CN103778412A - Face recognition method based on local ternary pattern adaptive threshold - Google Patents

Abstract

The invention discloses a face recognition method based on a local ternary pattern adaptive threshold, and pertains to the technical field of pattern recognition. The method comprises the following steps: at first, segmentation treatment is performed on face database images; epsilon-LTP characteristics in segmented regions are calculated for training set and test set images separately; the obtained epsilon-LTP characteristic matrix is decomposed into two layers including positive and negative pattern layers; information entropy weight is solved and obtained according to the characteristics of each layer; the characteristics of each layer are converted into the form of a weighted histogram by combining with the information entropy weight; the chi<2> (chi-square) distance function is adopted in the histogram to calculate the characteristic similarity between test samples and training samples; and classification recognition is performed on the test samples through a third-order neighbor classifier. The method of the invention has the advantage of good face recognition accuracy.

Description

A kind of face identification method based on local three binarization mode adaptive thresholds

Technical field

The present invention relates to relate to mode identification technology, particularly a kind of face identification method based on local three binarization modes (LTP) adaptive threshold.

Background technology

In recent years, recognition of face has obtained very much progress in the research in the field such as pattern-recognition, image processing and analysis, and existing part face identification system drops into actual use at present.Various face identification methods are exactly to find a kind of effectively face describing mode from essence, but will arrive, a kind of can not to be subject to the describing mode that various factors affects be very difficult.Illumination variation, expression shape change, the factor such as block all can be on the impact to some extent of obtaining of facial image, and the interference wherein bringing with illumination variation is the most serious.In actual conditions, affected by the uncontrollable factors such as illumination, can there are some nonlinearities change in face, even can cause the difference between the diversity ratio different people of same people under different light also large, and this has greatly increased the difficulty of recognition of face work.

Local binary patterns (Local Binary Pattern, LBP) is proposed by Ojala, measures pixel value size and extract texture information in image local neighborhood, and illumination variation is had to robustness.Easy, the anti-illumination interference of its calculating, discriminating power are strong, are widely used in the recognition of face under illumination variation.But in the time of illumination acute variation, LBP cannot represent the severe degree changing, and therefore reliability declines to a great extent, and the people such as Tan has proposed again local three binarization modes (Local Ternary Pattern, LTP) on this basis.

LTP operator improves LBP operator, adopts three value codings, to improve the classification capacity of whole feature space.At the window of 3 × 3, self-defined threshold value t, compares pixel in neighborhood and center pixel, and pixel value difference is mapped in g _cin the region that be quantified as 0, width is [t ,+t], it is+1 that difference is greater than this Interval Coding, and it is-1 that difference is less than this Interval Coding, and difference is encoded to 0 in interval range.Like this, in neighborhood, can produce the scale-of-two signed number of 8, then give different weights by its position, and its summation be obtained to part three binarization modes (LTP) eigenwert of this window, describe the texture information in this region by this number.

By the research to LBP and improvement, LTP has solved the identification problem under illumination acute variation, and the image-forming condition (as noise etc.) to acute variation has robustness.But LTP self adopts self-defined threshold value, need look for according to priori, set optimal threshold, ageing meeting is influenced, and meanwhile, threshold value cannot be taken into account the difference between sample, also has Problem of Universality.Therefore, improve the discrimination of LTP operator to recognition of face, the optimization of threshold value becomes a desirable direction.

Summary of the invention

For above deficiency of the prior art, the object of the present invention is to provide a kind of face identification method based on local three binarization mode adaptive thresholds that effectively solves threshold optimization, further improves the precision to recognition of face, technical scheme of the present invention is as follows: a kind of face identification method based on local three binarization mode adaptive thresholds, comprises the following steps:

101,, using facial image to be measured as test set, the known face database of selected part, as training set, is divided into n piece by the each facial image in training set and test set;

102, on the each piecemeal of test set after dividing equally in step 101, to each neighborhood (P, R), R represents the radius of neighbourhood, and P represents neighborhood point number, respectively the pixel g of computing center _cwith corresponding neighborhood point g _i(i=0,1 ..., P-1) pixel comparison value, try to achieve the dispersion σ of this group correlative value, by this dispersion σ as adaptive threshold ε=σ;

103, upper at threshold interval [ε ,+ε], to each neighborhood point g _i(i=0,1 ..., P-1) and central pixel point g _cgray scale difference carry out LTP coding, solve LTP adaptive threshold eigenwert, and be holotype characteristic layer and negative mode characteristic layer by this LTP adaptive threshold Eigenvalues Decomposition, the facial image of holotype characteristic layer composition is holotype characteristic layer eigenface, and the facial image of negative mode characteristic layer composition is negative mode characteristic layer eigenface;

