CN102708364A - Cascade-classifier-based fingerprint image classification method - Google Patents

Abstract

The invention discloses a cascade-classifier-based fingerprint image classification method, and mainly aims to solve the problems of low classification accuracy and low classifier robustness of the prior art. The method is implemented by the following steps of: (1) extracting block directional fields of a fingerprint image by using a gradient method, calculating the mass of each block directional field, establishing a normalized directional field diffusion model, and calculating a normalized fingerprint image directional field; (2) performing complex filtering on the fingerprint image according to the normalized fingerprint image directional field; and (3) extracting the ridge flow direction of the fingerprint image, and designing a cascade classifier by taking the fingerprint image directional field, the response of the complex filtering and the ridge flow direction as characteristics to classify a fingerprint. The method has the advantages of high accuracy of the extracted directional fields, high discrimination performance of the extracted fingerprint characteristics, high classifier robustness and high classification accuracy, and can be used for classifying the fingerprint in an automatic fingerprint identification system.

Description

Fingerprint image sorting technique based on the classification device

Technical field

The invention belongs to the digital image processing techniques field, relate to the fingerprint image classification, specifically is a kind of fingerprint image sorting technique based on the classification device.

Background technology

The targeted customer is discerned the fingerprint image that needs the targeted customer to the robotization fingerprint recognition system and the fingerprint image in the fingerprint base compares one by one.Fingerprint base in a lot of law courts and the product for civilian use very conference causes the increase greatly in processing time and the reduction of degree of accuracy.The universal method that addresses this is that is exactly that the fingerprint image with similar features is classified.

Most sorting algorithm all is based on E.Henry at article Classification and Uses of Finger Prints, Rouledge, the Galton-Henry principle of classification that proposes in 1900.Common fingerprint is divided into 5 types, is respectively arch form A, tent camber T, left-handed type L, dextrorotatory form R and screw-type W.

The fingerprint the simplest a kind of method of classifying is based on the method for singular point, often has two types of singular points in the fingerprint, be respectively core point and trigpoint.M.Liu is at article Fingerprint classification based on adaboost learning from singu-larity features; Propose a kind of self-adaptation among Pattern Recognition 43 (3) (2010) 1062 – 1070 and strengthen learning algorithm; Singular point through detecting under multiple convergent-divergent is set up proper vector, then self-adaptation is strengthened learning algorithm and is used in design category device in the decision tree.This method is simple to operate, only needs can classify to fingerprint easily through quantity, type and the relevant position of singular point.But; To such an extent as to singular point is just not high to the confidence level of the very responsive singular point that extracts of noise itself; The situation of extracting and extracting by mistake can appear leaking; The another one problem is a singular point, and especially trigpoint maybe not can in fingerprint image occurs, and variety of problems causes this method cisco unity malfunction.

When the method based on singular point can't accurately be divided time-like, can flow to through the tracking crestal line to remedy, can follow the tracks of out crestal line based on the curvature of local direction field and flow to.J.-H.Chang and K.-C.Fan; At A new model for fingerprint classification by ridge distribution sequences; Defined 10 kinds of basic ridge pattern as characteristic among Pattern Recognition 35 (6) (2002) 1209 – 1223, set up sorter with distributing through the shape of analyzing crestal line.S.C.Dass and A.K.Jain are at article fingerprint classification using orientation field flow curves; In:InProceedings of ICVGIP; 2004, among the pp.650 – 655 according to one o'clock from a terminal point traverse into another terminal point methods analyst the plane tangent line the contour map picture.The shortcoming that crestal line flows to method is to distinguish arch and tent camber, and when trigpoint and core point vicinity, this method can not distinguish from the tent camber with dextrorotatory form left-handed again.Therefore, crestal line flow to method through be commonly used to remedy based on the not enough of the fingerprint classification method of singular point or and other characteristic bindings get up with training classifier.

