CN103839072A - False fingerprint detecting method based on naive Bayes classifiers - Google Patents

Abstract

A false fingerprint detecting method based on naive Bayes classifiers comprises the following steps of (1) training library dividing; (2) image normalization; (3) feature extraction, wherein the feature extraction comprises (3.1) discrete wavelet transformation, (3.2) noise removal, (3.3) wavelet reconstruction, (3.4) noise pattern estimation, (3.5) standard deviation chart calculation, and (3.6) standard deviation chart dividing, and image feature obtaining through statistic conducting; (4) feature dividing; (5) classifier training; (6) classifier performance estimation; (7) classifier fusion: utilizing a construction method of the naive Bayes classifiers for conducting fusion to obtain a new classifier. According to the method, the requirement for the performance of a single classifier is not high, but the effect of the classifiers after fusion is quite good.

Description

A kind of false fingerprint detection method based on Naive Bayes Classifier

Technical field

The present invention relates to the technical fields such as image processing, machine learning, pattern-recognition, especially a kind of true and false detection method for fingerprint image.

Background technology

Image processing, feature extraction, sorter training and Images Classification etc. are the important steps in false fingerprint detection method.

Y.S.Moon etc. have proposed a kind of false fingerprint detection algorithm based on noise, and the true and false fingerprint image that is 1000dpi to 200 resolution tests, and all classification is correct.L.F.A.Pereria etc. are according to the paper of Moon, the false fingerprint detection algorithm based on noise of the fingerprint image that further to have proposed for resolution be 500dpi, and the sagem storehouse of LivDet2011 is tested, on average sentencing wrong rate is 12.8%.

Detection side's ratio juris based on noise, has introduction in the paper of Y.S.Moon, and general principle is: the material of the false fingerprint of making of main flow is as clay and gelatin, in the process of the false fingerprint of making, owing to comprising large organic molecule, generally can condense into piece, this has just caused the coarse of false fingerprint.Therefore, in general, the finger tip surface of false fingerprint is more coarse than live body.This characteristic can be used as the foundation of false fingerprint detection algorithm, its showing as in data: the noise of the true fingerprint image of noise ratio of false fingerprint image has larger standard deviation.Due to the algorithm proposing in the Moon paper fingerprint image that is 1000dpi for resolution, and current generally use is the fingerprint acquisition instrument of 500dpi, in order to check the applicability of its algorithm to 500dpi image, L.F.A.Pereria tests with the sagem fingerprint base of LivDet2011, on average sentencing wrong rate is 42.8%, differs greatly with current main flow algorithm performance.Therefore L.F.A.Pereria etc. improves this algorithm, propose to analyze (Spatial surface coarseness analysis based on spatial surface roughness, be called for short SSCA) false fingerprint detection method, mainly feature extraction is improved, by noise pattern piecemeal and ask standard deviation, to the standard deviation figure piecemeal obtaining statistics, last statistical value is its feature the most, comprehensive all fritters, obtain the feature of fingerprint image.It is 12.8% that this algorithm is tested the wrong rate of on average sentencing obtaining on the sagem of LivDet2011 fingerprint image storehouse.

Support vector machine is a kind of method of supervising formula study, can be widely used in statistical classification and regretional analysis.Support vector machine belongs to vague generalization linear classifier, is also considered to put forward a special case of clo husband standardization (Tikhonov Regularization) method.The feature of this sorter be can simultaneous minimization experience error with maximize Geometry edge district, therefore support vector machine is also referred to as maximal margin region class device.

Naive Bayes Classifier is the Bayesian simple probability sorter of a kind of application based on independent hypothesis, and also can describe more accurately this potential probability model is independent characteristic model.The basis of Bayes's classification is probability inference, exactly at the probability of occurrence of only knowing each condition, and in the uncertain situation whether it exists, how to complete reasoning and decision task.And Naive Bayes Classifier is based on independent hypothesis, suppose that the each feature of sample is uncorrelated with other features, even these features interdepend or some feature by other characteristics determined, all think that these attributes are independently in decision probability distribution.Naive Bayes Classifier relies on accurate natural probability model, in the sample of supervised learning, can obtain extraordinary classifying quality.In many practical applications, model-naive Bayesian parameter estimation is used maximum Likelihood, does not use Bayesian probability or any Bayesian model.Although be by simple thought and the hypothesis of too simplifying, Naive Bayes Classifier still can be obtained goodish effect in a lot of complicated real-world situation.2004, one piece of article of analyzing Bayes classifier problem has disclosed some Naive Bayes Classifiers theoretically obtained the reason of mysterious classifying quality.An advantage of Naive Bayes Classifier is only need to just can estimate necessary parameter according to a small amount of training data.Can know by simple probabilistic method analysis, the current ubiquitous shortcoming of false fingerprint detection method is only to classify with single classifier, but the result of the false fingerprint detection algorithm of single classifier is reliable not.

