q _LocBe respectively that hot spot, eyelid block, eyelashes block and the assessment result of Iris Location, E _Bright, E _Eyelid, E _EyelashAnd E _LocBe respectively that hot spot, eyelid block, eyelashes block and the weights factor of Iris Location assessment result, each weights factor span interval is [0,1], and E _Bright+ E _Eyelid+ E _Eyelash+ E _Loc=1.

2.3) iris normalization.Be deployed on the polar coordinate system from the normalization of Descartes's rectangular coordinate system filtering the qualified iris image of assessment, conversion process through type (4) obtains:

$\left\{\begin{array}{c}x(r,\mathrm{\θ})=(1-r)x_0\left(\mathrm{\θ}\right)+r{x}_1\left(\mathrm{\θ}\right)\\ y(r,\mathrm{\θ})=(1-r)y_0\left(\mathrm{\θ}\right)+r{y}_1\left(\mathrm{\θ}\right)\end{array}\right.---\left(4\right)$

In the formula: r ∈ [0,1], θ ∈ [0,2 π], (x ₀(θ), y ₀(θ)), (x ₁(θ), y ₁(θ)) represent pupil and (being the inside and outside edge of iris) marginal point of sclera on the θ direction respectively.Iris image after the expansion, as shown in Figure 2.Common same iris is after repeatedly gathering, and the position of iris and size all can change in each iris image.Normalization launches iris image is adjusted to identical size and correspondence position, thereby has eliminated translation, convergent-divergent and the selection influence to iris recognition, for necessary base has been established in next step feature extraction.

Step 3, use are extracted original iris features based on the improvement algorithm of odd symmetry 2D Log-Gabor wave filter.

3.1) deriving 2D Log-Gabor filter function, 2D Log-Gabor is defined under the log polar coordinates, and function is expressed through type (5) and is provided:

$G_{(s,t)}(\mathrm{\ρ},\mathrm{\θ})=\exp \left(\frac{-{(\mathrm{\ρ}-{\mathrm{\ρ}}_s)}^2}{2{\mathrm{\σ}}_{\mathrm{\ρ}}^2}\right)\exp \left(\frac{-{(\mathrm{\θ}-{\mathrm{\θ}}_{(s,t)})}^2}{2{\mathrm{\σ}}_{\mathrm{\θ}}^2}\right)---\left(5\right)$

$\left\{\begin{array}{c}{\mathrm{\ρ}}_s={\log }_{2}n-s\\ {\mathrm{\θ}}_{(s,t)}=\left\{\begin{array}{cc}\frac{\mathrm{\π}}{n_t}t& \mathrm{if\; s\; is\; odd}\\ \frac{\mathrm{\π}}{n_t}(t+\frac{1}{2})& \mathrm{if\; s\; is\; even}\end{array}\right.\\ ({\mathrm{\σ}}_{\mathrm{\ρ}},{\mathrm{\σ}}_{\mathrm{\θ}})=0.996(\sqrt{\frac{2}{3}},\frac{\mathrm{\π}}{\sqrt{2}{n}_t})\end{array}\right.$

In the formula (5): (ρ is polar coordinates θ), is scale with Log2, wave filter in the multiple frequence scope, n _sThe=5th, the number of division of multiresolution, n _tBe direction numerical value, its scope is between 3-20, and preferred value is 8, belongs to 8 anisotropic filters, s ∈ 1 ...., n _s, t ∈ 1 ...., n _t, be used for the frequency spectrum and the direction of mark wave filter, (ρ _s, θ _{(s, t)}) be the coordinate center of wave filter, (σ _ρ, σ _θ) be the bandwidth under ρ and θ respectively.