Field of direction image is a most frequently used category feature that comes to flow to crestal line the training classifier of joining together, and most of existing fingerprint classification algorithm have all been used field of direction image, and in fact, singular point and crestal line flow graph can extract from field of direction image.A lot of direct service orientation field picture of method are extracted characteristic as characteristic through the direction field picture being carried out simple gridding, and this method causes the very proper vector of higher-dimension.For reducing EMS memory occupation and computing time, a lot of dimensionality reduction technologies are used to reduce intrinsic dimensionality.The Karhunen-Loeve conversion often is used to carry out the dimensionality reduction operation, and a kind of MKL of improving one's methods of Karhunen-Loeve conversion then not only can be used for carrying out the dimensionality reduction operation and can also classify.Use non-linear discriminating analysis also can carry out dimensionality reduction and classification to field of direction vector.But because the existence of small differences in the fingerprint image, feasible sorting technique based on field of direction image also is a great challenge.

Maturation day by day along with nerual network technique; A lot of machine learning methods are suggested the fingerprint characteristic classification that is used for fixing size; Neural network method also has been used widely in fingerprint classification and has been obtained reasonable effect, and these methods comprise multilayer level artificial neural networks, probabilistic neural network and self organizing neural network.Jain etc. are at A multichannel approach to fingerprintclassification; Propose among IEEE Transactions on Pattern Analysis and Machine Intel-ligence 21 (4) (1999) 348 – 359 to use the K nearest neighbor classifier to find out two types the most similar in the fingerprint code proper vector, train a special neural net method that it is distinguished then.But in the inferior quality fingerprint image, the degree of accuracy of said method all can be greatly affected.

Summary of the invention

The objective of the invention is to deficiency, propose a kind of fingerprint image sorting technique, to improve the degree of accuracy of fingerprint image classification based on the classification device to above-mentioned prior art.

For realizing above-mentioned purpose, fingerprint classification method of the present invention comprises the steps:

(1) use Fast Fourier Transform (FFT) that fingerprint image is strengthened, the fingerprint image I after strengthening be divided into the piece of w * w, and with gradient method take the fingerprint image I piece field of direction θ (x, y), x wherein, y is the horizontal ordinate and the ordinate of presentation video piece respectively;

(2) calculate the quality q of each piece fingerprint image according to the piece field of direction _{(x, y)}

The piece field of direction of the fingerprint image that (3) step (1) is extracted is converted into a continuous vector field:

v _(x，y)＝(v _1(x，y),v _2(x，y))，

Wherein, b _{1 (x, y)}=cos (2 θ (x, y)), v _{2 (x, y)}=sin (2 θ (x, y));

(4) utilize the continuous vector field of step (3) extraction and the picture quality of step (2) calculating to set up the associating energy function:

J(u _(x，y))＝D(u _(x，y))+με(u _(x，y))，

Wherein, u _{(x, y)}=(u _{1 (x, y)}, u _{2 (x, y)}) be the field of direction after the normalization to be asked, u _{1 (x, y)}Expression waits to ask the cosine value of the piece field of direction, u _{2 (x, y)}Expression waits to ask the sine value of the piece field of direction,

$D\left(u_{(x,y)}\right)=\frac{1}{2}{\∫}_{\mathrm{\Ω}}{q}_{(x,y)}{\left|\right|u_{(x,y)}-v_{(x,y)}\left|\right|}^2{d}_x{d}_y,$ Be source vector field v _{(x, y)}With the vector field u after the normalization _{(x, y)}Between the difference item,

Be penalty term, Ω representes the effective coverage of fingerprint image, and μ is the normalized parameter of decision difference item and penalty term relation, || ... || ²Represent 1 norm square;

(5) find the solution feasible associating energy function J (u _{(x, y)}) obtain the u of minimum value _{(x, y)}, draw the field of direction after the normalization

Atan wherein ^-1Represent arc tangent, draw the field of direction image B of fingerprint image;

(6) with the field of direction after the normalization fingerprint image I is carried out complex filter, and judges the type of fingerprint:

(6a) response and its maximum response threshold value E with every bit in the fingerprint image behind the complex filter compares, if certain any response, judges then that this is a core point greater than maximum response threshold value E, otherwise this point is not a core point, and wherein 0.4 < E < 0.8;