Summary of the invention

In order to make up the insecure deficiency of single sorter classification results of existing false fingerprint detection method, the invention provides the good false fingerprint detection method based on Naive Bayes Classifier of a kind of classification results reliability.

The technical solution adopted for the present invention to solve the technical problems is:

Based on a false fingerprint detection method for Naive Bayes Classifier, comprise the following steps:

1) training storehouse is divided: training storehouse image is divided into equal two parts at random, is designated as A, B;

2) image normalization: utilize fingerprint image normalized function that picture size is unified into m pixel × n pixel (m, the multiple that the value of n is 4 can get 256,512 etc. conventionally);

3) feature extraction: image is carried out to noise extraction and processing, comprise following process:

3.1) image is carried out to wavelet transform, obtain a low frequency part and six HFSs;

3.2) to hyperbolic shrinkage method denoising for six HFSs after conversion;

3.3) according to step 3.2) six HFSs and the low frequency part before that obtain carry out wavelet reconstruction, obtains the image after denoising;

3.4) by image subtraction after original image and denoising, obtain noise pattern;

3.5) noise pattern is divided into individual fritter, wherein p _xfor the width of fritter and the ratio of original image width, p _yfor the ratio of length and the original image length of fritter, calculate the standard deviation of each piece, obtain standard deviation figure;

3.6) calculate the maximal value in standard deviation figure, be designated as S, interval [0, S] is divided into

(be designated as k) equal portions, obtain k interval, standard deviation figure is divided into

individual fritter, wherein q _xfor the width of fritter and the ratio of original image width, q _yfor the ratio of length and the original image length of fritter, add up each fritter Plays difference and fall into the number in k interval, obtain k value, using this as this little block feature, comprehensively each piece, obtains this characteristics of image;

4) feature is divided: feature is divided into 4 parts, corresponds respectively to four parts in upper left, upper right, lower-left, bottom right of image;

5) sorter training: train respectively the feature after division with SVM, obtain sorter, can obtain altogether 8 sorters for A, two training storehouses of B;

6) classifier performance assessment:

6.1) for 4 sorters in A, the feature obtaining with correspondence position in B is respectively tested, and adds up sentencing the wrong rate of sentencing of rate and true fingerprint of false fingerprint;

6.2) in like manner, for 4 sorters in B, the feature obtaining with correspondence position in A is respectively tested, and adds up sentencing the wrong rate of sentencing of rate and true fingerprint of false fingerprint;

7) Multiple Classifier Fusion:

7.1) for a unknown fingerprint image, think that it is that the probability of false fingerprint is prior probability, after the classification results of comprehensive 8 sorters, in the time being judged to the probability of false fingerprint and being more than or equal to probability threshold value T, think that this fingerprint is false fingerprint;

7.2) for 8 sorters that obtain before, be designated as respectively f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇, f ₈, note C is that { 0,1}, wherein 0 represents false fingerprint, and 1 represents true fingerprint in classification set;

7.3) for these 8 sorters, think between them it is separate, therefore, in the result of known 8 sorters, and the performance of 8 sorters, then use Bayesian formula, note P1=∏ P (C=0|f _i(X)), P ₂=∏ P (C=1|f _i(X)), obtaining a fingerprint image is false fingerprint probability P=P1/ (P1+P2), when P is greater than T, is judged to 0, otherwise is 1, P (C=0|f wherein _i(X)) all can be obtained by Bayesian formula the sorter F after therefore being merged;

8) false fingerprint detection: the image that detect is carried out to 1), 2), 3) operation of step, then the proper vector obtaining is classified with sorter F.