104, the holotype characteristic layer eigenface in difference calculation procedure 103 and the information entropy weights W of negative mode characteristic layer eigenface _j,

$W_j=E_j/\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}{E}_j(j=1,...,s),$

Wherein information entropy E _jcalculating formula is:

distribution probability

$P_j^i=\frac{\mathrm{\&Sigma;T}\{f(x,y)=i\}}{\underset{i=0}{\overset{n-1}{\mathrm{\&Sigma;}}}\mathrm{\&Sigma;T}\{f(x,y)=i\}}(i=\mathrm{0,1},...,n-1);$ F (x, y) represent distribution function

105, the holotype characteristic layer eigenface in step 103 and negative mode characteristic layer eigenface are weighted to histogram conversion, form the enhancing histogram of test set; Repeating step 102-105, the enhancing histogram of composing training collection

106, adopt card side χ ²distance function calculates the enhancing histogram of facial image to be measured and the histogrammic χ of enhancing of training set facial image ²distance, the three rank Nearest Neighbor Classifier methods of employing select the classification under facial image to be measured, complete recognition of face.

Further, in step 102, the dispersion σ acquiring method of each neighborhood correlative value is:

A, ask for central pixel point g according to formula (1) _ccontrast value in neighborhood (P, R),

△g _i=g _i-g _c,(i=0,1,...,P-1) (1)

B, ask for central pixel point g according to formula (2) _cthe average of the interior contrast value of neighborhood (P, R):

$\stackrel{\&OverBar;}{\mathrm{\&Delta;g}}=\left(\underset{i=0}{\overset{P-1}{\mathrm{\&Sigma;}}}{\mathrm{\&Delta;g}}_i\right)/P---\left(2\right)$

C, calculate the variance of this group contrast value according to formula (3):

$D=\left(\underset{i=0}{\overset{P-1}{\mathrm{\&Sigma;}}}{({\mathrm{\&Delta;g}}_i-\stackrel{\&OverBar;}{\mathrm{\&Delta;g}})}^2\right)/P---\left(3\right)$

D, according to formula (3) variance estimation dispersion:

$\mathrm{\&sigma;}=\sqrt{D}---\left(4\right)$

Further, the face database described in step 101 comprises Extended YALE B and PIE face database.

Further, the χ in step 106 ²distance function is: wherein, indicate the respectively test set and the corresponding spatial enhance histogram of the training set face H that contrast of S, M _i, i is sample identification.

Advantage of the present invention and beneficial effect are as follows:

The present invention, mainly for the limitation of the self-defined threshold value of local three binarization modes (LTP) method, has designed a kind of face identification method based on LTP adaptive threshold (ε-LTP).Because the threshold value defining by priori exist optimize and Problem of Universality, here threshold value is improved to by based in neighborhood between pixel the dispersion of correlative value define.Adaptive threshold feature extraction algorithm based on LTP, has solved searching optimal threshold and threshold value Problem of Universality, and combining information entropy is weighted fusion to the characteristic layer histogram decomposing simultaneously, and this method has obtained higher accuracy of identification.

Accompanying drawing explanation

Fig. 1 is the face identification method process flow diagram that the present invention is based on LTP adaptive threshold (ε-LTP) and layering Weighted Fusion;

Fig. 2 is the cataloged procedure that adopts LTP adaptive threshold (ε-LTP) operator;

Fig. 3 decomposes LTP adaptive threshold (ε-LTP) for the two-layer eigenface of positive and negative pattern;

Fig. 4 asks for each layer of feature weight of LTP adaptive threshold (ε-LTP) based on information entropy;

Fig. 5 is that on Extended Yale B database, subset 1 is made training sample (%);

On Fig. 6 Extended Yale B database, subset 4 is made training sample (%);

Fig. 7 is the discrimination (%) of distinct methods on CMU PIE database.

Embodiment

The invention will be further elaborated to provide the embodiment of an indefiniteness below in conjunction with accompanying drawing.