(6b) peak response and its minimum response threshold value F with complex filter compares, if the peak response of complex filter less than minimum response threshold value F, and does not have core point in the fingerprint image; Judge that then fingerprint is arch form A, otherwise, execution in step (6c); Wherein 0.3 < F < 0.7, and F < E;

(6c) through piece field of direction image B being carried out the take the fingerprint some field of direction of image I of bilinear interpolation computing,, ask every pair of vector of unit length on the neighbouring sample point line direction according to the field of direction sampled point on the crestal line that takes the fingerprint;

The vector of unit length of (6d) setting on first sampled point and second the sampled point line direction is an initial vector, is ordinate with the inner product of each vector of unit length and initial vector of unit length, and the sampled point sequence is a horizontal ordinate; Make the crestal line circle of equal altitudes; If two local maximums are arranged between 0.8 to 1 in the crestal line circle of equal altitudes, two local minimums judge that then fingerprint is screw-type W between-0.8 to-1; Otherwise, execution in step (6e);

(6e) set up proper vector according to the response of field of direction image B and complex filter every bit, with PCA proper vector is carried out dimensionality reduction, the proper vector after using K arest neighbors sorting algorithm to dimensionality reduction is carried out rough sort; Find out with the immediate K of an actual fingerprint image neighborhood, the type that wherein comprises two maximum neighborhood representatives of sample size is for immediate two types with true fingerprint pattern, if comprise left-handed type L or dextrorotatory form R in these two types; Do not comprise screw-type W, local maximum is between 0.8 to 1 in the crestal line circle of equal altitudes, and local minimum is between-0.8 to-1; And the terminal point of fingerprint ridge line up-sampling point is all in the left side of initial point; Judge that then fingerprint is left-handed type L, if the terminal point of sampled point judges then that all on the right side of initial point fingerprint is dextrorotatory form R; Otherwise, execution in step (6f);

(6f) use SVMs that two types of K arest neighbors sorting algorithm output are further classified; Import two types that K arest neighbors sorting algorithm rough sort is come out; The rough sort type of input is trained SVMs, and the classification results of the SVMs output after the training is final fingerprint pattern.

The present invention makes that the field of direction of extracting is more accurate owing to proposed a kind of normalization field of direction model; Simultaneously because the amalgamation and expression fingerprint characteristic that the present invention responds with the field of direction and complex filter; Promptly represent the directional information of crestal line with the field of direction; Express the characteristic of singular point with the complex filter response, the deficiency that makes the two can remedy the other side has mutually improved the differentiation property between fingerprint is all kinds of; Because the present invention proposes a classification device and comes fingerprint is classified, big difference and the little difference between all kinds of all possesses good robustness to this sorter to fingerprint in addition.Experimental result shows with method of the present invention fingerprint classification to be possessed better degree of accuracy.

Description of drawings

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

Fig. 2 flows to and the crestal line circle of equal altitudes for the crestal line among the present invention;

Fig. 3 is the present invention and the effect contrast figure of existing gradient method in the inferior quality fingerprint image.

Embodiment

With reference to Fig. 1, fingerprint classification method of the present invention comprises the steps:

Step 1, the field of direction of the image that takes the fingerprint.

(1.1) with Fast Fourier Transform (FFT) fingerprint image is strengthened, the fingerprint image I after strengthening is divided into the piece of w * w, with gradient method take the fingerprint image I piece field of direction θ (x, y), x wherein, y is the horizontal ordinate and the ordinate of presentation video piece respectively, w=32;

(1.2) calculate the gradient mean value M of each piece fingerprint image interior pixel _{(x, y)}With consistance coh _{(x, y)}:

$M_{(x,y)}=\sqrt{G_{\mathrm{xx}(x,y)}+G_{\mathrm{yy}(x,y)}},$

${\mathrm{coh}}_{(x,y)}=\frac{\sqrt{{(G_{\mathrm{xx}(x,y)}-G_{\mathrm{yy}(x,y)})}^2+{4G}_{\mathrm{xy}(x,y)}^2}}{G_{\mathrm{xx}(x,y)}+G_{\mathrm{yy}(x,y)}},$

