CN102779273A - Human-face identification method based on local contrast pattern - Google Patents

Abstract

The invention relates to a human-face identification method based on a local contrast pattern (LCP) and belongs to the technical field of pattern identification. The human-face identification method based on the LCP includes that human-face database images are preprocessed firstly so as to weaken influences of illumination on feature extraction; calculation of LCP feature matrixes is performed on training set images and testing set images respectively; the obtained LCP feature matrixes are converted into a column diagram; calculation of feature similarity of testing samples and training samples is performed on the column diagram by adopting a chi-square (X<2>) distance function; and the testing samples are classified and identified by using a nearest neighbor classifier. The human-face identification method based on the LCP has high human-face identification accuracy.

Description

A kind of face identification method based on local contrastive pattern

Technical field

The present invention relates to mode identification technology, particularly a kind of face identification method based on local contrastive pattern.

Background technology

In recent years, face recognition technology has been made significant headway, and various face recognition algorithms are suggested in succession and update, and has a lot of face identification systems to drop into actual use at present.But; Still have a lot of problems not solved well in the recognition of face research; Its reason is, facial image can receive such as illumination variation, expression shape change in acquisition process, the influence of factor such as block, and this interference that wherein brings with illumination variation again is the most serious.Wait out of doors under the environment that can not control illumination, face characteristic receives the influence of direction of illumination and intensity of illumination and produces nonlinearities change significantly, makes recognition of face become very difficult.Difference between the image that suitable people's face obtains under different illumination conditions is also bigger than the difference of different people face between the image that obtains under the identical illumination condition.

Eliminating illumination to aspect the influence of recognition of face, (Local Binary Pattern, LBP) method is just causing showing great attention to of people because of advantages such as easy, the anti-illumination interference of its calculating, discriminating power are strong to local binary pattern.

Local binary pattern (LBP) is proposed by Ojala, and this method is showed the textural characteristics of image local by the relation between the local pixel gray-scale value.Local binary pattern (LBP) operator definitions is at one 3 * 3 window; With the window center gray values of pixel points is threshold value, the gray-scale value of 8 adjacent neighbor pixel points is compared with it, if the surrounding pixel value is greater than center pixel value; Then this location of pixels is marked as 1, otherwise is 0.Like this, 8 points in 3 * 3 neighborhoods can produce one 8 scale-of-two unsigned number, give different weights by its position again, and its summation is promptly obtained local binary pattern (LBP) eigenwert of this window, and describe this regional texture information with this number.

Because local binary pattern (LBP) method has only been considered the magnitude relationship between central pixel point and the neighbor pixel point gray-scale value; It is different fully image local area pixel gray-scale value to occur; But the identical situation of local binary pattern (LBP) eigenwert that they obtain; Thereby influenced discrimination based on the face identification method of local binary pattern (LBP); This is because local binary pattern (LBP) method is not considered the comparative information of gray-scale value between central pixel point and the neighbor pixel point, and the difference of comparative information is distinguished a very important characteristic of regional area texture exactly.Therefore, improve the discrimination based on the face identification method of local binary pattern (LBP), the introducing of comparative information becomes a kind of desirable direction.

Summary of the invention

The present invention is based on the existing method that adopts local binary pattern (LBP) to carry out recognition of face and do not consider the comparative information of gray-scale value between central pixel point and the neighbor pixel point; Thereby have influence on this problem of precision of recognition of face; We have proposed a kind of based on the local contrastive pattern (face identification method of Local Contrast Pattern-LCP); This method can solve the problem that there is not comparative information between considered pixel in local binary pattern (LBP) method and influences discrimination effectively, obtains high recognition.

Technical scheme of the present invention is: a kind of face identification method based on local contrastive pattern (LCP), and this method step specifically comprises:

1. facial image pre-service

Sample in the face database of test is divided into training set and test set, and all samples are carried out pre-service, pre-service comprises gamma correction, difference of gaussian filtering and contrast equalization, to weaken the influence of illumination to feature extraction.Gamma correction is controlled the overall brightness of facial image through changing the Gamma parameter; Utilize the difference of gaussian wave filter to realize to the even smoothing processing of illumination unevenness of face image; The purpose of contrast equalization is that the gray level of whole facial image is regulated again, is a kind of standardization to integral image contrast and brightness variation.Through behind the above preprocessing process, can obtain illumination variation facial image relatively uniformly, thereby reduce the influence of illumination variation feature extraction.