Convert 2D Log-Gabor wave filter to rectangular coordinate from polar coordinates again, the frequency domain form after conversion is:

$G(u,v)=\exp \left(\frac{-{\left(\ln (u_1/f_0)\right)}^2}{2{\left(\ln ({\mathrm{\σ}}_u/f_0)\right)}^2}\right)\×\exp \left(\frac{-v_1^2}{2{\mathrm{\σ}}_v^2}\right)---\left(6\right)$

Wherein

f ₀Be centre frequency, σ _μControl μ ₁Bandwidth on the direction, σ _vControl v ₁Bandwidth on the direction, θ is the direction of wave filter, in order to guarantee the constant shape of wave filter, for different centre frequency f ₀, σ _u/ f ₀And σ _vMust remain unchanged.

3.2) 2D Log-Gabor wave filter is decomposed into Two-Dimensional Odd balanced-filter function, decomposition computation is tried to achieve by formula (7):

$G^o(u,v)=\exp \left(\frac{-{\left(\ln (u_1/f_0)\right)}^2}{2{\left(\ln ({\mathrm{\σ}}_u/f_0)\right)}^2}\right)\×\exp \left(\frac{-v_1^2}{2{\mathrm{\σ}}_v^2}\right)\sin \left(2\mathrm{\π}{f}_0{u}_1\right)---\left(7\right)$

Two-Dimensional Odd symmetry Log-Gabor wave filter can change the bandwidth that results from the angle direction on the 2 d texture direction bandwidth, amplifies because the influence of the periodicity of angle makes bandwidth be periodically, causes the generation of crescent shape frequency-domain waveform; But on the 2 d texture direction, excessive can't causing of bandwidth periodically amplifies, when two component bandwidth are excessive; The spatial domain waveform of function can produce distortion, equally also exceeds the coverage of organizing wave filter, so; Pass through wavelet transformation; Enlarged frequency range and rationally increased the covering frequency spectrum, lattice-shaped problem (like Fig. 3, shown in 4) appears in the iris feature image of having avoided extracting.

3.3) sub-piece division of iris and feature extraction.Iris texture image after normalization and the enhancing is divided into 32 * 8 sub-block by the wave filter size, and obtains the eigenwert F of each sub-piece _Mn, through type (8) is tried to achieve:

$F_{\mathrm{mn}}(i,j)=\underset{x=1}{\overset{N}{\mathrm{\Σ}}}\underset{y=1}{\overset{M}{\mathrm{\Σ}}}I(i+x-\frac{N}{2},j+y-\frac{M}{2}){G_{\mathrm{mn}}}^o(x,y)---\left(8\right)$

G wherein _Mn ^o(u, v) for yardstick is that m, direction are the odd symmetry 2D Log-Gabor wave filter of n, (i j) is sub-piece centre coordinate, and N, M are sub-block size.

Step 4, Feature Fusion.

Any outstanding feature extraction algorithm all absolutely fully is mingled with in texture feature extraction, the especially iris feature usually has a lot of invalid characteristics.Iris texture belongs to utmost point complex texture, and enhancing or denoising process are lost the real information of a lot of images easily, thereby produces error characteristic.In addition, the redundancy feature value of filling too much also is to cause discerning one of inaccurate reason.Based on above consideration, the primitive character that step 3 is extracted screens fusion, forms available characteristic.

Odd symmetry 2D Log-Gabor wave filter in the inventive method is chosen 5 yardsticks, and each yardstick is chosen 8 directions, and every sub-block provides 40 eigenwerts, chooses the validity feature of 20 conducts that wherein amplitude is bigger.

The Feature Fusion concrete steps are following:

4.1) extract the amplitude and the phase place of sub-piece.Extract (x, y) the eigenwert F of sub-piece through step 3 _Mn(x, y) after, through type (9) and formula (10) are extracted the amplitude information M of corresponding sub-piece _Mn(x is y) with phase information P _Mn(x, y):

$M_{\mathrm{mn}}(x,y)=\sqrt{F_{\mathrm{mn}}(x,y)\×\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}}---\left(9\right)$

$P_{\mathrm{mn}}(x,y)=\arctan \left(\frac{i\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}-F_{\mathrm{mn}}(x,y)}{\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}+F_{\mathrm{mn}}(x,y)}\right)---\left(10\right)$

Wherein

Expression F _Mn(x, conjugate function y).