Wherein,

expression gradient consistance in the horizontal direction

expression gradient consistance in vertical direction

The expression gradient in the horizontal direction with vertical direction on related consistance, W ₀The piece of representing fingerprint image to divide, G _{X (u, v)}And G _{Y (u, v)}Be respectively the gradient on level and the vertical direction, the horizontal ordinate of u remarked pixel point in piece, the ordinate of v remarked pixel point in piece;

(1.3), calculate the quality of each piece fingerprint image according to the gradient mean value and the consistance of each piece fingerprint image interior pixel:

q _(x，y)＝f(M _(x，y)，M ₁,M ₂)· ,C ₁,C ₂)，

In the formula, M ₁Be the lower limit of field of direction average in the piece, M ₂Be the upper limit of field of direction average in the piece, C ₁Be the conforming lower limit of the field of direction in the piece, C ₂Be the conforming upper limit of the field of direction in the piece,

$f(M_{(x,y)},{\mathrm{<figure-callout\; id="1"\; label="M"\; filenames="HDA00001713672600031.png"\; state="\{\{state\}\}">M</figure-callout>}}_1,M_2)=\left\{\begin{array}{cc}& \\ 0& \mathrm{If}{M}_{(x,y)}\&le;{\mathrm{<figure-callout\; id="1"\; label="M"\; filenames="HDA00001713672600031.png"\; state="\{\{state\}\}">M</figure-callout>}}_1\\ 1& \mathrm{If}{M}_{(x,y)}>M_2\\ \frac{M_{(x,y)}-M_1}{M_2-{\mathrm{<figure-callout\; id="1"\; label="M"\; filenames="HDA00001713672600031.png"\; state="\{\{state\}\}">M</figure-callout>}}_1}& \mathrm{Otherwise}\end{array}\right.,$ Be piece field of direction average normalized function,

$f({\mathrm{Coh}}_{(x,y)},{\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA00001713672600031.png"\; state="\{\{state\}\}">C</figure-callout>}}_1,C_2)=\left\{\begin{array}{cc}& \\ 0& \mathrm{If}{\mathrm{Coh}}_{(x,y)}\&le;{\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA00001713672600031.png"\; state="\{\{state\}\}">C</figure-callout>}}_1\\ 1& \mathrm{If}{\mathrm{Coh}}_{(x,y)}>C_2\\ \frac{{\mathrm{Coh}}_{(x,y)}-C_1}{C_2-{\mathrm{<figure-callout\; id="1"\; label="C"\; filenames="HDA00001713672600031.png"\; state="\{\{state\}\}">C</figure-callout>}}_1}& \mathrm{Otherwise}\end{array}\right.,$ Be piece field of direction consistance normalized function;

The piece field of direction of the fingerprint image that (1.4) step (1) is extracted is converted into a continuous vector field:

v _(x,y)＝(v _1(x，y),v _2(x，y))，

Wherein, v _{1 (x, y)}=cos (2 θ (x, y)), v _{2 (x, y)}=sin (2 θ (x, y));

(1.5) utilize the continuous vector field of step (1.4) extraction and the picture quality of step (1.3) calculating to set up the associating energy function:

J(u _(x，y))＝D(u _(x，y))+με(u _(x，y))，

Wherein, u _{(x, y)}=(u _{1 (x, y)}, u _{2 (x, y)}) be the field of direction after the normalization to be asked, u _{1 (x, y)}Expression waits to ask the cosine value of the piece field of direction, u _{2 (x, y)}Expression waits to ask the sine value of the piece field of direction,

$D\left(u_{(x,y)}\right)=\frac{1}{2}{\&Integral;}_{\mathrm{\&Omega;}}{q}_{(x,y)}{\left|\right|u_{(x,y)}-v_{(x,y)}\left|\right|}^2{d}_x{d}_y,$ Be source vector field v _{(x, y)}With the vector field u after the normalization _{(x, y)}Between the difference item,

Be penalty term, Ω representes the effective coverage of fingerprint image, and μ is the normalized parameter of decision difference item and penalty term relation, || ... || ²Represent 1 norm square;

(1.6) find the solution feasible associating energy function J (u _{(x, y)}) obtain the u of minimum value _{(x, y)}, draw the field of direction after the normalization Atan wherein ^-1Represent arc tangent, draw the field of direction image B of fingerprint image.

