CN102332086A - Facial identification method based on dual threshold local binary pattern - Google Patents

Abstract

The invention relates to a facial identification method based on a dual threshold local binary pattern. The method comprises the following steps that: (1), local binary characteristic operation on a facial image is carried out; (2), a histogram according to a local binary pattern is obtained; (3), facial identification is realized by utilizing characteristic matching; (4), image space transformation is carried out before characteristic extraction and operation on the facial image are carried out on the facial image. Advantages of the invention are as follows: a calculating speed is fast on the basis of a characteristic matching method; an identification precision is high; sensitive degrees on a gesture, illumination, an expression and an environmental change can be reduced; and certain limitation on a facial identification technology is broken through, wherein the limitation is caused by setting of a single threshold.

Description

A kind of face identification method based on dual threshold partial binary pattern

Technical field

The invention belongs to the face recognition technology field, be specifically related to a kind of face identification method based on the dual threshold partial binary.

Background technology

Recognition of face has huge theory significance and using value. and the research of recognition of face has huge impetus for the development in fields such as image processing, pattern-recognition, computer vision, computer graphics, also has a wide range of applications in every field such as biological characteristic authentication, video monitoring, safety simultaneously.The pertinent literature of recognition of face and method enormous amount, and a face identification system often comprises several kinds of diverse ways and technology, and this is to continue to use psychologic achievement in research: promptly people's vision system carries out recognition of face through global feature and local feature.

The university of lot of domestic and foreign and research institution; MIT [1], CMU [2 like the U.S.; 3], the Cambridge University [9-12] of UIUC [4,5], Maryland university [6-8], Britain, the Toshiba company [13-14] of Japan and domestic Tsing-Hua University and Institute of Automation, CAS have all carried out extensive and deep research to recognition of face

Existing face recognition algorithms can be divided into following several types in the document: (1) is based on the face identification method of geometric properties.The geometric properties that the method adopted is the eigenvector that is the basis with the shape of human face and geometric relationship, and its component generally includes people's face and specifies the distance of point-to-point transmission, curvature, angle etc.Shortcoming is from image, to extract relatively difficulty of stable characteristics, and relatively poor to the robustness of expression shape change and attitude variation.(2) based on the face identification method of template matches: the acquisition of the method template is difficult to obtain with traditional edge acquisition methods.(3) method of elastic graph coupling.This is a kind of method based on dynamic linking structure (DLA).(4) HMM (HMM) is the one group of statistical model that is used to describe signal statistics.HMM uses the variation of Markov connection mode analog signal statistical property, and this variation is to describe through observation sequence is old indirectly.(5) based on neural network method.(6) eigenface method: obtain characterizing one group of orthogonal basis (being eigenface) in people's face space through great amount of samples being carried out principal component analysis (PCA) (PCA) exactly based on the face recognition algorithms of eigenface; The characteristic of being extracted is exactly the projection vector of facial image in this sub spaces (promptly physically facial image in the position in people's face space), adopts the sorter classification then.For example; Can regard a matrix of forming by pixel value as for piece image, also can extend, regard a vector as; Image like a width of cloth N*N pixel can be regarded as the vector that length is N2; So just think that this width of cloth image is a point that is arranged in the N2 dimension space, the vector representation of this image is exactly original image space, but this space only be can represent or many spaces of detected image in one.No matter the concrete form of subspace how; The basic thought that this method is used for image recognition all is the same; At first select a suitable subspace; Image will be projected on this sub spaces, utilize then between this projection of image certain is measured to confirm the similarity between image, and modal is exactly various distance metrics.The main performance index of measuring recognition of face has two; Misclassification rate (False Accept Rate; FAR) and reject rate (False Reject Rate, FRR), the former is meant that the general latter that other people mistake is done the designated person is meant the probability of designated person's mistake being done other personnel.

