CN103218605B - A kind of fast human-eye positioning method based on integral projection and rim detection - Google Patents

Abstract

The invention discloses a kind of fast human-eye positioning method based on integral projection and rim detection, it mainly contains following step: the first step: turn gray scale to the facial image detected, and utilizes wave filter to its smoothing denoising; Second step: adopt integrate levels sciagraphy to obtain human eye Position Approximate; 3rd step: edge extracting is carried out to the image in the first step, and by image binaryzation; 4th step: calculate row, column complexity, accurately locate position of human eye; 5th step: correct, obtain final position of human eye.The method computing velocity is fast, and the impact that can effectively suppress the ornaments in facial image to be located human eye, has the advantage of fast and stable.

Description

A kind of fast human-eye positioning method based on integral projection and rim detection

Technical field

The invention belongs to the characteristic point positioning method in pattern-recognition, particularly relate to a kind of fast human-eye positioning method based on integral projection and rim detection, a kind of fast and convenient solution proposed mainly for human eye orientation problem in input picture, can locate human eye fast and effectively.

Background technology

Computer face identification is the research field enlivened very much in recent years.Being of wide application of it, as Sex, Age analysis, security system authentication, Expression analysis and video conference etc.It mainly comprises Face datection, feature location is extracted and the several step of feature identification.Can human eye, as a key feature of face, accurately have tremendous influence to the result of feature extraction and feature identification in location to it.

Herein based on half-tone information and the marginal information of image, one human eye detection algorithm is fast and effectively proposed.This algorithm can quick position human eye, and can be good at the impact that suppresses illumination and jewelry to bring to positioning result.

Summary of the invention

The invention provides the fast human-eye positioning method based on integral projection and rim detection that a kind of succinct and accuracy is high.

In order to realize this target, the present invention takes following technical scheme:

Step 1: initialization, reads in an image collected containing face ,

Step 2: utilize Adaboost algorithm to the digital picture collected carry out Face datection operation, get facial image wherein ,

Step 3: to the facial image acquired in step 2 carry out pre-service, method is as follows:

Step 3.1: by the facial image acquired be converted into facial image gray level image, and by facial image gray level image be normalized to W the image of H , wherein W, H are positive integer, represent facial image respectively gray level image be normalized to W the image of H line number and columns,

Step 3.2: utilize Gaussian filter to W the image of H smoothing, denoising, concrete grammar is as follows: first, the image after level and smooth, denoising the gray-scale value of each boundary pixel be respectively smoothly, W before denoising the image of H the gray-scale value of each boundary pixel; Secondly, for W the image of H the gray-scale value of non-border pixel, then with W the image of H in any one non-border pixel centered by pixel, choose 3 the Gaussian template of 3, obtains the gray-scale value of described center pixel, that is:

In formula, represent row-coordinate, represent row coordinate, for image in the gray-scale value of point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some lower-left angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some upper left angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some bottom right angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some upper right angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of the horizontal left direction point of point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of point immediately below point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of the horizontal right direction point of point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of point directly over point, for after gaussian filtering the gray-scale value of point, obtains image after process , its size is still W h, traversal W the image of H in all non-border pixel,

Step 4: to image carry out integrate levels projection, projection computing formula is as follows:

In formula, represent the the integrate levels projection result of row, for image 's the gray-scale value at some place, W is picturedeep, and H is picturewide,

From the integrate levels projection result that W is capable, choose the minimum integrate levels projection result of gray-scale value, and be designated as , for making integrate levels projection computing formula when obtaining minimum value corresponding line number, from image in select vertical coordinate scope and be region, as the region to be detected that human eye may exist, wherein, , represent value be round value downwards, for the vertical coordinate interval estimation parameter that human eye may exist,

Step 5: to image carry out edge extracting, specific implementation process is as follows:

Step 5.1: first, carries out the gradient magnitude matrix correspondence image after gradient calculation the grey scale pixel value of each frontier point be image before carrying out gradient calculation the grey scale pixel value of frontier point, secondly, for image middle non-border pixel point , choose , , for horizontal direction edge detection operator, for vertical direction edge detection operator, calculate non-border pixel point the horizontal direction at place, the first-order partial derivative of vertical direction, and non-border pixel point the gradient magnitude at place and gradient direction, computing formula is as follows:

In formula representative image middle non-border pixel point gray-scale value, i is image in the horizontal coordinate of non-frontier point, j is image in the vertical coordinate of non-frontier point, for image in with non-border pixel point point is top left hand element grid in be positioned at the gray-scale value of the pixel in the grid lower left corner, for image in with non-border pixel point point is top left hand element grid in be positioned at the gray-scale value of the pixel in the grid upper right corner, for image in with non-border pixel point point is top left hand element grid in be positioned at the gray-scale value of the pixel in the grid lower right corner, , represent non-border pixel point respectively the horizontal direction at place, the first-order partial derivative of vertical direction, represent non-border pixel point gradient magnitude, represent non-border pixel point gradient direction,

Step 5.2: by the gradient direction value of each non-border pixel point obtained in step 5.1 carry out discretize and obtain new gradient direction value , choose with non-border pixel point centered by 3 3 windows, right value process, computing formula is as follows:

In formula, with represent respectively with non-border pixel point centered by 3 in 3 windows along the gradient magnitude of two pixels in direction, for the gradient magnitude after above formula process, with the gradient magnitude after processing as central element, image frontier point grey scale pixel value as boundary element structural matrix, this matrix correspondence produces image ,

Step 6: to the image obtained in step 5.2 carry out binaryzation, concrete grammar is as follows:

Step 6.1: adopt maximum variance between clusters definite threshold, the process of definite threshold is as follows:

Statistical picture in the gray-scale value of each pixel, be integer using the highest gray-scale value as most high grade grey level m, m, then image grayscale range is made up of each round values in interval [0, m], and gray level is that the number of pixels of t is set to , then total number of pixels , the probability of each gray-scale value is if use integer gray scale is divided into two groups , , for gray level is less than or equal to gray level group, for gray level is greater than be less than gray level group, utilize the variance between following formulae discovery two gray level groups:

In formula, for general image mean value, for the mean value of group, for the probability of group, chooses the value of each round values in interval [0, m] as k respectively, then calculates each variance yields corresponding to k value respectively , and maximum variance yields is selected from m+1 variance yields, then using the k value corresponding to maximum variance yields as threshold value T,

Step 6.2: use the threshold value T obtained in step 6.1 to image carry out binary conversion treatment, image its gray-scale value is set to 255 by the pixel that middle gray-scale value is more than or equal to threshold value T, and the grey scale pixel value lower than threshold value T is set to 0, obtains bianry image ,

Step 7: horizontal level and the upright position of accurately determining human eye, specific implementation process is as follows:

Step 7.1: for bianry image , calculate row, column complexity function, computing formula is as follows:

In formula, represent row complexity function, represent row complexity function, representative image in be positioned at the pixel value of the pixel at place, , for the human face region parameter obtained in step 4,

Step 7.2: adopt known mean filter, respectively to row complexity function , row complexity function carry out one dimension low-pass filtering and obtain new row complexity function , new row complexity function , find new row complexity function by calculating maximum of points and new row complexity function maximum point determination position of human eye coordinate, through calculating function a maximum of points , function two maximum points , , obtain pixel , pixel ,

Step 7.3: the new image obtained in step 3.2 in, select the pixel that step 7.2 obtains respectively , pixel 6 6 neighborhoods, minimum two pixels of the gray-scale value in this field are set as human eye center, and the position coordinates of minimum for gray-scale value two pixels is defined as eyes coordinate , .

Compared with prior art, feature of the present invention is:

First this algorithm carried out integral projection process to image before accurately locating human eye, so just roughly can determine the region that human eye may exist, eliminate other interference region, the impact that such as face, nose bring, can greatly save calculating required time like this, also can be good at the positioning result improving final human eye simultaneously.In addition, because the change of human eye area is compared to other organ more complicated of face, edge variation, than stronger, adopts the method for rim detection can identify human eye area simply and effectively.

Accompanying drawing explanation

Fig. 1 is fast human-eye location algorithm process flow diagram.

Fig. 2 is gradient direction angular discretization standard drawing.

Fig. 3 is facial image integrate levels perspective view.

Fig. 4 is the edge detection results figure of human eye area and this human eye area.