Step 2 is carried out complex filter according to the field of direction after the normalization to fingerprint image and is extracted characteristics of image, and aims at field of direction image B, sets up proper vector.

(2.1) fingerprint image is carried out first order complex filter, the polynomial expression that uses the Gaussian window function to generate is approximately complex filter: (x+iy) ^mg _{σ (x, y)}, wherein

σ representes the variance of Gaussian function, and x representes the horizontal ordinate of fingerprint image, and y representes the ordinate of fingerprint image, and x+iy representes the plural number of constructing, and m representes the progression of complex filter, and the present invention only uses first order filtering, m=± 1;

(2.2) with the forward function T of first order complex filter _cWith the negative sense function T _dBe expressed as respectively:

T _c＝(x+iy)g _σ(x,y)，

T _d＝(x+iy) ^-1g _σ(x,y)；

(2.3) the tensor z of the field of direction in complex field is expressed as:

z＝cos(2θ)+isin(2θ)；

(2.4) according to the function of complex filter and the tensor of complex field, the filter response of calculated complex wave filter:

The forward function and the field of direction of complex filter are carried out convolution at the tensor of complex field, and the filter response that draws forward is: R _c=z*T _c,

The negative sense function and the field of direction of complex filter are carried out convolution at the tensor of complex field, and the filter response that draws negative sense is: R _d=z*T _d

(2.5) field of direction image is aimed at filter response:

(2.5a) setting the weights function is:

${g^{\&prime;}}_{(x,y)}=\exp (-\frac{{(x-x_0)}^2+{(y-y_0)}^2}{2{\mathrm{\&sigma;}}^{\&prime;}}),$

X wherein ₀And y ₀The coordinate at expression fingerprint image center, σ '=[min (width, height)/3] ², expression gaussian kernel parameter, width representes the width of fingerprint image, and height representes the height of fingerprint image, and min representes to get minimum value;

(2.5b) according to forward filter response R _cWith weights function g' _{(x, y)}, the structure filter response at the projection function of vertical direction is:

h _(x，y)＝R _c··，

Wherein, n _(pi/2)Vector of unit length on the expression vertical direction;

(2.5c) find the solution and make projection function h _{(x, y)}Get peaked x, the y coordinate figure is making h _{(x, y)}Get peaked point and be made as RP, move to picture centre to RP to realize aiming at of field of direction image and filter response;

(2.6) with complex filter response with aim at after field of direction image be characteristic establishment proper vector p:

p＝vec[real(t)?image(t)]，

Wherein, vec [] representes vector operations, the real part of real () expression fingerprint image, and the imaginary part of image () expression fingerprint image, t representes fingerprint image characteristics:

t＝vec[vec(z)?vec(R _c)?vec(R _d)]。

Step 3, design classification device is classified to fingerprint.

(3.1) response and its maximum response threshold value E with every bit in the fingerprint image behind the complex filter compares, if certain any response, judges then that this is a core point greater than maximum response threshold value E, otherwise this point is not a core point, and wherein 0.4 < E < 0.8;

(3.2) peak response and its minimum response threshold value F with complex filter compares, if the peak response of complex filter less than minimum response threshold value F, and does not have core point in the fingerprint image; Judge that then fingerprint is arch form A, otherwise, execution in step (3.3); Wherein 0.3 < F < 0.7, and F < E;

(3.3) through piece field of direction image B being carried out the take the fingerprint some field of direction of image I of bilinear interpolation computing; According to the field of direction sampled point on the crestal line that takes the fingerprint; Ask every pair of vector of unit length on the neighbouring sample point line direction; If neighbouring sample point is a; B, a then, the vector of unit length on the b line direction is

and calculate by following formula:

Wherein representes the vector of sampled point a to sampled point b line direction, | ... | the expression subtend is measured mould.

