CN103942540A

CN103942540A - False fingerprint detection algorithm based on curvelet texture analysis and SVM-KNN classification

Info

Publication number: CN103942540A
Application number: CN201410143345.2A
Authority: CN
Inventors: 张永良; 方珊珊; 谢瑜
Original assignee: HANGZHOU JINGLIANWEN TECHNOLOGY Co Ltd
Current assignee: HANGZHOU JINGLIANWEN TECHNOLOGY Co Ltd
Priority date: 2014-04-10
Filing date: 2014-04-10
Publication date: 2014-07-23

Abstract

The invention discloses a false fingerprint detection algorithm based on curvelet texture analysis and SVM-KNN classification. The false fingerprint detection algorithm based on the curvelet texture analysis and the SVM-KNN classification comprises the following steps of curvelet transformation; curvelet reconstruction; curvelet coefficient feature extraction, extraction of coefficient energy, entropy and the like, and coefficient first-order statistical magnitude extraction; textural feature extraction, and extraction of a first-order statistical magnitude, a gray-level co-occurrence matrix, MRF features and the like; training of classifiers; performance evaluation of the classifiers; false fingerprint detection, wherein testing is carried out on samples by utilizing the trained classifiers. The false fingerprint detection algorithm based on the curvelet texture analysis and the SVM-KNN classification is suitable for various fingerprint acquisition instruments of 500 dpi resolution, analysis is respectively carried out on image high-frequency noise information and texture information after denoising, and the noise and texture differences between true fingerprints and false fingerprints are quantified.

Description

False fingerprint detection algorithm based on cockle texture analysis and SVM-KNN classification

Technical Field

The invention relates to the field of image processing, pattern recognition and other related technologies, in particular to a false fingerprint detection algorithm based on texture analysis and SVM-KNN classification, and aims to detect fingerprints to distinguish true fingerprints from false fingerprints.

Background

The false fingerprint generally refers to a fingerprint which is made of materials such as silica gel, latex, plasticine and the like and simulates human fingerprint textures, and the algorithm mainly comprises the important links such as image processing, feature extraction, classifier training, image classification and the like.

With the development of economy and society, biometric identification technology, especially fingerprint technology, is being widely used. However, the appearance of false fingerprints made of cheap materials and the improvement of the manufacturing technology thereof bring potential safety hazards to fingerprint feature identification. Research shows that the traditional fingerprint acquisition instrument can be deceived by using a fake fingerprint film made of cheap materials such as silica gel, gelatin, plasticine and the like according to fingerprints left by a user. Attempts to create unauthorized access, known as spoofing, are made using a false fingerprint containing information about the identity of the authorized user. Such attack methods against the acquisition instrument are often employed when one wishes to conceal the true identity or gain the privilege of an authorized user. In order to resist the attack of false fingerprints on the fingerprint identification system, a false fingerprint detection technology is developed. The false fingerprint detection technology is a method for judging whether a fingerprint sample comes from a live fingerprint, and the detection method can be divided into two types: the first type uses characteristics of finger temperature, skin conductivity, pulse blood oxygen and the like, which can be detected by adding extra hardware equipment on a fingerprint acquisition instrument, but the cost of the acquisition instrument is increased, so that the method is called a hardware-based false fingerprint detection method; the second method is to perform additional processing on the fingerprint sample picture, such as gray level analysis, texture feature analysis, etc., in order to detect activity information of the fingerprint image, and is called a software-based method. Software-based methods include static and dynamic feature detection methods, where static features are extracted from one or more images (a finger is placed on the acquirer one or more times for acquisition), and dynamic features are extracted from multiple fingerprint image frames (a finger is placed on the acquirer for a period of time to acquire a sequence of images for analysis). The software-based approach is low cost, less intrusive to the user, and can be used with existing fingerprint collectors. Therefore, the method has great practical value and popularization significance for the research of the software-based false fingerprint detection method.