4.2) fusion feature.Through type (11) finds n/2 the maximum direction sequence number s of amplitude under the m yardstick, tries to achieve yardstick by formula (12) and encodes h with the effective iris feature under these directions _r, h _iRepresent the 1bit coding that real part and imaginary part generate respectively.

$s=\arg \underset{n}{\max }\left(M_{\mathrm{mn}}(x,y)\right)---\left(11\right)$

$\left\{\begin{array}{cc}(h_r,h_i)=\left(\mathrm{1,1}\right),& 0\≤P_{\mathrm{ms}}(x,y)\≤\mathrm{\π}/2\\ (h_r,h_i)=\left(\mathrm{0,1}\right),& \mathrm{\π}/2\≤P_{\mathrm{ms}}(x,y)\≤\mathrm{\π}\\ (h_r,h_i)=\left(\mathrm{0,0}\right),& \mathrm{\π}\≤P_{\mathrm{ms}}(x,y)\≤3\mathrm{\π}/2\\ (h_r,h_i)=\left(\mathrm{1,0}\right),& 3\mathrm{\π}/2\≤P_{\mathrm{ms}}(x,y)\≤2\mathrm{\π}\end{array}\right.---\left(12\right)$

Through the Feature Fusion formula; Be not that all unique points are all participated in comparison identification, and through Feature Fusion, the quantization characteristic point of participating in identification all be at the regional level in the most representative characteristic; Thereby simplified characteristic quantity to greatest extent, increased recognition efficiency.

Step 5, characteristic matching classification.

The process of characteristic matching classification is exactly an identifying; Being about to the image of collection and the iris image in the database compares; Calculate " similarity " of two iris image codings; Whether relatively differentiate these two codings to " similarity " value and a fixing lower limit from same iris then, reach final identifying purpose.

The present invention introduces the weighting hamming distance that distributes based on texture and is used for the characteristic matching classification, and the concrete steps of characteristic matching classification are following:

5.1) feature coding.At first, characteristic area is divided into 32 * 8 sub-block, every sub-block is chosen 5 yardsticks and 8 directions, obtains the 20480bit data; Secondly, through step 4 Feature Fusion, obtain the 10240bit characteristic sequence; Afterwards, this sequence according to the hexadecimal code mode, is once encoded, just produced the feature coding of 1280 bytes for per 8.

5.2) the hamming range coding.Hamming distance (Hamming Distance HD) can be used as the similarity of weighing between two iris-encodings, and the hamming distance between two iris-encodings is obtained by formula (13):

$\mathrm{HD}=\frac{1}{L}\underset{j=1}{\overset{L}{\mathrm{\Σ}}}{A}_j\⊕B_j---\left(13\right)$

L representes the length of whole iris-encoding, and A, B represent the iris feature coded sequence, and j representes the j position of iris-encoding, and A, B do XOR and calculate, and work as A _jWith B _jBe 0 when identical, otherwise be 1.

The inventive method utilization is calculated based on the weighting hamming distance that texture distributes: at first, obtain the hamming distance h of each subband according to formula (13) ₁, h ₂..., h _n, n is the subband number, i is big more, the subband h of expression _iThe closer to the iris inner boundary, find the solution the weighted sum of subband sea prescribed distance by formula (14) again:

$\mathrm{WHD}=\frac{1}{n}(a_1{h}_1+a_2{h}_2+...+a_n{h}_n)---\left(14\right)$

Weighting factor satisfies in the formula: a ₁＜a ₂＜...＜a _nAnd a ₁+ a ₂+ ...+a _n=n.

For similar iris; Because of the hamming of each subband distance basic identical; The Weighted distance that through type (14) obtains is the same basically with the hamming distance that formula (13) is calculated; And for inhomogeneous iris, little because of the more inner border of the hamming of outer boundary subband distance subband, the Weighted distance that the result obtains is big than the definition distance of hamming distance.Therefore, after the weighting hamming of employing formula (14) distance, the class spacing of iris can be bigger, thereby help the differentiation between the different irises.