In recent years, based on local two advance binarization mode (Local Binary Pattern, face identification method LBP) receives people's attention, this method derives from the texture analysis field.LBP extracts partial transformation characteristics such as image border, angle point in essence, and these are important for distinguishing the different people face.It is each pixel and the order relation of its local neighborhood point in brightness in the computed image at first, and the two-value order relation is encoded forms local binary pattern then, adopts the feature description of multizone histogram as image at last.We in depth analyze and study LBP, and change single shortcoming to LBP, have proposed a kind of based on dual threshold local binary pattern (Double-Threshold LBP, face identification method DTLBP).At first the gray scale difference of each pixel and its local neighborhood point in the computed image forms DTLBP through selecting different threshold codings, adopts the multizone histogram vectors to carry out face characteristic then and describes, and last many threshold values of obfuscation matching result carries out recognition of face.Experimental result on face database shows that the DTLBP method all has robustness preferably to variations such as expression, background, distances.

Traditional face recognition technology generally all is to set a threshold value, promptly sets a threshold value for identification instrument in advance, when identification instrument is discerned people's face; When resolution met or exceeded this threshold value, identification instrument was discerned successfully, on the contrary when resolution is lower than prior preset threshold; Identification is unsuccessful, has prevented to a certain extent to forge and steal, but all might bring deviation to identification because preset threshold is higher or on the low side; As when threshold value is higher; Though originally people's face of typing is with instant to know others face be same people's face, owing to weather, health degree, hair style, factor such as woman'ss persona, probably identification can not could be discerned successfully through maybe passing through the long period; When threshold value is on the low side; People's face that some are approximate or make it just can discern success easily with the approximate people's face of identification instrument typing people face through artificial processing, therefore, single threshold setting has caused certain limitation to face recognition technology.

Summary of the invention

The technical matters that the present invention solved is: a kind of face identification method based on the dual threshold partial binary is provided, to solve the problem in the background technology.

A kind of face identification method based on the dual threshold partial binary, this method comprises:

(1) facial image is carried out the partial binary characteristic operation;

(2) obtain histogram from the partial binary pattern;

(3) utilize characteristic matching to realize recognition of face;

(4) before facial image being done the feature extraction computing, carry out the image space conversion.

In the present invention, people's face conversion of being adopted has the Gabor conversion, and Gassian conversion and Harr conversion have also comprised in the scheme input picture is carried out squelch, spatial resolution standardization, a plurality of pre-processing module such as space piecemeal.

A kind of face identification method based on the dual threshold partial binary, realize that this programme needs following steps:

1, with all people's face typing identification instruments to be identified, and the relevant information of typing simultaneously, be scientific two threshold values of setting of people's face identification instrument then; One high slightly; One low slightly, when people's face to be identified advances pedestrian's face when identification at the recognition of face instrument, when its resolution is higher than high slightly threshold value; Discern successfully this moment, and the identified person can directly pass through; When resolution was lower than slightly end threshold value, identification can not be successful, and the identified person can not pass through, and is given the alarm to administrative center by the recognition of face instrument; And resolution is when being between the high slightly and low slightly threshold value, and this moment, the identified person was identified the manually relevant information of the prior typing of input of instrument requirement, imports when correct when the information of all requirements, discern successfully, discerns on the contrary and gets nowhere; In recognition of face; Because the influence of the factors such as inclination, deflection and illumination variation of head; Bring very big influence can for problems such as follow-up analysis and identification; Therefore must carry out the normalization of position, size and illumination to people's face, because the human eye interocular distance is more stable, usually with this standard as geometrical normalization;