Moon provides a false fingerprint detection method for analyzing the roughness degree of the surface of a high-resolution fingertip image. The method is only effective for high-resolution acquisition instruments (1000 dpi and above), and has low applicability to common 500dpi acquisition instruments popular in the market. Due to the existence of sweat pores and the sweat phenomenon, the gray level of the ridge line of the real fingerprint is obviously changed, while the gray level of the ridge line of the artificial fingerprint made of artificial materials such as gelatin and silica gel is slowly changed. Based on the above idea, Nikam and Agrwal propose a texture-based method for analyzing the liveness of a fingerprint image using the gray levels associated with the fingerprint pixels. This method shows good performance when the center point (core point) is accurately located, whereas existing center point detection algorithms do not perform well when dealing with low quality images and too dry or too wet fingerprint images. Abhyankar and Schuckers propose a method based on multi-resolution texture analysis and texture frequency analysis to quantify the gray level distribution changes as the physical structure changes using different texture characteristics. However, this method has a limitation in practical application because the calculation of the local ridge frequency is influenced by weather and is also related to different skin conditions.

At present, texture analysis is a mainstream method for false fingerprint detection, filtering and denoising are generally performed on a fingerprint image before texture analysis, and common texture features include first-order statistics, Local Binary Pattern (LBP) and the like. However, the LBP operator has a limitation of a small spatial support area, and it is not preferable to perform a pure texture analysis instead of discarding the analysis of the noise difference of the acquired image due to the difference of materials between true and false fingerprints.

Disclosure of Invention

The invention aims to solve the problems and provides a false fingerprint detection algorithm based on curved fingerprint analysis and SVM-KNN classification, the algorithm is suitable for various fingerprint collection instruments with the resolution of 500dpi, high-frequency noise information of images and denoised texture information are respectively analyzed, noise and texture differences between true and false fingerprints are quantified, and compared with an LBP algorithm, the algorithm can better express the tiny deviation between noise distribution of the true fingerprints and the false fingerprints caused by false fingerprint manufacturing materials.

The technical conception of the invention is as follows: currently, the mainstream method for false fingerprint detection is texture analysis, for example, in the text "a comprehensive description of texture features with classification based on fed distribution" of t.ojala, correlation analysis is performed on false fingerprint detection using Local Binary Pattern (LBP) features, LBP is an effective texture description operator, measures and extracts Local texture information of an image, the illumination is not changed, the textures of the true and false fingerprints have no obvious difference, the expression forms of the textures have slight difference, the fingerprint images are quantitatively analyzed by extracting LBP characteristics to distinguish the true and false fingerprints, the test results of this algorithm showed an average false acceptance rate of 14.85%, but the LBP operator had the limitation of a small spatial support region, and the method carries out simple texture analysis, and abandons the analysis of the noise difference of the collected images caused by different materials of the true and false fingerprints, which is also not advisable; because the material of the false fingerprint is different from the skin of a human body, the noise distribution of the fingerprint image acquired usually has larger difference, the invention provides the false fingerprint detection algorithm for carrying out SVM-KNN classification by utilizing the curvelet coefficient characteristic and the curvelet reconstruction image texture characteristic. Firstly, performing curvelet transformation on a fingerprint image to extract coefficient characteristics of each scale and each direction domain, wherein curvelet coefficients can be divided into a plurality of scale layers, the innermost layer, namely the first layer, is called a Coarse scale layer and is a matrix composed of low-frequency coefficients, the outermost layer is called a Fine scale layer and is a matrix composed of high-frequency coefficients, the middle layers are called Detail scale layers, each layer of coefficients is a matrix composed of medium-high frequency coefficients, and the high-frequency and medium-high frequency coefficients in the Fine layer and the Detail layer contain noise information of the image; reconstructing a fingerprint image and extracting texture features and coefficient features such as first-order statistics, a gray level co-occurrence matrix (GLCM) and a Markov Random Field (MRF) to form a feature vector; and then training by using an SVM (support vector machine), and introducing SVM-KNN (support vector machine-K nearest neighbor) classification to detect false fingerprints due to instability of the SVM in classifying samples near the hyperplane.