5.3) selection hamming distance threshold.Obtain the hamming range coding of iris through above-mentioned steps after, need confirm the hamming distance threshold through the coupling rule.The selection of threshold value is very important, and it has direct influence to the reject rate and the misclassification rate of recognizer.Can find that through statistics be distributed between 0.4 to 0.5 in the hamming distance set of different iris images, the hamming range distribution of identical iris concentrates between 0.15 to 0.35 by Fig. 5.Choose the threshold value between 0.35 to 0.4, recognition result is added up, can get misclassification rate (FAR) and equate at 0.38 place with reject rate (FRR).Therefore, can confirm that 0.38 is the hamming distance threshold.

5.4) iris feature identification.Confirm after the threshold value, the feature coding of gathering sample in iris and the iris database is compared, if its hamming distance is then thought same iris smaller or equal to threshold value; Otherwise, then think different irises.

The invention provides a kind of iris identification method based on odd symmetry 2D Log-Gabor wave filter; To the demand of long distance iris identification in practical application; Adopt adaptive iris capturing system, can adjust automatically through the iris quality assessment result of iris method for evaluating quality feedback; Filter gathering iris image simultaneously, avoid of the influence of inferior quality iris image recognition result; Use improved feature extracting method, remedied the crescent shape defective that Log-Gabor wave filter and 2D Log-Gabor wave filter cause owing to the conversion of angle direction bandwidth cycle property based on odd symmetry 2D Log-Gabor wave filter; Use weighting hamming distance to carry out iris feature coupling and identification, guaranteed the high efficiency and the robustness of identification.The inventive method has broken through 2 meters of existing iris capturing distances with interior restriction, has improved the accuracy and the reliability of long distance iris identification.

Among the embodiment, select for use CASIA-IrisV3 iris storehouse to comprise that 2655 width of cloth iris image samples and the UBIRS iris storehouse of totally 396 eyes comprise 241 people, 1877 width of cloth iris sample images establishment test sample book of totally 246 eyes to 249 people.UBIRIS iris storehouse comprises coloured image; Different with the CASIA gray-scale map; After pre-processing module, the performance of the iris image texture of UBIRIS is poor slightly, and it is outstanding inadequately that main cause is that capillary color and iris background color show; Through behind the gaussian filtering, discrimination is not as good as the recognition effect in CASIA storehouse again.

2D Gabor algorithm, Log-Gabor complex filter algorithm and the inventive method contrast test to commonly used compare; With reference to Fig. 6, Fig. 7 and table 1, be its false acceptance rate of iris authentication system (FAR) of adopting three kinds of algorithms, false rejection rate (FRR), etc. the error rate performance index tables of comparisons such as (ERR).

Table 1, adopt three kinds of algorithms false rejection rate, etc. the error rate table of comparisons

In the contrast experiment, also find, use the odd symmetry wave filter and improve about 4%-5% than using even symmetry wave filter recognition performance.Though the coding effect is similar, matching performance still is that the odd symmetry wave filter is slightly high.Use the odd symmetry wave filter higher by about 0.04 than coupling between use even symmetry filtering class; Exceed about 0.02 than a type interior coupling; The main cause that causes this phenomenon is that iris texture more is partial to point symmetry in local frequency domain distribution; Rather than line symmetry, after the iris image convolution, efficient coding is manyed 0.3-0.6 than the efficient coding that obtains with the line symmetry.

