CN102156887A - Human face recognition method based on local feature learning - Google Patents

Abstract

The invention discloses a human face recognition method based on local feature learning, which comprises the following steps of: (a) partitioning a known classified human face sample into blocks, and calculating data of the human face sample of each block through an LBP (Length Between Perpendiculars) operator and an LTP (Long-Term Potentiation) operator to obtain a local histogram vector a of data of the human face sample of each block; (b) performing chi-square histogram distance calculation on local histogram vectors of the same positions on any two different human face samples of the same person in all human face samples to obtain a positive sample feature library; and (c) performing chi-square histogram distance calculation on local histogram vectors of the same positions on any two human face samples of different persons in all human face samples to obtain a negative sample feature library. The human face recognition method based on the local feature learning, provided by the invention, has the advantages of quick response, high accuracy and good recognition effect.

Description

A kind of face identification method based on local feature study

Technical field

The present invention relates to a kind of biological identification technology, particularly a kind of face identification method based on local feature study.

Background technology

Biometrics identification technology is meant the technology of utilizing intrinsic physiological characteristic of biosome (mainly referring to the people) or behavioural characteristic to carry out identity authentication.With traditional identity authentication technology such as certificate, magnetic card, passwords etc. are compared, biometrics identification technology has made full use of individual's intrinsic biological characteristic, stopped the forgery of identity from the source and stolen, effective more, reliable, safety have obtained application more and more widely in information safety system.

In biometrics identification technology, compare iris recognition, fingerprint recognition etc., recognition of face has naturality, substantivity, friendly and advantage such as untouchable, be subjected to people's favor degree height, wide application prospect arranged at aspects such as national security, public security, criminal investigation, judicial domain, information security and Self-Services.

Traditional face identification method based on local feature study all can obtain certain recognition effect, but the recognition of face effect big to the illumination luminance difference, that expression shape change is apparent in view is not ideal enough.

Summary of the invention

To the objective of the invention is in order addressing the above problem, to have designed a kind of face identification method based on local feature study.

Realize that above-mentioned purpose technical scheme of the present invention is, a kind of face identification method based on local feature study, this method comprises sets up training pattern and two steps of model of cognition.

Wherein the foundation based on training pattern in the face identification method of local feature study may further comprise the steps:

(a) known classification is good people's face sample carries out the piecemeal processing, and by LBP operator (local binary pattern) and LTP operator (local three binarization modes) people's face sample data of every is calculated, and draws local histogram's vector of every people's face sample data;

(b) to the local histogram's vector side of card histogram distance calculation of the same position on any two the different people's face samples of same individual in everyone the face sample, obtain positive sample characteristics storehouse;

(c) to the local histogram's vector side of card histogram distance calculation of the same position on any two people's face samples of different people in everyone the face sample, obtain the negative sample feature database;

(d) will carry out computing in the input of the data in the positive and negative samples feature database AdaBoost cascade classifier, obtaining any two the people's face samples of differentiation is others' AdaBoost cascade classifier database of same class.

Foundation based on the model of cognition of the face identification method of local feature study may further comprise the steps;

(e) operation that facial image to be identified is carried out step (a) obtains local histogram's vector of every people's face sample data;

(f) local histogram's vector of every people's face sample data on the facial image to be identified and local histogram vector the block side histogram distance calculation of everyone face sample on same position in the AdaBoost cascade classifier database are obtained characteristic;

(g) will carry out computing in the described characteristic input AdaBoost cascade classifier and compare judgement with data in the AdaBoost cascade classifier database, the people's face classification under people's face sample that judgment data comes to the same thing be exactly the classification that people's face figure to be identified belongs to.

People's face sample data described in the more excellent step (a)-(e) is meant that the size of people's face sample and gray scale are carried out data normalization handles resulting data value.

The present invention compared with prior art has following beneficial effect: in extracting the face characteristic process, the face characteristic that utilizes LBP operator and LTP operator that the method for multiple local feature is extracted is more accurate, illumination, attitude variation are had good robustness, and discrimination is higher; The Adaboost cascade classifier is incorporated in the many classification problems of people's face, thereby utilization one-to-many strategy is converted into two class problems to the multiclass problem, has improved travelling speed, makes the recognition of face performance obtain significant raising.