2. extract local contrastive pattern (LCP) eigenface of facial image

To the impartial piece of drawing of pretreated facial image; Adopt local contrastive pattern (LCP) method to each pixel computation of characteristic values respectively to each piecemeal; Original gray value with eigenwert replacement respective pixel obtains the new facial image of a width of cloth, is called local contrastive pattern (LCP) eigenface.

3. local contrastive pattern (LCP) eigenface is carried out the histogram conversion

Convert local contrastive pattern (LCP) eigenface into histogram respectively according to the piecemeal of being divided, and all piecemeals are unified into an enhancing histogram by row priority, facial image is described as characteristic.

4. ask the card side (χ between the histogram ²) distance

To the resulting enhancing histogram of step 3, adopt χ ²Distance function calculates the enhancing histogram of test set facial image and the histogrammic χ of enhancing of all training set facial images ²Distance.

5. with nearest neighbor classifier the test set image is carried out Classification and Identification

With the χ between the histogram ²Distance compares, and therefrom selects minimum one type of distance, as the classification under the test set image.

The present invention is directed to the only size of neighbor pixel and center pixel gray-scale value relatively of local binary pattern (LBP) method; And do not consider the shortcoming of the concrete difference between its comparative information; It is improved to the mode of asking for eigenwert based on pixel contrast information; Designed a kind of face identification method based on local contrastive pattern (LCP), adopted nearest neighbor classifier to carry out Classification and Identification, this method has obtained higher discrimination.

Description of drawings

Fig. 1 is based on local contrastive pattern's (LCP) face identification method process flow diagram;

Fig. 2 is the block method of drawing of facial image;

Fig. 3 adopts local contrastive pattern (LCP) operator to ask for the process of eigenwert.

Embodiment

Below in conjunction with accompanying drawing and instance the present invention is described further.

As shown in Figure 1, a kind of face identification method based on local contrastive pattern may further comprise the steps:

1. facial image pre-service

Sample in the face database of test is divided into training set and test set, and all samples are carried out pre-service, this preprocess method comprises following three steps:

1) gamma correction

Gamma correction is the nonlinear transformation that original-gray image I is adopted index or log-transformation.As use I ^γ, perhaps log (I) replaces original-gray image I, wherein, and γ>0, γ ∈ [0,1] can weaken the influence that illumination variation is brought to a certain extent through index or log-transformation, and optimum desirable γ=0.2 is as default value.

2) difference of gaussian filtering DoG (Difference of Gaussian Filtering)

Gamma correction can not be eliminated the influence that global illumination brightness changes, the hatching effect that is for example caused by surface structure fully.The DoG wave filter can filter out some redundant informations under the situation mostly, improves the performance of total system.According to the gray image signals through gamma correction is done difference of gaussian filtering, the transport function of difference of gaussian wave filter is the poor of two different in width Gaussian functions, concrete formula:

$G\left(U\right)={\mathrm{Ae}}^{-u^2/{{\mathrm{\&sigma;}}_1}^2}-{\mathrm{Be}}^{-u^2/{{\mathrm{\&sigma;}}_2}^2},A\&GreaterEqual;B,{\mathrm{\&sigma;}}_1\&le;{\mathrm{\&sigma;}}_2---\left(1\right)$

Wherein, A, B are standardization coefficient, σ ₁, σ ₂Be respectively the width of two wave filters, optimum can be provided with σ ₀=1.0, σ ₁=2.0 is default parameters.Obtain new image I (x, y), wherein x, y are image pixel residing horizontal ordinate position in image do difference of gaussian filtering through the gray image signals u of gamma correction.