2, rotation: the slant correction of facial image is a benchmark with two lines, and image is rotated, and makes right and left eyes in postrotational image, remain on same horizontal level; The consistance that this has guaranteed people's face direction has embodied the unchangeability of people's face in the rotation of the plane of delineation, yardstick: to the facial image after the process cutting; Carry out the image zoom conversion, obtain the calibration image of unified size, for calibration back image; Suppose that w and h are the width and the height of facial image, set w=h, binocular interval is w/2; The distance on eyes and border, the image left and right sides is w/4, and the distance of eyes and image coboundary and lower boundary is respectively h/4 and 3h/4.If the size of regulation calibration image is 80 * 80; According to the position of human eye parameter; Obtain two coordinates in calibration image; Then the convergent-divergent multiple is f=(original image two interocular distances)/(calibration image two interocular distances), has guaranteed people's little consistance of being bold, and has embodied the yardstick unchangeability of people's face in the plane of delineation; In the practical application of recognition of face, illumination is uneven often, and the light conditions of people's face place environment is Protean often; People's face is detected for the high light that polarisation, sidelight cause and bright excessively, dark excessively and shade etc. and the recognition of face rate descends significantly, in program, has used the method for AMFG workshop at ICCV 2007, and the gamma correction has mainly been taked in the processing of illumination; Filtering and contrast equalization method, LBP operator have obtained application extensive in recognition of face, classify with it; Not needing learning training, is a kind of method of characteristic matching, more meets the requirement of real-time; Several kinds of LBP operators below in program, having used; The LBP operator is used to measure the contrast of topography at first, and the citation form of the initial definition of this operator acts in the 8-neighborhood of pixel, specifically; Gray-scale value with centre of neighbourhood pixel is a threshold value; 8 pixels in the neighborhood are carried out the thresholding processing, multiply by corresponding weights then, the LBP code value that has just obtained respective pixel after the results added;

The present invention has been applied to LBP in realizing _{P, R}With

Two kinds of extend types, at first, the texture T that defines in the local neighborhood in the panchromatic texture image is the joint distribution of the individual pixel corresponding grey scale value of P (P＞1):

T＝t(g _c，g ₀，g ₁，……，g _P-1) (1)

G wherein _cThe gray-scale value of center pixel in corresponding the local neighborhood, g _p(p=0,1 ..., P-1) be positioned at the gray-scale value of P pixel on the annulus that radius is R (R＞0) around corresponding the neighborhood, and this P pixel has been formed a round symmetric neighborhood collection;

If g _cCoordinate be (0,0), g so _pCoordinate figure then be:

(-Rsin(2πp/P)，Rcos(2πp/P)) (2)

If the pixel that neighborhood is concentrated is not positioned at the center of image pixel, then estimate its gray-scale value through linear interpolation;

As the first step in the derivation, do not lose any information, just with the pixel g in the circle symmetric neighborhood collection _p(p=0,1 ..., gray-scale value P-1) deducts center pixel g _cGray-scale value, obtain

T＝t(g _c，g ₀-g _c，g ₁-g _c……，g _P-1-g _c) (3)

Then, suppose difference value g _p-g _cWith g _cBetween be separate, can formula (3) be decomposed into so

T≈t(g _c)t(g ₀-g _c，g ₁-g _c……，g _P-1-g _c) (4)

In fact, above-mentioned independently is not proper independence, and therefore, the distribution after the decomposition is a kind of approximate of original joint distribution.Yet we can accept this approximate of partial information loss that cause, because it allows us to obtain the grey scale change unchangeability of operator.That is to say the t (g in the formula (4) _c) distribute and described the overall brightness of image, it and local grain have nothing to do, and also just to texture analysis useful information can't be provided.So in the defined associating intensity profile of formula (1), most of texture properties just converts following distribution into:

T≈t(g ₀-g _c，g ₁-g _c……，g _P-1-g _c) (5)

People such as Ojala find; Some specific partial binary pattern frequency of occurrences in observed image is higher than other patterns far away; Simultaneously, these partial binary patterns often have consistent loop structure: only comprised a spot of space conversion (the mutual conversion of adjacent position 0/1) in the pattern.Therefore, they are named as " consistent (Uniform) " pattern again.For the consistent pattern of formal definition, there is document to introduce the consistency metric U of space conversion times in the reflection pattern.Simultaneously, specify and to have the U value to be not more than 2 partial binary pattern be consistent pattern.And; Proposed to have as follows the texture operator of gray scale and rotational invariance, be used for texture analysis to replace