Claims (3)

1. iris image feature extracting method based on odd symmetry 2D Log-Gabor wave filter is characterized in that this method is implemented according to following steps:

Step 1, collection iris image: the camera that adopts autozoom under the irradiation of halogen light modulation cooperates, collects iris image as iris capturing equipment;

Step 2, iris image pre-service:

2.1) Iris Location, through level and vertical Gray Projection to iris home position (x ₀, y ₀) carry out coarse positioning, use Gauss's gradient rounded edge to detect the location that operator carries out iris inward flange and outer boundary respectively then, suppose that (x is that coordinate is (x, the eye image intensity of y) locating, and the modelling center of circle at iris edge is at (x y) to I ₀, y ₀) locate, radius of circle is r, then tries to achieve central coordinate of circle and radius by formula (1):

Wherein dS is the differential of the long S of circular arc, and * represents convolution algorithm,

Be with r ₀Be filter center, σ is the Gaussian function of standard deviation;

2.2) the iris image quality assessment, concrete steps are following:

2.21) feedback with evaluation, at first the brightness and the fog-level of images acquired are assessed, make Q _Light, Q _BlurRepresent brightness, fuzzy two assessment results, then the assessment result Q of feedback respectively _FeedbackFor:

Q _feedback＝Q _Light+Q _Blur (13)

2.22) filter assessment, through step 2.21) harvester is adjusted to optimum condition after, the assessment result Q of use formula (14) _IrisIris image to gathering filters, and with the image filtering that does not meet quality requirements, remaining iris image gets into next step processing,

Q _iris＝E _Bright×Q _Bright+E _Eyelid×Q _Eyelid+E _Eyelash×Q _Eyelash+E _Loc×Q _Loc (14)

Wherein, Q _Bright, Q _Eyelid, Q _EyelashAnd Q _LocBe respectively that hot spot, eyelid block, eyelashes block and the assessment result of Iris Location, E _Bright, E _Eyelid, E _EyelashAnd E _LocBe respectively that hot spot, eyelid block, eyelashes block and the weights factor of Iris Location assessment result, each weights factor span interval is [0,1], and E _Bright+ E _Eyelid+ E _Eyelash+ E _Loc=1;

2.3) iris normalization, being deployed on the polar coordinate system from the normalization of Descartes's rectangular coordinate system filtering the qualified iris image of assessment, conversion process is obtained by formula (2):

$\left\{\begin{array}{c}x(r,\mathrm{\θ})=(1-r)x_0\left(\mathrm{\θ}\right)+{\mathrm{rx}}_1\left(\mathrm{\θ}\right)\\ y(r,\mathrm{\θ})=(1-r)y_0\left(\mathrm{\θ}\right)+{\mathrm{ry}}_1\left(\mathrm{\θ}\right)\end{array}\right.---\left(2\right)$

In the formula (2), r ∈ [0,1], θ ∈ [0,2 π], (x ₀(θ), y ₀(θ)), (x ₁(θ), y ₁(θ)) represent pupil and the marginal point at the inside and outside edge of sclera on the θ direction respectively;

Step 3, use are extracted original iris features based on the improvement algorithm of odd symmetry 2D Log-Gabor wave filter:

3.1) deriving 2D Log-Gabor filter function, 2D Log-Gabor is defined under the log polar coordinates, and function is expressed and is provided by formula (3):

$G_{(s,t)}(\mathrm{\ρ},\mathrm{\θ})=\exp \left(\frac{{-(\mathrm{\ρ}-{\mathrm{\ρ}}_s)}^2}{{2\mathrm{\σ}}_{\mathrm{\ρ}}^2}\right)\exp \left(\frac{{-(\mathrm{\θ}-{\mathrm{\θ}}_{(s,t)})}^2}{{2\mathrm{\σ}}_{\mathrm{\θ}}^2}\right)---\left(3\right)$

$\left\{\begin{array}{c}\\ {\mathrm{\ρ}}_s={\log }_{2}n-s\\ {\mathrm{\θ}}_{(s,t)}=\left\{\begin{array}{cc}\frac{\mathrm{\π}}{n_t}t& \mathrm{ifsisodd}\\ \frac{\mathrm{\π}}{n_t}(n+\frac{1}{2})& \mathrm{ifsiseven}\end{array}\right.\\ ({\mathrm{\σ}}_{\mathrm{\ρ}},{\mathrm{\σ}}_{\mathrm{\θ}})=0.966(\sqrt{\frac{2}{3}},\frac{\mathrm{\π}}{\sqrt{2}{n}_t})\end{array}\right.$