As shown in Figure 1, a kind of face identification method based on LTP adaptive threshold comprises the following steps:

1. extract multilayer LTP adaptive threshold (ε-LTP) eigenface of facial image

Sample in the face database of test is divided into training set and test set, and to all sample extraction LTP adaptive thresholds (ε-LTP) eigenface, this extracting method step is as follows:

1) to facial image piecemeal

For taking into account the local feature information of 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), draw block mode as shown in accompanying drawing 1 the first from left figure.This accompanying drawing is a face in Extended YALE B face database, and resolution is 240 × 220, and the block size by 8 × 8 carries out equalization and divides.

2) ask the adaptive threshold in each neighborhood of pixel points

The P neighborhood sampling of each pixel, gets 3 × 3 neighborhood pieces conventionally, removes P=8 neighbor pixel outside center pixel, radius of neighbourhood R=1.In neighborhood, the contrast of statistics central pixel point and interior the each point of neighborhood (P, R), according to the statistical property of discrete values, dispersion σ that must the interior contrast value of this neighborhood, specifically calculation procedure is as follows.

A) g _cin neighborhood (P, R), ask each contrast value:

△g _i=g _i-g _c,(i=0,1,...,P-1) (1)

B) ask for the average of this group contrast value:

$\stackrel{\&OverBar;}{\mathrm{\&Delta;g}}=\left(\underset{i=0}{\overset{P-1}{\mathrm{\&Sigma;}}}{\mathrm{\&Delta;g}}_i\right)/P---\left(2\right)$

C) calculate the variance of this group contrast value:

$D=\left(\underset{i=0}{\overset{P-1}{\mathrm{\&Sigma;}}}{({\mathrm{\&Delta;g}}_i-\stackrel{\&OverBar;}{\mathrm{\&Delta;g}})}^2\right)/P---\left(3\right)$

D) according to variance estimation dispersion:

$\mathrm{\&sigma;}=\sqrt{D}---\left(4\right)$

In computation process, dispersion sigma reaction is sensitive, and result can change with the variation of each value.Meanwhile, σ is subject to the impact of sampling variation less, and the dispersion of each sample is more stable in the ordinary course of things, and rate of change is not high.Based on above-mentioned characteristic, divide neighborhood to ask for dispersion to sample here, define the optimal threshold that this dispersion σ is LTP, obtain the threshold epsilon=σ of the each neighborhood of self-adaptation.

3) calculate LTP adaptive threshold (ε-LTP) eigenwert

Obtain after adaptive threshold ε, to neighborhood point g _i(i=0,1 ..., P-1) and central pixel point g _cgray scale difference value encode.According to the definition of LTP operator, the difference of neighborhood territory pixel is mapped in g _cin the interval that be quantified as 0, codomain is [ε ,+ε].It is+1 that pixel value difference is greater than this Interval Coding, and it is-1 that pixel value difference is less than this Interval Coding, and pixel value difference is encoded to 0 in interval range, and computing formula is as follows:

$s^{`}(g,g_c,t)=\left\{\begin{array}{c}1,(g-g_c)\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="t"\; filenames="HDA0000457795060000012.png,HDA0000457795060000013.png"\; state="\{\{state\}\}">t</figure-callout>}\\ 0,|g-g_c|<t\\ -1,(g-g_c)<-t\end{array}\right.---\left(5\right)$

After the interval conversion of contrast, be self-adaptation ε-LTP cataloged procedure of certain neighborhood as shown in Figure 2.

4) decompose LTP adaptive threshold (ε-LTP) feature

Be generally to simplify and calculate, while extracting LTP feature, can be decomposed into positive and negative pattern two-layer, adopt equally the method here.Obtain after ε-LTP coding, the coding result that extracts "+1 " is designated as " 1 ", and all the other are designated as " 0 ", obtain holotype feature by LBP coded system; The coding result that extracts " 1 " is designated as " 1 ", and all the other are designated as " 0 ", obtain negative mode feature by LBP coded system.Finally obtain the eigenface of two-layer different texture, coding decomposable process as shown in Figure 3.

2. ask the information entropy weight of each layer of LTP adaptive threshold (ε-LTP) eigenface

Information entropy proposes the earliest for measuring definite probabilistic quantity of information size of stochastic variable, and this uncertainty is just similar to randomness from the angle of stochastic variable, therefore calculates the stochastic distribution that information entropy is just equivalent to determine stochastic variable.In the two-layer eigenface of extracting at the self-adaptation ε shown in accompanying drawing 3-LTP, can find out that the expressed information of the different characteristic layer of image is different, and information entropy just can be expressed the quantity of information size of each layer.Measure the percentage contribution of eigenwert to image expression by characteristic layer information entropy, be characteristic layer information entropy weight.