:

$\mathrm{LB}{P}_{P,R}^{\mathrm{riu}2}=\left\{\begin{array}{cc}{\mathrm{\Σ}}_{p=0}^{P-1}\mathrm{sign}(g_p-g_c),& \mathrm{if\; U}\left({\mathrm{LBP}}_{P,R}\right)\≤2\\ P+1,& \mathrm{otherwise}\end{array}\right.---\left(6\right)$

Wherein

$U\left({\mathrm{LBP}}_{P,R}\right)=|\mathrm{sign}(g_{P-1}-g_c)-\mathrm{sign}(g_0-g_c)|+\underset{p=1}{\overset{P-1}{\mathrm{\Σ}}}|\mathrm{sign}(g_p-g_c)-\mathrm{sign}(g_{p-1}-g_c)|---\left(7\right)$

The corresponding U value of consistent pattern that subscript riu2 explanation has rotational invariance is not more than 2.According to definition, can accurately obtain P+1 consistent pattern in the round symmetric neighborhood of P pixel composition.Formula (9) is in fact distributed to the unique sign of each consistent pattern according to the quantity of pattern meta set, and (0～P), it is one type that other " non-unanimity (nonuniform) " patterns are then gathered, and allocation identification (P+1).During use, from LBP _{P, R}Arrive The total P+2 of a mapping different output valve, this shines upon available one 2 ^PThe look-up tables'implementation that individual element is formed.

Use of the description of LBP histogram vectors as face characteristic; Promptly at first image is carried out LBP feature extraction conversion; Use of the description of LBP histogram vectors then, occur many improving one's methods afterwards, promptly at first facial image is carried out piecemeal as face characteristic; Carry out LBP texture analysis, statistics with histogram to every respectively, at last the histogram vectors of each piece is connected into the histogram vectors of entire image.

Beneficial effect:

The advantage of the inventive method is: based on the method for characteristic matching, computing velocity is fast; Accuracy of identification is high; Can reduce sensitivity to attitude, illumination, expression and environmental change; Overcome single threshold setting and caused certain limitation to face recognition technology.

Description of drawings

Fig. 1 is people's face geometrical normalization synoptic diagram;

Fig. 2 is the calculating synoptic diagram of partial binary pattern;

Fig. 3 is the construction process of recognition of face proper vector;

Fig. 4 is the symmetric neighborhood collection under the different PR situation;

Fig. 5 is the process flow diagram of face identification system;

Fig. 6 is the process flow diagram of recognition methods of the present invention;

Embodiment

For technological means, creation characteristic that the present invention is realized, reach purpose and effect and be easy to understand and understand, below in conjunction with specific embodiment, further set forth the present invention.

Embodiment mainly comprises following module, geometrical normalization module, illumination pretreatment module, LBP operator processing module and identification module.

In recognition of face,, therefore must carry out the normalization of position, size and illumination to people's face because the influence of the factors such as inclination, deflection and illumination variation of head brings very big influence can for problems such as follow-up analysis and identification.Because the human eye interocular distance is more stable, usually with this standard as geometrical normalization.

Rotation: the slant correction of facial image is a benchmark with two lines, and image is rotated, and makes right and left eyes in postrotational image, remain on same horizontal level.The consistance that this has guaranteed people's face direction has embodied the unchangeability of people's face in the rotation of the plane of delineation.Yardstick: to through the facial image after the cutting, carry out the image zoom conversion, obtain the calibration image of unified size.For calibration back image, suppose that w and h are the width and the height of facial image, set w=h, binocular interval is w/2, and the distance on eyes and border, the image left and right sides is w/4, and the distance of eyes and image coboundary and lower boundary is respectively h/4 and 3h/4.If the size of regulation calibration image is 80 * 80; According to the position of human eye parameter; Obtain two coordinates in calibration image; Then the convergent-divergent multiple is f=(original image two interocular distances)/(calibration image two interocular distances), has guaranteed people's little consistance of being bold, and has embodied the yardstick unchangeability of people's face in the plane of delineation.