(3.4) vector of unit length of setting on first sampled point and second the sampled point line direction is an initial vector, is ordinate with the inner product of each vector of unit length and initial vector of unit length, and the sampled point sequence is a horizontal ordinate; Make the crestal line circle of equal altitudes; The crestal line circle of equal altitudes is as shown in Figure 2 with its corresponding fingerprint pattern, and wherein Fig. 2 (a) representes the crestal line flow graph of screw-type fingerprint image, the corresponding crestal line circle of equal altitudes of Fig. 2 (b) expression screw-type fingerprint image; The crestal line flow graph of Fig. 2 (c) expression dextrorotatory form fingerprint image; The corresponding crestal line circle of equal altitudes of Fig. 2 (d) expression dextrorotatory form fingerprint image, the crestal line flow graph of the left-handed type fingerprint image of Fig. 2 (e) expression, the corresponding crestal line circle of equal altitudes of the left-handed type fingerprint image of Fig. 2 (f) expression; The crestal line flow graph of Fig. 2 (g) expression arch form fingerprint image, the corresponding crestal line circle of equal altitudes of Fig. 2 (h) expression arch form fingerprint image;

(3.5) if two local maximums are arranged between 0.8 to 1 in the crestal line circle of equal altitudes, two local minimums judge that then fingerprint is screw-type W between-0.8 to-1, otherwise, execution in step (3.5);

(3.6) with PCA the proper vector p of setting up in the step (2.6) is carried out dimensionality reduction, promptly earlier a plurality of variablees among the proper vector p are carried out linear transformation, set up incoherent in twos significant variable; Again set up proper vector with the significant variable of negligible amounts then, the proper vector after using K arest neighbors sorting algorithm to dimensionality reduction is carried out rough sort, finds out two types near true fingerprint pattern; If comprise left-handed type L or dextrorotatory form R in these two types, do not comprise screw-type W, local maximum is between 0.8 to 1 in the crestal line circle of equal altitudes; Local minimum is between-0.8 to-1, and the terminal point of fingerprint ridge line up-sampling point judges then that all in the left side of initial point fingerprint is left-handed type L; If the terminal point of sampled point is all on the right side of initial point; Judge that then fingerprint is dextrorotatory form R, otherwise, execution in step (3.6);

(3.7) two types that K arest neighbors sorting algorithm rough sort are come out are input to SVMs, and SVMs is trained, and the classification results of its SVMs output is final fingerprint pattern.

Effect of the present invention can specify through following experiment

1. experiment condition:

The experimental result that the field of direction is extracted among the present invention is tested on the online fingerprint algorithm test platform of FVC-on Going, and the database that the field of direction is extracted and fingerprint classification is used is the NIFT-4 fingerprint database.

2. experiment content:

Experiment 1 is extracted the field of direction with the inventive method and existing gradient method, and its result is as shown in Figure 3, and it is as shown in table 1 that the field of direction is extracted experimental data;

Experiment 2 is classified to fingerprint with the inventive method and existing fingerprint classification method, and the fingerprint classification experimental data is as shown in table 2.

3. experimental result:

Fig. 3 has provided field of direction extraction algorithm commonly used and field of direction extraction algorithm effect contrast figure of the present invention; Wherein Fig. 3 (a) representes the source fingerprint image, Fig. 3 (b) expression source field of direction image, the field of direction image that Fig. 3 (c) expression gradient method is extracted; The field of direction image that the method that Fig. 3 (d) expression the present invention proposes is extracted; Can see that from Fig. 3 the field of direction image effect that method proposed by the invention is extracted is best, and is the highest with source field of direction image similarity.

Table 1 has provided the experimental data of the inventive method and the existing gradient method extraction field of direction; Every kind of method is tested 2 data sets in the NIFT-4 fingerprint database respectively in the test, and each data acquisition comprises image and 50 low-quality images of 10 panel height quality.