3) contrast equalization

After gamma correction and difference of gaussian filtering, still there are Gao Guang, blurred picture edge pixel in the image, the purpose of contrast equalization is that image is readjusted gray shade scale, makes the brightness of image and contrast reach standardization.The present invention adopts simply fast that method realizes the contrast equalization, specifically comprises following three formulas:

$I(x,y)\&LeftArrow;\frac{I(x,y)}{{\left(\mathrm{mean}\left({\left|I(x^{\&prime;},y^{\&prime;})\right|}^{\mathrm{\&alpha;}}\right)\right)}^{1/\mathrm{\&alpha;}}}---\left(2\right)$

$I(x,y)\&LeftArrow;\frac{I(x,y)}{{\left(\mathrm{mean}\left(\min (\mathrm{\&tau;},|I(x^{\&prime;},y^{\&prime;}){|}^{\mathrm{\&alpha;}}\right)\right)}^{1/\mathrm{\&alpha;}}}---\left(3\right)$

$I(x,y)\&LeftArrow;\mathrm{\&tau;}\tanh (I(x,y)/\mathrm{\&tau;})---\left(4\right)$

Wherein, (x y) is image after the last step difference of gaussian Filtering Processing to I, and x, y are image pixel residing horizontal ordinate position in image; Mean averages to the entire image gray-scale value; Min gets small function; τ is a cutoff frequency, be used for filtering existence than the high-gray level value; α is the high compression index, can reduce the influence of very big gray-scale value; Tanh is a hyperbolic tangent function, can be further with gray-scale value control in the reasonable scope.Optimum can be provided with α=0.1, τ=10 are default parameters.

Respectively the facial image sample is carried out gamma correction, difference of gaussian filtering and the operation of contrast equalization obtain illumination variation facial image relatively uniformly, are used for feature extraction.

2. extract local contrastive pattern (LCP) eigenface of facial image

Pretreated facial image is divided into impartial n piece (value of n is chosen according to experiment effect according to the size of facial image), draws block mode shown in accompanying drawing 2.Accompanying drawing 2 is a people's face in the PIE face database, and resolution is 64 * 64, its block size by 8 * 8 is carried out equalization divide.Adopt local contrastive pattern (LCP) method to each pixel computation of characteristic values respectively to each piecemeal.

Adopt the T operator to describe the local textural characteristics of facial image, the T operator representation is:

$T=(s(g_0/g_c),s(g_1/g_c),...,s(g_{p-1}/g_c))---\left(5\right)$

Wherein, g _cRepresent the gray-scale value of regional area central pixel point, g ₀To g _P-1Represent in the regional area with g _cIt for the center of circle, radius the gray-scale value of an equally distributed P neighbor pixel on the circular arc of R.Through to each s (g _p/ g _c) (ratio of neighbor pixel and center pixel gray-scale value), specify a binomial factor 2 ^p, calculate:

${\mathrm{LCP}}_{P,R}=\underset{p=0}{\overset{P-1}{\mathrm{\&Sigma;}}}\frac{g_p}{g_c}{2}^p\frac{1}{g_c},p=\mathrm{0,1,2}...P-1---\left(6\right)$

Wherein P is the neighbor pixel number, multiply by 1/g in the formula _cBe because work as the g of two different regional areas _p/ g _cWhen equating, the situation that eigenwert equates still can occur, add 1/g this moment _cThen can they be differentiated.Through formula (6), can obtain describing local contrastive pattern (LCP) eigenwert of image local textural characteristics, its computation process is shown in accompanying drawing 3.

Local contrastive pattern's characteristic has reflected the local grain characteristic of facial image, can obtain the overall textural characteristics of picture in its entirety according to formula (2).To a width of cloth resolution is the image (M, N are respectively the number of pixels of image delegation and row) of M * N, has M * N local contrastive pattern (LCP) eigenwert, and these eigenwerts all are 2 ^pIn the individual round values one.This M * N local contrastive pattern (LCP) eigenwert formed the matrix of an expression picture, forms local contrastive pattern (LCP) eigenface f _LBP(x, y).Eigenface f _LBP(x is the matrix of forming with the LCP eigenwert of representing picture y), can it be regarded as by the facial image of eigenwert as grey scale pixel value, and be used to Texture classification and recognition of face.