In the practical application of recognition of face, illumination is uneven often.The light conditions of people's face place environment is Protean often, and the high light that polarisation, sidelight cause and bright excessively, dark excessively and shade etc. all can make detection of people's face and recognition of face rate descend significantly.In program, used the method for AMFG workshop at ICCV 2007, the gamma correction has mainly been taked in the processing of illumination, filtering and contrast equalization method.

The LBP operator has obtained application extensive in recognition of face, use its classification, does not need learning training, is a kind of method of characteristic matching, more meets the requirement of real-time.Several kinds of LBP operators below in program, having used.The LBP operator is used to measure the contrast of topography at first.The citation form of the initial definition of this operator acts in the 8-neighborhood of pixel.Specifically, be threshold value with the gray-scale value of centre of neighbourhood pixel, 8 pixels in the neighborhood are carried out the thresholding processing, multiply by corresponding weights then, the LBP code value that has just obtained respective pixel after the results added, as shown in Figure 2.The present invention has been applied to LBP in realizing _{P, R}With

Two kinds of extend types.At first, the texture T that defines in the local neighborhood in the panchromatic texture image is the joint distribution of the individual pixel corresponding grey scale value of P (P＞1):

T＝t(g _c，g ₀，g ₁，……，g _P-1)

(1)

G wherein _cThe gray-scale value of center pixel in corresponding the local neighborhood, g _p(p=0,1 ..., P-1) be positioned at the gray-scale value of P pixel on the annulus that radius is R (R＞0) around corresponding the neighborhood, and this P pixel has been formed a round symmetric neighborhood collection.

If g _cCoordinate be (0,0), g so _pCoordinate figure then be:

(-Rsin(2πp/P)，Rcos(2πp/P))

(2)

That Fig. 4 has shown is various (P, the round symmetric neighborhood collection under R).If the pixel that neighborhood is concentrated is not positioned at the center of image pixel, then estimate its gray-scale value through linear interpolation.

As the first step in the derivation, do not lose any information, just with the pixel g in the circle symmetric neighborhood collection _p(p=0,1 ..., gray-scale value P-1) deducts center pixel g _cGray-scale value, obtain

T＝t(g _c，g ₀-g _c，g ₁-g _c……，g _P-1-g _c)

(3)

Then, suppose difference value g _p-g _cWith g _cBetween be separate, can formula (3) be decomposed into so

T≈t(g _c)t(g ₀-g _c，g ₁-g _c……，g _P-1-g _c)

(4)

In fact, above-mentioned independently is not proper independence, and therefore, the distribution after the decomposition is a kind of approximate of original joint distribution.Yet we can accept this approximate of partial information loss that cause, because it allows us to obtain the grey scale change unchangeability of operator.That is to say the t (g in the formula (4) _c) distribute and described the overall brightness of image, it and local grain have nothing to do, and also just to texture analysis useful information can't be provided.So in the defined associating intensity profile of formula (1), most of texture properties just converts following distribution into:

T≈t(g ₀-g _c，g ₁-g _c……，g _P-1-g _c)

(5)

People such as Ojala find; Some specific partial binary pattern frequency of occurrences in observed image is higher than other patterns far away; Simultaneously, these partial binary patterns often have consistent loop structure: only comprised a spot of space conversion (the mutual conversion of adjacent position 0/1) in the pattern.Therefore, they are named as " consistent (Uniform) " pattern again.For the consistent pattern of formal definition, there is document to introduce the consistency metric U of space conversion times in the reflection pattern.Simultaneously, specify and to have the U value to be not more than 2 partial binary pattern be consistent pattern.And; Proposed to have as follows the texture operator of gray scale and rotational invariance, be used for texture analysis to replace :