In the time merging multilayer feature, again information entropy is defined according to the significance level of each layer of characteristic information,, using information entropy proportion as each layer of weight, computation process is as follows:

1) distribution probability of statistical nature layer j on component i:

$P_j^i=\frac{\mathrm{\&Sigma;T}\{f(x,y)=i\}}{\underset{i=0}{\overset{n-1}{\mathrm{\&Sigma;}}}\mathrm{\&Sigma;T}\{f(x,y)=i\}}(i=\mathrm{0,1},...,n-1)---\left(6\right)$

2) computing information entropy E _j:

$E_j=-\underset{i=0}{\overset{n-1}{\mathrm{\&Sigma;}}}{p}_j^i\&CenterDot;\log p_j^i---\left(7\right)$

3) the information entropy proportion of statistical nature layer j, is defined as weights W _j, characteristic layer s=2:

$W_j=E_j/\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}{E}_j(j=1,...,s)---\left(8\right)$

Information entropy weight can evaluate the significance level of each layer of feature in image characteristics extraction quantitatively, in many feature extractions and Classification and Identification, utilizes information entropy weighting to have comparatively significantly advantage.

3. each layer of LTP adaptive threshold (ε-LTP) eigenface is weighted to histogram conversion

Each layer of LTP adaptive threshold (ε-LTP) eigenface is converted to respectively to histogram according to divided piecemeal, and each layer of blocked histogram conversion regime is as follows:

$H^j=(H_0^j,H_1^j,...H_{m-1}^j),$ Wherein j=1,2 (9)

Wherein, m represents m value in histogram,

the corresponding eigenwert quantity of the each value of j layer eigenface:

$H_i^j={\mathrm{\&Sigma;}}_{x,y}I\{f_{\mathrm{\&epsiv;}-\mathrm{LTP}}^j(x,y)=i\}i=0,...m-1,j=\mathrm{1,2}---\left(10\right)$

In formula, obtain by adding up 1 number

value, by all

obtain the histogram that obtains this layer of eigenface.If

in eigenwert equate that with corresponding histogrammic value I{A} is 1, otherwise be 0:

$I\left\{A\right\}=\left\{\begin{array}{cc}1& \mathrm{Ais\; true}\\ 0& \mathrm{Ais\; false}\end{array}\right.---\left(11\right)$

Then by each layer of blocked histogram cascade, form and strengthen histogram:

H=(W ₁·H ¹,...,W _j·H ^j)(j=2) (12)

Wherein, the number of plies that j is eigenface.Finally this is strengthened to histogram as the proper vector of describing facial image.

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

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

${\mathrm{\&chi;}}^2(S,M)=\underset{i}{\mathrm{\&Sigma;}}\frac{{(S_i-M_i)}^2}{S_i+M_i}---\left(13\right)$

Wherein, S, M indicate to be respectively two the corresponding spatial enhance histogram of face H that contrast _i, i is sample identification.

5. with Nearest Neighbor Classifier, test pattern is carried out to Classification and Identification

Three rank Nearest Neighbor Classifiers are a kind of sorting techniques of simple general-purpose in recognition of face, by the χ between all histograms that obtain ²(card side) distance compares, and therefrom selects three minimum classes of distance, then selects therein optimal result as classification under test set image.

One embodiment of the present of invention are as follows:

Adopt Extended YALE B and PIE face database as experimental data base.

The facial image that Extended YALE B face database comprises 38 people, everyone has the 64 width direct pictures of taking under different light.The face position of all images is extracted and forms new face database, and often magnifying little is 192 × 168.

PIE face database comprises 68 people, everyone comprises different attitudes, expression and illumination subset, totally 41368 photos, adopt illumination subset (C27) wherein, everyone in this subset comprises 21 photos under different light, the face position of all images is extracted and forms new face database, and often magnifying little is 64 × 64.

(1) in Extended YALE B, according to the angle of illumination, strong and weak variation, this storehouse is divided into five subsets of Set1 to Set5, and experiment adopts Set1, Set4 to carry out as training set, and all the other each subsets are made test set successively.The training set image of PIE experiment is followed successively by 21 width images under everyone different light, and residual image is made test set.