Technical conceive of the present invention is: in to existing false fingerprint detection algorithm paractical research, find, with single classifier, image is classified, the classification results obtaining is reliable not.Can treat this problem from the angle of probability, when detection in real time, suppose that fingerprint to be measured is that the probability of false fingerprint is 0.01, there is a sorter, it sentences probability to be a, to be b to the wrong probability of sentencing of true fingerprint, obtain so according to Bayesian formula to false fingerprint, test fingerprint is false fingerprint, and to be classified that device is judged to be the probability of false fingerprint

further hypothesis is judged to false fingerprint in the time of p> prior probability 0.5, so, can obtain a relational expression about a, b: a > 99 × b, because a<1, so b<1/99.For this precision, the detection method of current single classifier can't reach, and has therefore just considered the method merging here.And this process also can be analyzed from another angle, single classifier is judged to false fingerprint, make to judge that this fingerprint is that the probability of false fingerprint has improved, but this probability is also not enough to judge that it is false fingerprint.According to this thinking, fusion method has been proposed, first determine a series of separate sorter, the required independently feature of structure using them as Naive Bayes Classifier, fingerprint is classified with each sorter, for the result of each classification, can increase or reduce it is the probability of false fingerprint.Therefore, as long as there is the sorter with certain classification capacity of sufficient amount, just can meet the requirement that true and false fingerprint is classified, this is also the thought of Naive Bayes Classifier structure.

On the basis of the method based on SSCA that the present invention proposes at L.F.A.Pereria, first construct multiple independently sorters, independent characteristic using each sorter as false fingerprint image, then merge and obtain last sorter by the method for structure Naive Bayes Classifier.

Beneficial effect of the present invention is: less demanding to single classifier performance, but effect after Multiple Classifier Fusion can be very good, and classification results reliability is good.

Accompanying drawing explanation

Fig. 1 is a kind of false fingerprint detection method process flow diagram based on Naive Bayes Classifier.

Fig. 2 is the process flow diagram of feature extraction.

Fig. 3 is the schematic diagram of feature extraction.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1, a kind of false fingerprint detection method based on Naive Bayes Classifier, described fingerprint detection method comprises the following steps:

1) training storehouse is divided: training storehouse image is divided into equal two parts at random, is designated as A, B;

2) image normalization: utilize fingerprint image normalized function that picture size is unified into m pixel × n pixel;

Just there have been the concrete steps of following feature extraction with reference to Fig. 2 and Fig. 3.

3) feature extraction: be the process that picture noise is extracted and processed, comprise following process:

A) image f (x, y) is carried out to two-dimensional discrete wavelet conversion, obtain a low frequency part and six HFS g _k(x, y), k ∈ { 1,2,3,4,5,6};

B) to hyperbolic shrinkage method denoising for six HFSs after conversion:

$\mathrm{\&delta;}=\sqrt{2\log \left(N\right)\mathrm{\&sigma;}}$

Wherein N is g _kthe signal length of (x, y), σ is the standard deviation by 3 details of the first dimension of conversion decomposition;

C) six HFSs and the low frequency part before that obtain according to step a) are carried out wavelet reconstruction, obtain image f'(x, y after denoising);

D) by image subtraction after original image and denoising, obtain noise pattern;

E) noise pattern is divided into

individual fritter, calculates the standard deviation of each piece, obtains standard deviation figure;

F) calculate the maximal value in standard deviation figure, be designated as S, interval [0, S] is divided into (be designated as k) equal portions, obtain k interval.Standard deviation figure is divided into

individual fritter, adds up each fritter Plays difference and falls into the number in k interval, obtains k value, and using this as this little block feature, comprehensively each piece, obtains this characteristics of image;

4) feature is divided: feature is divided into 4 parts, corresponds respectively to four parts in upper left, upper right, lower-left, bottom right of image;

5) sorter training: for the feature of each class, all train with SVM, obtain sorter, so just altogether obtain 8 sorters;

6) classifier performance is estimated:

A) for 4 sorters in A, the feature obtaining with correspondence position in B is respectively tested, and adds up sentencing the wrong rate of sentencing of rate and true fingerprint of false fingerprint;

B) in like manner, for 4 sorters in B, the feature obtaining with correspondence position in A is respectively tested, and adds up sentencing the wrong rate of sentencing of rate and true fingerprint of false fingerprint;

7) Multiple Classifier Fusion:

A) for a unknown fingerprint image, consider that it is that the probability (being prior probability) of false fingerprint is 0.5, and at the classification results of comprehensive 8 sorters, when being the probability of false fingerprint while being more than or equal to probability threshold value T, think false fingerprint;