$\mathrm{LB}{P}_{P,\mathrm{<figure-callout\; id="2"\; label="R\; riu"\; filenames="HSA00000518005500011.png"\; state="\{\{state\}\}">R</figure-callout>}}^{\mathrm{riu}}=\left\{\begin{array}{cc}{\mathrm{\&Sigma;}}_{p=0}^{P-1}\mathrm{sign}(g_p-g_c),& \mathrm{if\; U}\left({\mathrm{LBP}}_{P,R}\right)\&le;2\\ P+1,& \mathrm{otherwise}\end{array}\right.---\left(6\right)$

Wherein

$U\left({\mathrm{LBP}}_{P,R}\right)=|\mathrm{sign}(g_{P-1}-g_c)-\mathrm{sign}(g_0-g_c)|+\underset{p=1}{\overset{P-1}{\mathrm{\&Sigma;}}}|\mathrm{sign}(g_p-g_c)-\mathrm{sign}(g_{p-1}-g_c)|---\left(7\right)$

The corresponding U value of consistent pattern that subscript riu2 explanation has rotational invariance is not more than 2.According to definition, can accurately obtain P+1 consistent pattern in the round symmetric neighborhood of P pixel composition.Formula (9) is in fact distributed to the unique sign of each consistent pattern according to the quantity of pattern meta set, and (0～P), it is one type that other " non-unanimity (nonuniform) " patterns are then gathered, and allocation identification (P+1).During use, from LBP _{P, R}Arrive

The total P+2 of a mapping different output valve, this shines upon available one 2 ^PThe look-up tables'implementation that individual element is formed.

Use of the description of LBP histogram vectors as face characteristic; Promptly at first image is carried out LBP feature extraction conversion, use of the description of LBP histogram vectors then, occur many improving one's methods afterwards as face characteristic; Promptly at first facial image is carried out piecemeal; Carry out LBP texture analysis, statistics with histogram to every respectively, at last the histogram vectors of each piece is connected into the histogram vectors of entire image, concrete process is as shown in Figure 3.

With certain enterprise's attendance recorder is example, and whole flow process comprises people's face to be identified typing, relevant information typing, dual threshold setting, identification, judgement, manually input, success or warning.At first, with all (employee) people's face 1 typing to be identified (attendance recorder) identification instruments 2 of certain enterprise, and the relevant information of typing simultaneously 3, relevant information 3 comprises job number, identification card number, cell-phone number or the like; Be two threshold values of setting of (attendance recorder) recognition of face instrument 2 sciences then; One 70%; One 80%, when (employee) people's face 1 to be identified advances pedestrian's face when identification at (attendance recorder) recognition of face instrument 2, its resolution reached or greater than 80% o'clock; Discern successfully this moment, and (employee) identified person 1 can directly pass through; Resolution is lower than at 70% o'clock, and identification can not be successful, and (employee) identified person 1 can not pass through, and is given the alarm to administrative center 4 by (attendance recorder) recognition of face instrument 2; And resolution be in 70% and 80% between the time, (employee) identified person 1 is required the manually relevant information 3 of the prior typing of input by (attendance recorder) identification instrument 2 at this moment, imports when correct when the information of all requirements, discern successfully, discerns on the contrary and gets nowhere.

The above only is a preferred implementation of the present invention, and protection scope of the present invention also not only is confined to the foregoing description, and all technical schemes that belongs under the thinking of the present invention all belong to protection scope of the present invention.Should be pointed out that for those skilled in the art in the some improvement and the retouching that do not break away under the principle of the invention prerequisite, these improvement and retouching also should be regarded as protection scope of the present invention.

Claims (2)

1. the face identification method based on the dual threshold partial binary is characterized in that, this method comprises:

(1) facial image is carried out the partial binary characteristic operation;

(2) obtain histogram from the partial binary pattern;

(3) utilize characteristic matching to realize recognition of face;

(4) before facial image being done the feature extraction computing, carry out the image space conversion.

2. a kind of face identification method based on the dual threshold partial binary according to claim 1 is characterized in that the conversion of described people's face has the Gabor conversion, Gassian conversion and Harr conversion.

Images