In order to achieve the purpose, the invention adopts the following technical scheme:

the false fingerprint detection algorithm based on the moire texture analysis and SVM-KNN classification comprises the following steps:

s1, curvelet transformation

Carrying out discrete curvelet transformation on the target image by utilizing a Wrap algorithm to obtain a curvelet system c (j, l, k);

s2, curvelet reconstruction

Denoising before curvelet transformation, and then reconstructing curvelet to obtain a reconstructed image f;

s3, coefficient feature extraction

Extracting coefficient characteristics of the curvelet coefficient c (j, l, k), which comprises the following processes:

s3.1, dividing the curvelet coefficient into a plurality of scale layers, wherein the innermost layer, namely the first layer is called a Coarse scale layer and is a matrix formed by low-frequency coefficients; the outermost layer, also called Fine scale layer, is a matrix composed of high frequency coefficients; the middle layers are called Detail scale layers, and the coefficient of each layer is a matrix consisting of medium and high frequency coefficients; the high-frequency and medium-high-frequency coefficients in the Fine layer and the Detail layer contain graphical noise information, coefficient characteristics are extracted from the Fine layer and the Detail layer, energy E, a maximum value MAX, a minimum value MIN, a Mean value Mean and a variance Var of each layer of system are counted, and the square sum of absolute values of the coefficients of each layer is calculated to serve as energy;

s3.2, performing first-order statistic feature extraction on each layer of series, wherein the first-order measurement comprises the following steps:

energy:

entropy:

variance:

skewness:

kurtosis:

s4, extracting texture features

Three types of texture features are extracted from the reconstructed image f, namely a first-order measurement matrix, a gray level co-occurrence matrix and an MRF (maximum likelihood ratio), and the method comprises the following steps:

s4.1, calculating the variation degree among pixels through a histogram, extracting first-order statistics, and quantifying the variation of gray level distribution when the physical structure of the image is changed to measure the probability of the occurrence of a certain gray level value at a random position in the image, wherein the correlation among the pixels can indicate the truth of a fingerprint;

s4.2, the pixel with the gray level x in the image is located at (i, k), the distance between the pixel and the pixel is D, the direction is theta, and the pixel with the gray level y is counted (i + D)_i,k+D_k) The mathematical expression of the number of occurrences p (x, y, d, θ) is:

p(x,y,d,θ)={[(i,k),(i+D_i,k+D_k)|f(i,k)=x,

f(i+D_i,k+D_k)=y]}

wherein x, y =1, 2, ·, L denotes a gray level in an image;

i, K =1, 2, ·, K denotes pixel coordinates;

d is the step length for generating the gray level co-occurrence matrix;

D_i，D_kis the position offset;

the generation direction theta can be any direction, so that the co-occurrence matrixes in different directions are generated, and the gray level co-occurrence matrix is normalized:

the extraction distance d is 1-4, theta is respectively taken in four directions of 0 degree, 45 degrees, 90 degrees and 135 degrees, and 16 gray level co-occurrence matrixes are counted; then respectively calculating corresponding energy, contrast, entropy, local stationary, autocorrelation and dissimilarity:

energy:

contrast ratio:

entropy:

local stability:

self-correlation:

dissimilarity:

wherein,

the texture features can effectively describe the texture features of the fingerprint image, and have better identification capability, and the gray level co-occurrence matrix of the image can reflect the comprehensive information of the gray level of the image about the change amplitude, direction and adjacent intervals, and is the basis for analyzing the local pattern and the arrangement rule of the image;

s4.3, from the structural analysis perspective, MRF textural features are to find out interdependencies between texture primitives and texture primitives, the primitive relationships of the MRF textural features can be represented by a conditional probability model, the MRF is a two-dimensional lattice, each point can be described by the probability model, the assumption of the MRF is that the pixel value of each point in the lattice only depends on the pixel value of pixels in the field, and the Markov random field can be described by the following local conditional Probability Density (PDF):

p(z(s)|z(m),m=1,2,…,N×M,s≠m)