B) for 8 sorters that obtain before, be designated as respectively f1, f2, f3, f4, f5, f6, f7, f8, note C is that { 0,1}, wherein 0 represents false fingerprint, and 1 represents true fingerprint in classification set;

C) for these 8 sorters, think between them it is separate, therefore, in the result of known 8 sorters, and the performance of 8 sorters, then use Bayesian formula, note P1=∏ P (C=0|f _i(X)), P2=∏ P (C=1|f _i(X)), can obtain a fingerprint image is false fingerprint probability (being posterior probability) P=P1/ (P1+P2), when P is greater than T, is judged to 0, otherwise is 1, P (C=0|f wherein _i(X)) all can be obtained by Bayesian formula the sorter F after therefore being merged;

8) false fingerprint detection: the image that detect is carried out to 1), 2), 3) operation of step, then the proper vector obtaining is classified with sorter F.

Last table 1 is the test result to LivDet2011 match image library by this detection method.

Table 1.

Claims (1)

1. the false fingerprint detection method based on Naive Bayes Classifier, is characterized in that: described false fingerprint detection method comprises the following steps:

1) training storehouse is divided: training storehouse image is divided into equal two parts at random, is designated as A, B;

2) image normalization: utilize fingerprint image normalized function that picture size is unified into m pixel × n pixel, m, the multiple that the value of n is 4;

3) feature extraction: image is carried out to noise extraction and processing, comprise following process:

3.1) image is carried out to wavelet transform, obtain a low frequency part and six HFSs;

3.2) to hyperbolic shrinkage method denoising for six HFSs after conversion;

3.3) according to step 3.2) six HFSs and the low frequency part before that obtain carry out wavelet reconstruction, obtains the image after denoising;

3.4) by image subtraction after original image and denoising, obtain noise pattern;

3.5) noise pattern is divided into

individual fritter, wherein p _xfor the width of fritter and the ratio of original image width, p _yfor the standard deviation of the length of fritter and each piece of ratio calculation of original image length, obtain standard deviation figure;

3.6) calculate the maximal value in standard deviation figure, be designated as S, interval [0, S] is divided into k part,

obtain k interval, standard deviation figure is divided into

individual fritter, wherein q _xfor the width of fritter and the ratio of original image width, q _yfor the ratio of length and the original image length of fritter, add up each fritter Plays difference and fall into the number in k interval, obtain k value, using this as this little block feature, comprehensively each piece, obtains this characteristics of image;

4) feature is divided: feature is divided into 4 parts, corresponds respectively to four parts in upper left, upper right, lower-left, bottom right of image;

5) sorter training: train respectively the feature after division with SVM, obtain sorter, can obtain altogether 8 sorters for A, two training storehouses of B;

6) classifier performance assessment:

6.1) for 4 sorters in A, the feature obtaining with correspondence position in B is respectively tested, and adds up sentencing the wrong rate of sentencing of rate and true fingerprint of false fingerprint;

6.2) in like manner, for 4 sorters in B, the feature obtaining with correspondence position in A is respectively tested, and adds up sentencing the wrong rate of sentencing of rate and true fingerprint of false fingerprint;

7) Multiple Classifier Fusion:

7.1) for a unknown fingerprint image, think that it is that the probability of false fingerprint is prior probability, after the classification results of comprehensive 8 sorters, in the time being judged to the probability of false fingerprint and being more than or equal to probability threshold value T, think that this fingerprint is false fingerprint;

7.2) for 8 sorters that obtain before, be designated as respectively f ₁, f ₂, f ₃, f ₄, f ₅, f ₆, f ₇, f ₈, note C is that { 0,1}, wherein 0 represents false fingerprint, and 1 represents true fingerprint in classification set;

7.3) for these 8 sorters, think between them it is separate, therefore, in the result of known 8 sorters, and the performance of 8 sorters, then use Bayesian formula, note P1=∏ P (C=0|f _i(X)), P ₂=∏ P (C=1|f _i(X)), obtaining a fingerprint image is false fingerprint probability P=P1/ (P1+P2), when P is greater than T, is judged to 0, otherwise is 1, P (C=0|f wherein _i(X)) all can be obtained by Bayesian formula the sorter F after therefore being merged;

8) false fingerprint detection: the image that detect is carried out to 1), 2), 3) operation of step, then the proper vector obtaining is classified with sorter F.

Images