Description of drawings

Fig. 1 is the schematic flow sheet based on the face identification method of local feature study that the present invention proposes;

Fig. 2 is basic LBP operator transformation synoptic diagram;

Fig. 3 is basic LTP operator transformation synoptic diagram;

Embodiment

For ease of the understanding of technical solution of the present invention, be introduced below in conjunction with concrete embodiment.As shown in Figure 1, a kind of face identification method based on local feature study, this method comprises sets up training pattern and two steps of model of cognition, wherein setting up training pattern may further comprise the steps: (a) that known classification is good people's face sample carries out the piecemeal processing, and by LBP operator and LTP operator people's face sample data of every is calculated, draw local histogram's vector of every people's face sample data; (b) to the local histogram's vector side of card histogram distance calculation of the same position on any two the different people's face samples of same individual in everyone the face sample, obtain positive sample characteristics storehouse; (c) to the local histogram's vector side of card histogram distance calculation of the same position on any two people's face samples of different people in everyone the face sample, obtain the negative sample feature database; (d) will carry out computing in the input of the data in the positive and negative samples feature database AdaBoost cascade classifier, obtaining any two the people's face samples of differentiation is others' AdaBoost cascade classifier database of same class.This method is set up model of cognition and be may further comprise the steps; (e) operation that facial image to be identified is carried out step (a) obtains local histogram's vector of every people's face sample data; (f) local histogram's vector of every people's face sample data on the facial image to be identified and local histogram vector the block side histogram distance calculation of everyone face sample on same position in the AdaBoost cascade classifier database are obtained characteristic; (g) will carry out computing in the described characteristic input AdaBoost cascade classifier and compare judgement with data in the AdaBoost cascade classifier database, the people's face classification under people's face sample that judgment data comes to the same thing be exactly the classification that people's face figure to be identified belongs to.

Be meant that in the people's face sample data described in the step (a)-(e) size of people's face sample and gray scale are carried out data normalization handles resulting data value, wherein the size of people's face sample is according to the center cutting of eyes, thereby and is meant the gray level image that the influence that utilized histogram equalization to operate to eliminate illumination obtains for the gray scale of people's face sample.Described histogram equalization operation is gray level γ in the hypothesis piece image _mThe probability that occurs is: p _γ(γ _m)=n _i/ n, m=0,1,2 ..., L-1, n are the summation of pixel in the image, n _mBe that gray level is γ _mNumber of pixels, L is the gray level sum in the image.Histogram equalization form with this understanding is:

$s_m=T\left({\mathrm{\γ}}_m\right)=\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{p}_{\mathrm{\γ}}\left({\mathrm{\γ}}_i\right)=\underset{i=0}{\overset{m}{\mathrm{\Σ}}}{n}_i/n,$ k＝0，1，2，...，L-1；

Described LBP operator is local binary pattern operator, and the LTP operator is meant local three binarization mode operators.All characteristics that calculate by described LBP operator and LTP operator all are to represent with LBP operator and the corresponding decimal data of LTP operator, then by operation program AdaBoost cascade classifier, LBP operator and LTP operator characteristic of correspondence data are classified, obtain to distinguish people's face of same individual and be not the characteristic set of people's face of same individual.Simultaneously, the AdaBoost cascade classifier is converted into two person-to-person people's face classification problems with people's face classification problem of a plurality of people.The specific implementation method of described AdaBoost cascade classifier classification is:

1. given every layer maximum negative sample false drop rate f and minimum positive pattern detection rate d, the negative sample false drop rate target F that cascade classifier will reach _TarWith positive pattern detection rate D _Tar, P, N are respectively positive and negative sample set;

2. set F ₀=1.0, D ₀=1.0, i=0;

3. work as F _i＞F _TarThe time, i=i+1, n _i=0, F _i=F _I-1Work as F _i＞f * F _I-1The time, n _i=n _i+ 1.

4. in computing collection P and N, have the strong classifier of n feature with the computing of Adaboost algorithm; Calculate the F of current cascade classifier _iAnd D _iAdjust the threshold value of current strong classifier, up to the verification and measurement ratio of current cascade classifier more than or equal to d * D _I-1, N is nonempty set;

5. if F _i＞F _Tar, the negative sample image that the current cascade classifier that obtains is sorted in other is judged, there not being judicious image to put into N.