Table 1

The average error rate that can find out direction of fingerprint of the present invention field method for distilling from table 1 is lower than existing method, and particularly on low-quality fingerprint image, direction of fingerprint of the present invention field method for distilling significantly is superior to additive method on degree of accuracy.

Table 2 provides the experimental data that the inventive method and existing fingerprint classification method are classified to fingerprint.

Table 2

Algorithm	Characteristic	Sorter	5-class	4-class	Experiment is provided with
						Cappelli etc.	?OI	MKL?and?SPD	95.2％	96.3％	2nd?half
Candela etc.	?OI	NN	-	88.6％	2nd?half
						Jain etc.	?GF	k-NN?and?NN	90％	95.8％	2nd?half?1.8％rejects
Liu	?SP	AbDT	94.1％	95.7％	2nd?half
						Li etc.	?SP?and?OI	SVM	93.5％	95％	2nd?half
Hong etc.	?SP?and?GF	SVM?and?NB	90.8％	94.9％	2nd?half
						Zhang?and?Yan	?SP?and?RF	Rule?based	84.3％	92.7％	2nd?half
Dass?and?Jain	?RF	Rule?based	-	94.4％	Whole
						Chang?and?Fan	?RF	Rule?based	94.8％	-	First?impression
Algorithm of the present invention	?RF.?OI?and?CF	k-NN，Rule?based?S	95.9％	97.2％	2nd?half

OI represents the direction of fingerprint field in the table 2, and SP represents singular point, and on behalf of crestal line, RF flow to; CF represents complex filter, and GF represents the gabor wave filter, and NN represents neural network; K-NN represents k arest neighbors sorting algorithm; SVM represents SVMs, and NB represents Bayes classifier, and arch and tent camber are not distinguished among the 4-class.

Can find out that from table 2 degree of accuracy of the present invention in 5-class and 4-class all is superior to other existing algorithms.

Claims (6)

1. one kind is passed through the method that the classification device is realized fingerprint classification, may further comprise the steps:

(1) use Fast Fourier Transform (FFT) that fingerprint image is strengthened, the fingerprint image I after strengthening be divided into the piece of w * w, and with gradient method take the fingerprint image I piece field of direction θ (x, y), x wherein, y is the horizontal ordinate and the ordinate of presentation video piece respectively;

(2) calculate the quality q of each piece fingerprint image according to the piece field of direction _{(x, y)}

The piece field of direction of the fingerprint image that (3) step (1) is extracted is converted into a continuous vector field:

v _(x，y)＝(v _1(x，y),v _2(x，y))，

Wherein, v _{1 (x, y)}=cos (2 θ (x, y)), v _{2 (x, y)}=sin (2 θ (x, y));

(4) utilize the continuous vector field of step (3) extraction and the picture quality of step (2) calculating to set up the associating energy function:

J(u _(x，y))＝D(u _(x，y))+με(u _(x，y))，

Wherein, u _{(x, y)}=(u _{1 (x, y)}, u _{2 (x, y)}) be the field of direction after the normalization to be asked, u _{1 (x, y)}Expression waits to ask the cosine value of the piece field of direction, u _{2 (x, y)}Expression waits to ask the sine value of the piece field of direction,

Be source vector field v _{(x, y)}With the vector field u after the normalization _{(x, y)}Between the difference item,

Be penalty term, Ω representes the effective coverage of fingerprint image, and μ is the normalized parameter of decision difference item and penalty term relation, || ... || ²Represent 1 norm square;

(5) find the solution feasible associating energy function J (u _{(x, y)}) obtain the u of minimum value _{(x, y)}, draw the field of direction after the normalization Atan wherein ^-1Represent arc tangent, draw the field of direction image B of fingerprint image;

(6) with the field of direction after the normalization fingerprint image I is carried out complex filter, and judges the type of fingerprint:

(6a) response and its maximum response threshold value E with every bit in the fingerprint image behind the complex filter compares, if certain any response, judges then that this is a core point greater than maximum response threshold value E, otherwise this point is not a core point, and wherein 0.4 < E < 0.8;