In the formula (3), (ρ is polar coordinates θ), is scale with Log2, wave filter in the multiple frequence scope, n _sThe=5th, the number of division of multiresolution, n _tBe direction numerical value, its scope between 3-20, s ∈ 1 ...., n _s, s is used for the frequency spectrum of mark wave filter, t ∈ 1 ...., n _t, t is used for the direction of mark wave filter, (ρ _s, θ _{(s, t)}) be the coordinate center of wave filter, (σ _ρ, σ _θ) be the bandwidth under ρ and θ respectively;

Convert 2D Log-Gabor wave filter to rectangular coordinate from polar coordinates again, the frequency domain form after conversion is:

$G(u,v)=\exp \left(\frac{{-\left(\ln (u_1/f_0)\right)}^2}{2{\left(\ln ({\mathrm{\σ}}_u/f_0)\right)}^2}\right)\×\exp \left(\frac{{-v}_1^2}{{2\mathrm{\σ}}_v^2}\right)---\left(4\right)$

Wherein, $\left\{\begin{array}{c}{u}_1=u\mathrm{Cos}\mathrm{\θ}+v\mathrm{Sin}\mathrm{\θ}\\ v_1=-u\mathrm{Sin}\mathrm{\θ}+v\mathrm{Cos}\mathrm{\θ}\end{array}\right.,$ f ₀Be centre frequency, σ _uControl μ ₁Bandwidth on the direction, σ _vControl v ₁Bandwidth on the direction, θ is the direction of wave filter, for different centre frequency f ₀, σ _u/ f ₀And σ _vRemain unchanged;

3.2) 2D Log-Gabor wave filter is decomposed into Two-Dimensional Odd balanced-filter function, try to achieve by formula (5):

$G^o(u,v)=\exp \left(\frac{-{\left(\ln (u_1/f_0)\right)}^2}{2{\left(\ln ({\mathrm{\σ}}_u/f_0)\right)}^2}\right)\×\exp \left(\frac{{-v}_1^2}{{2\mathrm{\σ}}_v^2}\right)\sin \left(2\mathrm{\π}{f}_0{u}_1\right);---\left(5\right)$

3.3) sub-piece division of iris and feature extraction, the iris texture image after normalization and the enhancing is divided into 32 * 8 sub-block by the wave filter size, and obtains the eigenwert F of each sub-piece _Mn, try to achieve by formula (6):

$F_{\mathrm{mn}}(i,j)=\underset{x=1}{\overset{N}{\mathrm{\Σ}}}\underset{y=1}{\overset{M}{\mathrm{\Σ}}}I(i+x-\frac{N}{2},j+y-\frac{M}{2}){G_{\mathrm{mn}}}^o(x,y)---\left(6\right)$

G wherein _Mn ^o(u, v) for yardstick is that m, direction are the odd symmetry 2D Log-Gabor wave filter of n, (i j) is sub-piece centre coordinate, and N, M are sub-block size;

Step 4, Feature Fusion:

4.1) extract the amplitude and the phase place of sub-piece, extract (x, y) the eigenwert F of sub-piece through step 3 _Mn(x, y) after, through type (7) and formula (8) are extracted the amplitude information M of corresponding sub-piece _Mn(x is y) with phase information P _Mn(x, y):

$M_{\mathrm{mn}}(x,y)=\sqrt{F_{\mathrm{mn}}(x,y)\×\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}}---\left(7\right)$

$P_{\mathrm{mn}}(x,y)=\arctan \left(\frac{i\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}-F_{\mathrm{mn}}(x,y)}{\stackrel{\‾}{F_{\mathrm{mn}}(x,y)}+F_{\mathrm{mn}}(x,y)}\right)---\left(8\right)$