=p(z(s)|z(m),m∈N(s))

where z(s) is the pixel value of point s in the lattice nxm, N(s) is a domain set of pixel points centered at s, the value of p (z (s)) is affected by z (M), if the PDF obeys gaussian distribution, the markov random field is gaussian-markov random field (GMRF), and the neighborhood information is used to estimate the symmetric difference equation for the gray value of the pixel point:

z(s)=∑β_s,m[z(s+m)+z(s-m)]+e_s

wherein beta is_s,mWeight, e, of gray value of center pixel contributing to each neighborhood pixel_sIs the gaussian noise with mean 0, m is the deviation from the center point s, and is expressed as a matrix notation formula:

z(s)=β^TQ_s+e_s

wherein beta is represented by_s,mVector of composition, Q_sIs defined as follows:

Q_{s} = [\begin{matrix} z (s + m_{1}) + z (s - m_{1}) \\ z (s + m_{2}) + z (s - m_{2}) \\ z (s + m_{3}) + z (s - m_{3}) \\ . . . \end{matrix}]

the texture characteristic quantity is calculated by utilizing a least square method,

<math> <mrow> <mi>β</mi> <mo>=</mo> <msup> <mrow> <mo>[</mo> <munder> <mi>Σ</mi> <mrow> <mi>s</mi> <mo>&Element;</mo> <mi>f</mi> </mrow> </munder> <msub> <mi>Q</mi> <mi>s</mi> </msub> <msubsup> <mi>Q</mi> <mi>s</mi> <mi>T</mi> </msubsup> <mo>]</mo> </mrow> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </msup> <mo>[</mo> <munder> <mi>Σ</mi> <mrow> <mi>s</mi> <mo>&Element;</mo> <mi>f</mi> </mrow> </munder> <msub> <mi>Q</mi> <mi>s</mi> </msub> <mi>z</mi> <mrow> <mo>(</mo> <mi>s</mi> <mo>)</mo> </mrow> <mo>]</mo> </mrow> </math>

using GMRF to carry out second-order parameter estimation on the fingerprint image, wherein any 3 x 3 window in the image is the image sampling template range, thereby determining Q_sF represents a fingerprint image, each 3 multiplied by 3 window beta is a characteristic value, the MRF characteristic value is insensitive to gray level change, the MRF reflects different aggregations of texture elements, textures among different aggregations correspond to different statistical parameters, and random characteristics of the textures can be well described by applying the mathematical MRF;

s5.SVM-KNN classifier formation

For a sample alpha to be identified, alpha and two types of support vector representative points alpha are calculated⁺And alpha^-When the distance difference is smaller than a given threshold value, namely the alpha is closer to a separation interface, the SVM only calculates the distance between the SVM and a representative point taken by the two classes to classify the samples easily, and classification results are obtained by classifying the test samples by KNN and calculating the distance between the samples to be identified and each support vector;

s6, classifier training

Estimating relevant parameters of a classification model by analyzing a training set with known classes, deducing the class of test data according to the relevant parameters, training by using LIBSVM (LiBSVM), obtaining a training model A by using a feature vector, and classifying and scoring results by using SVM (support vector machine) -KNN (K nearest neighbor) to respectively test samples through the training model；

S7, evaluating performance of classifier

Dividing a fingerprint image database into two parts, wherein 50% of fingerprint image database is used for training, 50% of fingerprint image database is used for testing, and after testing, the different epsilon values of the database have certain difference, wherein epsilon =0.2 in the invention, and an SVM-KNN classifier is used for training and classifying, so that the classification performance of a simple SVM classifier is improved to a certain extent;

s8. false fingerprint detection

The image to be detected is subjected to the operations of steps S1, S2 and S3, and the obtained feature vectors are classified by a classifier.

The invention has the following beneficial effects: the texture and the noise information of the image are analyzed in multiple aspects, an SVM classifier is used for fusing KNN to improve the performance of the single SVM classifier, and the reliability of the classification result is good.

Drawings

FIG. 1 is a flow chart of a false fingerprint image detection algorithm in a false fingerprint detection algorithm based on a curved texture analysis and SVM-KNN classification.

FIG. 2 is a SVM-KNN classification schematic diagram in a fake fingerprint detection algorithm based on the curved wave texture analysis and SVM-KNN classification.