(6b) peak response and its minimum response threshold value F with complex filter compares, if the peak response of complex filter less than minimum response threshold value F, and does not have core point in the fingerprint image; Judge that then fingerprint is arch form A, otherwise, execution in step (6c); Wherein 0.3 < F < 0.7, and F < E;

(6c) through piece field of direction image B being carried out the take the fingerprint some field of direction of image I of bilinear interpolation computing,, ask every pair of vector of unit length on the neighbouring sample point line direction according to the field of direction sampled point on the crestal line that takes the fingerprint;

The vector of unit length of (6d) setting on first sampled point and second the sampled point line direction is an initial vector, is ordinate with the inner product of each vector of unit length and initial vector of unit length, and the sampled point sequence is a horizontal ordinate; Make the crestal line circle of equal altitudes; If two local maximums are arranged between 0.8 to 1 in the crestal line circle of equal altitudes, two local minimums judge that then fingerprint is screw-type W between-0.8 to-1; Otherwise, execution in step (6e);

(6e) set up proper vector according to the response of field of direction image B and complex filter every bit, with PCA proper vector is carried out dimensionality reduction, the proper vector after using K arest neighbors sorting algorithm to dimensionality reduction is carried out rough sort; Find out near two types of true fingerprint pattern,, do not comprise screw-type W if comprise left-handed type L or dextrorotatory form R in these two types; Local maximum is between 0.8 to 1 in the crestal line circle of equal altitudes; Local minimum is between-0.8 to-1, and the terminal point of fingerprint ridge line up-sampling point judges then that all in the left side of initial point fingerprint is left-handed type L; If the terminal point of sampled point is all on the right side of initial point; Judge that then fingerprint is dextrorotatory form R, otherwise, execution in step (6f);

(6f) use SVMs that two types of K arest neighbors sorting algorithm output are further classified, the classification results of its SVMs output is final fingerprint pattern.

2. the method through classification device realization fingerprint classification according to claim 1, the wherein said quality q that calculates each piece fingerprint image according to the piece field of direction of step (2) _{(x, y)}, calculate by following formula:

q _(x，y)＝f(M _(x，y)，M ₁,M ₂)· ,C ₁,C ₂)，

In the formula, M ₁Be the lower limit of field of direction average in the piece, M ₂Be the upper limit of field of direction average in the piece, C ₁Be the conforming lower limit of the field of direction in the piece, C ₂Be the conforming upper limit of the field of direction in the piece, M _{(x, y)}The average of the field of direction in the expression piece, coh _{(x, y)}The consistance of the field of direction in the expression piece,

is piece field of direction average normalized function

is piece field of direction consistance normalized function.

3. the method that realizes fingerprint classification through the classification device according to claim 1; Wherein step (6c) is described asks every couple of neighbouring sample point a, and the vector of unit length on the b line direction

calculates by following formula:

Wherein

representes the vector of sampled point a to sampled point b line direction, | ... | the expression subtend is measured mould.

4. the method through classification device realization fingerprint classification according to claim 1, wherein step (6e) is described carries out dimensionality reduction with PCA to proper vector, carries out as follows:

(6e1) a plurality of variablees in the proper vector of setting up in the step (6e) are carried out linear transformation, set up incoherent in twos significant variable;

(6e2) set up proper vector again with the significant variable of negligible amounts.

5. the method that realizes fingerprint classification through the classification device according to claim 1; Wherein the proper vector of the described use K of step (6e) arest neighbors sorting algorithm after to dimensionality reduction carried out rough sort; Find out with the immediate K of an actual fingerprint image neighborhood, the type that wherein comprises two maximum neighborhood representatives of sample size is immediate two types with true fingerprint pattern.

6. according to claim 1ly realize the method for fingerprint classification through the classification device, wherein the described use SVMs of step (6f) is further classified to two types of K arest neighbors sorting algorithm output, carries out as follows:

(6f1) import two types that K arest neighbors sorting algorithm rough sort is come out;

(6f2) SVMs is trained with the rough sort type of input;

(6f3) output of the SVMs after will training is as final fingerprint pattern.

Images