Wherein,

Expression F _Mn(x, conjugate function y);

4.2) fusion feature, find the concrete numerical value of n/2 maximum direction sequence number Ds of amplitude under the m yardstick by formula (9), try to achieve yardstick by formula (10) and encode h with the effective iris feature under these directions _r, h _iRepresent the 1bit coding that real part and imaginary part generate respectively:

$\mathrm{Ds}=\arg \underset{n}{\max }\left(M_{\mathrm{mn}}(x,y)\right)---\left(9\right)$

$\left\{\begin{array}{cc}(h_r,h_i)=\left(\mathrm{1,1}\right),& 0\≤P_{\mathrm{ms}}(x,y)\≤\mathrm{\π}/2\\ (h_r,h_i)=\left(\mathrm{0,1}\right),& \mathrm{\π}/2\≤P_{\mathrm{ms}}(x,y)\≤\mathrm{\π}\\ (h_r,h_i)=\left(\mathrm{0,0}\right),& \mathrm{\π}\≤P_{\mathrm{ms}}(x,y)\≤3\mathrm{\π}/2\\ (h_r,h_i)=\left(\mathrm{1,0}\right),& 3\mathrm{\π}/2\≤P_{\mathrm{ms}}(x,y)\≤2\mathrm{\π}\end{array}\right.;---\left(10\right)$

Step 5, characteristic matching classification: adopt the weighting hamming distance that distributes based on texture to be used for the characteristic matching classification, concrete steps are following:

5.1) feature coding: at first, characteristic area is divided into 32 * 8 sub-block, every sub-block is chosen 5 yardsticks and 8 directions, obtains the 20480bit data; Secondly, through step 4 Feature Fusion, obtain the 10240bit characteristic sequence; Afterwards, this sequence according to the hexadecimal code mode, is once encoded, obtained the feature coding of 1280 bytes for per 8;

5.2) the hamming range coding, apart from as the similarity of weighing between two iris-encodings, the hamming distance between two iris-encodings is obtained by formula (11) with hamming:

$\mathrm{HD}=\frac{1}{L}\underset{j=1}{\overset{L}{\mathrm{\Σ}}}{A}_j\⊕B_j---\left(11\right)$

L representes the length of whole iris-encoding, and A, B represent the iris feature coded sequence, and j representes the j position of iris-encoding, and A, B do XOR and calculate, and work as A _jWith B _jBe 0 when identical, otherwise be 1;

Utilization is calculated based on the weighting hamming distance that texture distributes: at first, obtain the hamming distance h of each subband according to formula (11) ₁, h ₂..., h _n, n is the subband number, i is big more, the subband h of expression _iThe closer to the iris inner boundary, through type (12) is found the solution the weighted sum of subband sea prescribed distance again:

$\mathrm{WHD}=\frac{1}{n}(a_1{h}_1+a_2{h}_2+...+a_n{h}_n)---\left(12\right)$

Weighting factor satisfies in the formula: a ₁＜a ₂＜...＜a _nAnd a ₁+ a ₂+ ...+a _n=n;

5.3) select the hamming distance threshold, obtain the hamming range coding of iris through above-mentioned steps after, according to the experiment statistics result, confirm that 0.38 is the hamming distance threshold;

5.4) iris feature identification, the hamming distance threshold that the step is confirmed in the employing compares the feature coding of gathering sample in iris and the iris database, if its hamming distance is then thought same iris smaller or equal to threshold value; Otherwise, then think different irises.

2. method according to claim 1 is characterized in that: said step 3.1), and n _tBe direction numerical value, preferred value is 8, belongs to 8 anisotropic filters.

3. method according to claim 1; It is characterized in that: in the said step 4, odd symmetry 2D Log-Gabor wave filter is chosen 5 yardsticks, and each yardstick is chosen 8 directions; Every sub-block provides 40 eigenwerts, chooses the validity feature of 20 conducts that wherein amplitude is bigger.

Images