Detailed Description

In order to make the technical means, the creation characteristics, the achievement purposes and the effects of the invention easy to understand, the invention is further explained below by combining the specific drawings.

False fingerprint detection algorithm based on the moire texture analysis and SVM-KNN classification, comprising the following steps (see FIG. 1):

s1, curvelet transformation: carrying out discrete curvelet transformation on the target image by utilizing a Wrap algorithm to obtain a curvelet system c (j, l, k);

where j denotes the scale, l denotes the angle, and k denotes the kth coefficient.

S2, curvelet reconstruction: the curvelet reconstruction process is a curvelet inverse transformation process, and before curvelet transformation, denoising is carried out firstly, and then curvelet reconstruction is carried out to obtain a reconstructed image f.

S3, coefficient feature extraction: extracting coefficient characteristics of the curvelet coefficient c (j, l, k), which comprises the following processes:

s3.1, dividing the curvelet coefficient into a plurality of scale layers, wherein the innermost layer, namely the first layer is called a Coarse scale layer and is a matrix formed by low-frequency coefficients; the outermost layer, also called Fine scale layer, is a matrix composed of high frequency coefficients; the middle layers are called Detail scale layers, and the coefficient of each layer is a matrix consisting of medium and high frequency coefficients; the high-frequency and medium-high-frequency coefficients in the Fine layer and the Detail layer contain graphical noise information, coefficient characteristics are extracted from the Fine layer and the Detail layer, energy E, a maximum value MAX, a minimum value MIN, a Mean value Mean and a variance Var of each layer of system are counted, and the square sum of absolute values of the coefficients of each layer is calculated to serve as the energy.

energy:

entropy:

variance:

skewness:

kurtosis:

s4, extracting texture features: three types of texture features are extracted from the reconstructed image f, namely a first-order measurement matrix, a gray level co-occurrence matrix and an MRF (maximum likelihood ratio), and the method comprises the following steps:

s4.1, calculating the variation degree among pixels through a histogram, extracting first-order statistics, and aiming at quantifying the variation of gray level distribution when the physical structure of the image is changed, wherein the first-order statistics can be applied to general one-dimensional data feature extraction and can also be applied to a two-dimensional data matrix of a similar picture to measure the probability of the occurrence of a certain gray value at a random position in the image, and the correlation among the pixels can indicate the truth of a fingerprint.

S4.2, counting a pixel (i + D) with a gray level of y and a distance of D, a direction of theta, wherein the position of the pixel is (i, k) and the gray level of the pixel is x in the image_i,k+D_k) The mathematical expression of the number of occurrences p (x, y, d, θ) is:

p(x,y,d,θ)={[(i,k),(i+D_i,k+D_k)|f(i,k)=x,

f(i+D_i,k+D_k)=y]}

wherein x, y =1, 2, ·, L denotes a gray level in an image;

i, K =1, 2, ·, K denotes pixel coordinates;

d is the step length for generating the gray level co-occurrence matrix;

D_i，D_kis the position offset;

energy:

contrast ratio:

entropy:

local stability:

self-correlation:

dissimilarity:

wherein,

p(z(s)|z(m),m=1,2,…,N×M,s≠m)

=p(z(s)|z(m),m∈N(s))

z(s)=∑β_s,m[z(s+m)+z(s-m)]+e_s

z(s)=β^TQ_s+e_s

Q_{s} = [\begin{matrix} z (s + m_{1}) + z (s - m_{1}) \\ z (s + m_{2}) + z (s - m_{2}) \\ z (s + m_{3}) + z (s - m_{3}) \\ . . . \end{matrix}]

using GMRF to carry out second-order parameter estimation on the fingerprint image, wherein any 3 x 3 window in the image is the image sampling template range, thereby determining Q_s(ii) a f represents a fingerprint image, each 3 multiplied by 3 window beta is a characteristic value, the MRF characteristic value is insensitive to gray level change, the MRF reflects different aggregations of texture elements, textures among different aggregations correspond to different statistical parameters, and random characteristics of the textures can be well described by applying the mathematical MRF.