3. local binary pattern (LCP) eigenface is carried out the histogram conversion

Convert local binary pattern (LCP) eigenface into histogram respectively according to the piecemeal of being divided, the blocked histogram conversion regime is following:

$H=(H_0,H_1,...H_{m-1})---\left(7\right)$

Wherein, m representes m value in the histogram, H _iBe the pairing eigenwert quantity of each value:

$H_i={\mathrm{\&Sigma;}}_{x,y}I\{f_{\mathrm{LBP}}(x,y)=i\}i=0,...m-1---\left(8\right)$

Obtain H through adding up 1 number in the formula _iValue, with all H _iObtain the histogram that promptly obtains this eigenface.If f _LBP(x, the eigenwert in y) equates that with corresponding histogrammic value then I{A} is 1, otherwise is 0:

$I\left\{A\right\}=\left\{\begin{array}{cc}1& \mathrm{Aistrue}\\ 0& \mathrm{Aisfalse}\end{array}\right.---\left(9\right)$

Then all blocked histograms are unified into the enhancing histogram of a m * n dimension by row priority:

$H_{i,j}={\mathrm{\&Sigma;}}_{x,y}I\{f_{\mathrm{LBP}}(x,y)=j\}i=0,...m-1j=\mathrm{0,1}...,m-1---\left(10\right)$

Wherein, m is the histogram number, and n is people's face image block number.To strengthen histogram describes facial image as characteristic.

4. ask the χ between the histogram ²(card side) distance

Adopt χ ²(card side) distance function calculates the enhancing histogram of resulting test set facial image and the histogrammic χ of enhancing of all training set facial images ²Distance.χ ²The computing formula of distance is following:

${\mathrm{\&chi;}}^2(p,q)={\mathrm{\&Sigma;}}_i\frac{{(p_i-q_i)}^2}{p_i+q_i}---\left(11\right)$

Wherein, p, q indicate to be the pairing space enhancing of two the people's faces histogram H of comparison respectively _{I, j}, i is for dividing block identification.

5. with nearest neighbor classifier the test set image is carried out Classification and Identification

Nearest neighbor classifier is a kind of sorting technique of simple, general-purpose in the recognition of face, with the χ between all histograms that obtain ²(card side) distance compares, and therefrom selects minimum one type of distance, as the classification under the test set image.

One embodiment of the present of invention are following:

Adopt YALE B, PIE and OUTDOOR face database are as experimental data base.

YALE B face database comprises 10 people's facial image, and everyone has 64 width of cloth direct pictures of taking under the different light.The people face position of all images is extracted the new face database of formation, and whenever magnifying little is 100 * 100.

The PIE face database comprises 68 people; Everyone comprises different attitudes, expression and illumination subclass; Totally 41368 photos adopt illumination subclass (C27) wherein, and in this subclass everyone comprises 21 photos under the different light; The people face position of all images is extracted the new face database of formation, and whenever magnifying little is 100 * 100.

The OUTDOOR face database comprises 68 people, and under uncontrollable natural lighting environment, everyone comprises 5 different illumination images, and the people face position with all images extracts the new face database of formation equally, and whenever magnifying little is 100 * 100

Experiment adopts four training sets of Set1 ~ Set4 to carry out, and the training set picture number of YALE B experiment is respectively everyone 15,25,35 and 45 images, and the test set picture number is 15.The training set picture number of PIE experiment is respectively everyone 6,10,15 and 20 images, and remaining image is the test set image.The training set picture number of OUTDOOR experiment is respectively everyone 1,2,3 and 4 images, and remaining image is the test set image.Adopt method of the present invention, above data set is tested.