Wherein the specific implementation of the Adaboost algorithm in the 4th step is:

1) given n computing sample (x ₁, y ₁) ..., (x _n, y _n), y _i=0,1 represents x respectively _iBe negative sample or positive sample.

2) initializes weights

Wherein positive number of samples is l, and the negative sample number is m.

3) t is from 1 to T, and following steps are carried out in circulation:

A) normalized weight ${\mathrm{\ω}}_{t,i}={\mathrm{\ω}}_{t,i}/{\mathrm{\Σ}}_{j=1}^n{\mathrm{\ω}}_{t,j};$

B) to each feature f _j, Weak Classifier h of computing _j, remember that this sorter error rate is

C) from all Weak Classifiers that previous step calculates, find out and have lowest error rate ω _tSorter h _t

D) upgrade weight

β wherein _t=ε _t/ (1-ε _t), if x _iBe classified correctly e _i=0, otherwise e _i=1.

4) obtain strong classifier at last: if

H (x)=1 then; Otherwise h (x)=0.Wherein

Carry out of the inventive method on famous FERET face database test, compare with the face identification method of classics, the discrimination of this method improves a lot.As shown in table 1, the inventive method with based on the method for principal component analysis, carried out the test and appraisal comparison based on the method for Bayes with based on the method for elastic bunch graph coupling.Wherein Fb is the expression shape change test set, and Fc is the illumination variation test set, and Dup I and Dup II change test set the time, and wherein Dup II is that the time difference in change is that year changes test set.Draw from table, the inventive method obviously is better than additive method, and for example in expression shape change test set Fc, the inventive method can reach 0.947 on discrimination, is much better than method discrimination 0.658 and all the other two kinds of methods based on principal component analysis (PCA).On computational complexity, the present invention has introduced cascade Adaboost method, and its computational complexity also is starkly lower than traditional method.Therefore, the inventive method is compared with existing face identification method, has not only improved the robustness of recognition of face, and has reduced the computational complexity of recognition of face, makes the recognition of face effect obtain significant raising.

Table 1

Technique scheme has only embodied the optimal technical scheme of technical solution of the present invention, those skilled in the art to some part wherein some changes that may make all embodied principle of the present invention, belong within the protection domain of invention.

Claims (3)

1. the face identification method based on local feature study comprises and sets up training pattern and two steps of model of cognition, and the foundation of this method training pattern may further comprise the steps:

(a) known classification is good people's face sample carries out the piecemeal processing, and by LBP operator and LTP operator people's face sample data of every is calculated, and draws local histogram's vector of every people's face sample data;

(b) to the local histogram's vector side of card histogram distance calculation of the same position on any two the different people's face samples of same individual in everyone the face sample, obtain positive sample characteristics storehouse;

(c) to the local histogram's vector side of card histogram distance calculation of the same position on any two people's face samples of different people in everyone the face sample, obtain the negative sample feature database;

(d) will carry out computing in the input of the data in the positive and negative samples feature database AdaBoost cascade classifier, obtaining any two the people's face samples of differentiation is others' AdaBoost cascade classifier database of same class.

2. the face identification method based on local feature study according to claim 1, the foundation of this method model of cognition may further comprise the steps;

(e) operation that facial image to be identified is carried out step (a) obtains local histogram's vector of every people's face sample data;

(f) local histogram's vector of every people's face sample data on the facial image to be identified and local histogram vector the block side histogram distance calculation of everyone face sample on same position in the AdaBoost cascade classifier database are obtained characteristic;

(g) will carry out computing in the described characteristic input AdaBoost cascade classifier and compare judgement with data in the AdaBoost cascade classifier database, the people's face classification under people's face sample that judgment data comes to the same thing be exactly the classification that people's face figure to be identified belongs to.

3. the face identification method based on local feature study according to claim 1 is characterized in that, the people's face sample data described in the step (a)-(g) is meant that the size of people's face sample and gray scale are carried out data normalization handles resulting data value.

Images