Concrete test result is as shown in Fig. 5, Fig. 6 and Fig. 7, wherein, LBP representation feature extraction step adopts local binary patterns (LBP) method, LTP representation feature extracts and adopts local three binarization modes (LTP) method, ε-LTP to represent local three binarization modes (ε-LTP) method of adaptive threshold.The discrimination contrast of three kinds of methods that Fig. 1, Fig. 2 have provided respectively the good Set1 of illumination condition trains and illumination condition is poor on Extended YALE B face database Set4 while training, Fig. 1, Fig. 2 show, discrimination on Extended YALE B face database under different illumination conditions, ε-LTP method is all better than LBP and LTP method; Fig. 3 has provided the discrimination contrast of three kinds of methods on the PIE face database, Fig. 3 shows, on PIE face database, at the place an order discrimination of sample training of different illumination conditions, L ε-LTP method is all better than LBP and LTP method, and especially under poor light source condition, this advantage is comparatively outstanding.

1. for whether test self-adaptation ε-LTP has robustness to noise, be one group of experiment under noise effect below.Two kinds of situations of plus noise image that result is divided into original image and completes by analogue noise.According to the variation of noiseless and noise grade λ, observe the noise robustness of each algorithm, on Extended Yale B face database and CMU PIE face database, can be found out by table 1,2 contrast recognition effect, the recognition performance of self-adaptation ε-LTP under noise circumstance is better than other two kinds of methods, when especially noise grade is higher.And the recognition performance of additive method declines very soon with the increase of noise grade λ, visible self-adaptation ε-LTP has further improved the robustness of noise.

Anti-noise jamming experiment (%) on table 1Extended Yale B database

Anti-noise jamming experiment (%) on table 2CMU PIE database

These embodiment are interpreted as being only not used in and limiting the scope of the invention for the present invention is described above.After having read the content of record of the present invention, technician can make various changes or modifications the present invention, and these equivalences change and modification falls into the inventive method claim limited range equally.

Claims (4)

1. the face identification method based on local three binarization mode adaptive thresholds, is characterized in that, comprises the following steps:

101,, using facial image to be measured as test set, the known face database of selected part, as training set, is divided into n piece by the each facial image in training set and test set;

102, in step 101 on each piecemeal of test set after dividing equally, to each neighborhood (P, R), R represents the radius of neighbourhood, and P represents neighborhood point number, respectively the pixel g of computing center _cwith corresponding neighborhood point g _i(i=0,1 ..., P-1) pixel comparison value, try to achieve the dispersion σ of this group correlative value, by this dispersion σ as adaptive threshold ε=σ;

103, upper at threshold interval [ε ,+ε], to each neighborhood point g _i(i=0,1 ..., P-1) and central pixel point g _cgray scale difference carry out LTP coding, solve LTP adaptive threshold eigenwert, and be holotype characteristic layer and negative mode characteristic layer by this LTP adaptive threshold Eigenvalues Decomposition, the facial image of holotype characteristic layer composition is holotype characteristic layer eigenface, and the facial image of negative mode characteristic layer composition is negative mode characteristic layer eigenface;

104, the holotype characteristic layer eigenface in difference calculation procedure 103 and the information entropy weights W of negative mode characteristic layer eigenface _j,

Wherein information entropy E _jcalculating formula is:

distribution probability

f (x, y) represent distribution function

105, the holotype characteristic layer eigenface in step 103 and negative mode characteristic layer eigenface are weighted to histogram conversion, form the enhancing histogram of test set; Repeating step 102-105, the enhancing histogram of composing training collection;

106, adopt card side χ ²distance function calculates the enhancing histogram of facial image to be measured and the histogrammic χ of enhancing of training set facial image ²distance, the three rank Nearest Neighbor Classifier methods of employing select the classification under facial image to be measured, complete recognition of face.

2. the face identification method based on local three binarization mode adaptive thresholds according to claim 1, is characterized in that, in step 102, the dispersion σ acquiring method of each neighborhood correlative value is:

A, ask for central pixel point g according to formula (1) _ccontrast value in neighborhood (P, R),

△g _i=g _i-g _c,(i=0,1,...,P-1) (1)

B, ask for central pixel point g according to formula (2) _cthe average of the interior contrast value of neighborhood (P, R):

C, calculate the variance of this group contrast value according to formula (3):

D, according to formula (3) variance estimation dispersion:

。

3. the face identification method based on local three binarization mode adaptive thresholds according to claim 1, is characterized in that, the face database described in step 101 comprises Extended YALE B and PIE face database.

4. the face identification method based on local three binarization mode adaptive thresholds according to claim 1, is characterized in that, the χ in step 106 ²distance function is:

wherein, indicate the respectively test set and the corresponding spatial enhance histogram of the training set face H that contrast of S, M _i, i is sample identification.

Images