Concrete test result is shown in table 1, table 2 and table 3, and wherein, LBP representation feature extraction step adopts local binary pattern (LBP) method, and the LCP representation feature extracts local contrastive pattern (LCP) method that adopts.Table 1 has provided the discrimination contrast of two kinds of methods on YALE B face database, and table 1 shows, on the YALE B face database on the discrimination and average recognition rate of Set1, Set2 and three branches experiments of Set4, the LCP method is superior to the LBP method; Table 2 has provided the discrimination contrast of two kinds of methods on the PIE face database, and table 2 shows, on the PIE face database on the discrimination and average recognition rate of Set1, Set2 and three branches experiments of Set3, the LCP method is superior to the LBP method; Table 3 has provided the discrimination contrast of two kinds of methods on the OUTDOOR face database, and table 3 shows, on the OUTDOOR face database on discrimination and average recognition rate that two branches of Set1 and Set2 are tested, the LCP method is superior to the LBP method.

The discrimination contrast of two kinds of methods of table 1 on YALE B face database

The discrimination contrast of two kinds of methods of table 2 on the PIE face database

The discrimination contrast of two kinds of methods of table 3 on the OUTDOOR face database

Claims (5)

1. the face identification method based on local contrastive pattern is characterized in that, comprises step: the sample in the face database is divided into training set and test set, adopts gamma correction, difference of gaussian filtering and contrast equalization that the facial image sample is carried out pre-service; To the impartial piece of drawing of pretreated facial image, adopt local contrastive pattern to each pixel computation of characteristic values respectively to each piecemeal, the original gray value with eigenwert replacement respective pixel obtains local contrastive pattern LCP eigenface; Convert local contrastive pattern's eigenface into histogram respectively according to the piecemeal of being divided, and all piecemeals are unified into the enhancing histogram by row priority, adopt card side χ ²Distance function calculates the enhancing histogram of test set facial image and the histogrammic χ of enhancing of all training set facial images ²Distance; With the χ between the histogram ²Distance compares, and selects minimum one type of distance, as the test set image.

2. face identification method according to claim 1 is characterized in that, gamma correction is the nonlinear transformation that original-gray image I is adopted index or log-transformation.

3. face identification method according to claim 1 is characterized in that, calls the transport function of difference of gaussian wave filter: $H_{i,j}=\underset{(x,y)\&Element;R_j}{\mathrm{\&Sigma;}}I\{f_{\mathrm{LBP}}(x,y)=i\}i=\mathrm{0,1},...,n-1j=\mathrm{0,1},...,m-1$ , the gray image signals through gamma correction is done difference of gaussian filtering, filter out redundant information, wherein, A, B are standardization coefficient, σ ₁, σ ₂Be respectively the width of two wave filters, u is a gray image signals.

4. face identification method according to claim 1 is characterized in that, adopts local contrastive pattern's method that each pixel computation of characteristic values is specially respectively to each piecemeal, adopts the T operator: T=(s (g ₀/ g _c), s (g ₁/ g _c) ..., s (g _P-1/ g _c)) the local textural characteristics of description facial image; Through ratio s (g to each neighbor pixel and center pixel gray-scale value _p/ g _c) binomial factor 2 of appointment ^p, calculate , obtain describing the local contrastive pattern LCP eigenwert of image local textural characteristics, wherein, g _cRepresent the gray-scale value of regional area central pixel point, g ₀To g _P-1Represent in the regional area with g _cBe the gray-scale value of an equally distributed P neighbor pixel on the circular arc of R for the center of circle, radius, P is the neighbor pixel number.

5. face identification method according to claim 1; It is characterized in that, be cascaded into the enhancing histogram and be specially, convert local binary pattern LCP eigenface into histogram respectively according to the piecemeal of being divided; With all blocked histograms, according to local contrastive pattern LCP eigenface function f _LBP(x, y) call formula: $H_{i,j}=\underset{(x,y)\&Element;R_j}{\mathrm{\&Sigma;}}I\{f_{\mathrm{LBP}}(x,y)=i\}i=\mathrm{0,1},...,n-1j=\mathrm{0,1},...,m-1$ Be unified into the enhancing histogram that a m * n ties up by row priority, m is the histogram number, and n is people's face image block number.

Images