S5, forming an SVM-KNN classifier: the traditional SVM classification algorithm has one defect: when the sample distance classification hyperplane is smaller than a given threshold epsilon, the classification accuracy is reduced. In the SVM classification process, the classification accuracy of the SVM is determined according to the error degree of the obtained representative points of each type of support vector, and samples which are easy to deviate are classified through KNN to improve the classification accuracy. Specifically, for a sample alpha to be identified, alpha and two types of support directions are calculatedQuantity representative Point α⁺And alpha^-When the distance difference is smaller than a given threshold value, namely the alpha is closer to the separation interface, the SVM only calculates the distance between the SVM and one representative point taken by the two classes to classify the samples, and classification is easy to be mistaken.

S6, classifier training: estimating relevant parameters of a classification model by analyzing a training set with known classes, deducing the class of test data according to the relevant parameters, training by using LIBSVM (LiBSVM), obtaining a training model A by using a feature vector, and classifying and scoring results by using SVM (support vector machine) -KNN (K nearest neighbor) to respectively test samples through the training model。

S7, evaluating the performance of the classifier: the fingerprint image database is divided into two parts, 50% of fingerprint image database is used for training, 50% of fingerprint image database is used for testing, and after testing, the different epsilon values of the database have certain difference, wherein epsilon =0.2, and an SVM-KNN classifier is used for training and classifying, so that the classification performance of a simple SVM classifier is improved to a certain extent.

Referring to fig. 2, the SVM-KNN classification principle means:

the traditional SVM classification algorithm has the following defects: when the sample distance classification hyperplane is smaller than a given threshold epsilon, the classification accuracy is reduced. In the SVM classification process, the classification accuracy of the SVM is determined according to the error degree of the obtained representative points of each type of support vector, and samples which are easy to deviate are classified through KNN to improve the classification accuracy. Specifically, for a sample alpha to be identified, alpha and two types of support vector representative points alpha are calculated⁺And alpha^-When the distance difference is smaller than a given threshold value, that is, α is closer to the separation interface and falls into a region III (as shown in fig. 2), the SVM only calculates the distance to a representative point of the two classes for classification, which is easy to be classified, and the KNN is used to classify the test sample, and the distance between the sample to be identified and each support vector is calculated to obtain the classification result.

S8. false fingerprint detection

The final table is the results of this test on the LivDet2011 game image library.

TABLE 1 false recognition rate of the algorithm for each false fingerprint data

TABLE 2 false reject rate of this algorithm to true fingerprint data

TABLE 3 false fingerprint average false acceptance rate for the algorithm herein

The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment, and all technical solutions belonging to the principle of the present invention belong to the protection scope of the present invention. Modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention.

Claims

1. The false fingerprint detection algorithm based on the cockle texture analysis and SVM-KNN classification is characterized by comprising the following steps:

s1, curvelet transformation

s2, curvelet reconstruction

s3, coefficient feature extraction

energy:

entropy:

variance:

skewness:

kurtosis:

s4, extracting texture features

p(x,y,d,θ)={[(i,k),(i+D_i,k+D_k)|f(i,k)=x,

f(i+D_i,k+D_k)=y]}

wherein x, y =1, 2, ·, L denotes a gray level in an image;

i, K =1, 2, ·, K denotes pixel coordinates;

d is the step length for generating the gray level co-occurrence matrix;

D_i，D_kis the position offset;

energy:

contrast ratio:

entropy:

local stability:

self-correlation:

dissimilarity:

wherein,

p(z(s)|z(m),m=1,2,…,N×M,s≠m)

=p(z(s)|z(m),m∈N(s))

z(s)=∑β_s,m[z(s+m)+z(s-m)]+e_s

z(s)=β^TQ_s+e_s

Q_{s} = [\begin{matrix} z (s + m_{1}) + z (s - m_{1}) \\ z (s + m_{2}) + z (s - m_{2}) \\ z (s + m_{3}) + z (s - m_{3}) \\ . . . \end{matrix}]

s5.SVM-KNN classifier formation

s6, classifier training

S7, evaluating performance of classifier

s8. false fingerprint detection