Fig. 5 is the row complexity function schematic diagram of human eye area edge detection results figure.

Fig. 6 is the row complexity function schematic diagram of human eye area edge detection results figure.

Fig. 7 is the result schematic diagram of row complexity function after low-pass filtering of human eye area edge detection results figure.

Fig. 8 is the result schematic diagram of row complexity function after low-pass filtering of human eye area edge detection results figure.

Embodiment

In a particular embodiment, will by reference to the accompanying drawings, the clear detailed implementation fully describing fast human-eye location algorithm,

A kind of fast human-eye positioning method, is characterized in that carrying out according to following steps:

Step 1: initialization, reads in an image collected containing face ,

Step 2: utilize Adaboost algorithm to the digital picture collected carry out Face datection operation, get facial image wherein ,

Step 3: to the facial image acquired in step 2 carry out pre-service, method is as follows:

Step 3.1: by the facial image acquired be converted into facial image gray level image, computing formula is as follows:

for processing the brightness value of rear pixel, R, G, B are the relative intensity of three primary colours, by facial image gray level image be normalized to W the image of H , wherein W, H are positive integer, represent facial image respectively gray level image be normalized to W the image of H line number and columns,

Step 3.2: utilize Gaussian filter to W the image of H smoothing, denoising, concrete grammar is as follows: first, the image after level and smooth, denoising the gray-scale value of each boundary pixel be respectively smoothly, W before denoising the image of H the gray-scale value of each boundary pixel; Secondly, for W the image of H the gray-scale value of non-border pixel, then with W the image of H in any one non-border pixel centered by pixel, choose 3 the Gaussian template of 3, obtains the gray-scale value of described center pixel, that is:

In formula, represent row-coordinate, represent row coordinate, for image in the gray-scale value of point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some lower-left angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some upper left angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some bottom right angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of some upper right angle point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of the horizontal left direction point of point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of point immediately below point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of the horizontal right direction point of point, for image in with centered by point 3 be positioned in 3 grids the gray-scale value of point directly over point, for after gaussian filtering the gray-scale value of point, obtains image after process , its size is still W h, traversal W the image of H in all non-border pixel,

Step 4: to image carry out integrate levels projection, projection computing formula is as follows:

In formula, represent the the integrate levels projection result of row, for image 's the gray-scale value at some place, W is picturedeep, and H is picturewide,

From the integrate levels projection result that W is capable, choose the minimum integrate levels projection result of gray-scale value, and be designated as , for making integrate levels projection computing formula when obtaining minimum value corresponding line number, from image in select vertical coordinate scope and be region, as the region to be detected that human eye may exist, wherein, , represent value be round value downwards, for the vertical coordinate interval estimation parameter that human eye may exist,

Step 5: to image carry out edge extracting, specific implementation process is as follows:

Step 5.1: first, carries out the gradient magnitude matrix correspondence image after gradient calculation the grey scale pixel value of each frontier point be image before carrying out gradient calculation the grey scale pixel value of frontier point, secondly, for image middle non-border pixel point , choose , , for horizontal direction edge detection operator, for vertical direction edge detection operator, calculate non-border pixel point the horizontal direction at place, the first-order partial derivative of vertical direction, and non-border pixel point the gradient magnitude at place and gradient direction, computing formula is as follows:

In formula representative image middle non-border pixel point gray-scale value, i is image in the horizontal coordinate of non-frontier point, j is image in the vertical coordinate of non-frontier point, for image in with non-border pixel point point is top left hand element grid in be positioned at the gray-scale value of the pixel in the grid lower left corner, for image in with non-border pixel point point is top left hand element grid in be positioned at the gray-scale value of the pixel in the grid upper right corner, for image in with non-border pixel point point is top left hand element grid in be positioned at the gray-scale value of the pixel in the grid lower right corner, , represent non-border pixel point respectively the horizontal direction at place, the first-order partial derivative of vertical direction, represent non-border pixel point gradient magnitude, represent non-border pixel point gradient direction,

Step 5.2: by the gradient direction value of each non-border pixel point obtained in step 5.1 carry out discretize according to the discretize standard of accompanying drawing 2 and obtain new gradient direction value , choose with non-border pixel point centered by 3 3 windows, right value process, computing formula is as follows:

In formula, with represent respectively with non-border pixel point centered by 3 in 3 windows along the gradient magnitude of two pixels in direction, for the gradient magnitude after above formula process, with the gradient magnitude after processing as central element, image frontier point grey scale pixel value as boundary element structural matrix, this matrix correspondence produces image ,

Step 6: to the image obtained in step 5.2 carry out binaryzation, concrete grammar is as follows:

Step 6.1: adopt maximum variance between clusters definite threshold, the process of definite threshold is as follows:

Statistical picture in the gray-scale value of each pixel, be integer using the highest gray-scale value as most high grade grey level m, m, then image grayscale range is made up of each round values in interval [0, m], and gray level is that the number of pixels of t is set to , then total number of pixels , the probability of each gray-scale value is if use integer gray scale is divided into two groups , , for gray level is less than or equal to gray level group, for gray level is greater than be less than gray level group, utilize the variance between following formulae discovery two gray level groups:

In formula, for general image mean value, for the mean value of group, for the probability of group, chooses the value of each round values in interval [0, m] as k respectively, then calculates each variance yields corresponding to k value respectively , and maximum variance yields is selected from m+1 variance yields, then using the k value corresponding to maximum variance yields as threshold value T,

Step 6.2: use the threshold value T obtained in step 6.1 to image carry out binary conversion treatment, image its gray-scale value is set to 255 by the pixel that middle gray-scale value is more than or equal to threshold value T, and the grey scale pixel value lower than threshold value T is set to 0, obtains bianry image ,

Step 7: horizontal level and the upright position of accurately determining human eye, specific implementation process is as follows:

Step 7.1: for bianry image , calculate row, column complexity function, computing formula is as follows:

In formula, represent row complexity function, represent row complexity function, representative image in be positioned at the pixel value of the pixel at place, , for the human face region parameter obtained in step 4,

Step 7.2: adopt known mean filter, respectively to row complexity function , row complexity function carry out one dimension low-pass filtering and obtain new row complexity function , new row complexity function , find new row complexity function by calculating maximum of points and new row complexity function maximum point determination position of human eye coordinate, through calculating function a maximum of points , function two maximum points , , obtain pixel , pixel ,

Step 7.3: the new image obtained in step 3.2 in, select the pixel that step 7.2 obtains respectively , pixel 6 6 neighborhoods, minimum two pixels of the gray-scale value in this field are set as human eye center, and the position coordinates of minimum for gray-scale value two pixels is defined as eyes coordinate , .

Claims (1)

1., based on a fast human-eye positioning method for integral projection and rim detection, it is characterized in that carrying out according to following steps:

Step 1: initialization, reads in an image I collected containing face ₁,

Step 2: utilize Adaboost algorithm to the digital picture I collected ₁carry out Face datection operation, get facial image I wherein ₂,

Step 3: to the facial image I acquired in step 2 ₂carry out pre-service, method is as follows:

Step 3.1: by the facial image I acquired ₂be converted into facial image I ₂gray level image, and by facial image I ₂gray level image be normalized to the image I of W × H ₃, wherein W, H are positive integer, represent facial image I respectively ₂gray level image be normalized to the image I of W × H ₃line number and columns,

Step 3.2: utilize Gaussian filter to the image I of W × H ₃smoothing, denoising, concrete grammar is as follows: first, the image I after level and smooth, denoising ₄the gray-scale value of each boundary pixel be respectively smoothly, the image I of W × H before denoising ₃the gray-scale value of each boundary pixel; Secondly, for the image I of W × H ₃the gray-scale value of non-border pixel, then with the image I of W × H ₃in any one non-border pixel centered by pixel, choose the Gaussian template of 3 × 3, obtain the gray-scale value of described center pixel, that is:

g ₄(x,y)＝{g ₃(x-1,y-1)+g ₃(x-1,y+1)+g ₃(x+1,y-1)+

g ₃(x+1,y+1)+[g ₃(x-1,y)+g ₃(x,y-1)+

g ₃(x+1,y)+g ₃(x,y+1)]×2+g ₃(x,y)×4}/16

In formula, y represents row-coordinate, and x represents row coordinate, g ₃(x, y) is image I ₃in (x, y) gray-scale value of putting, g ₃(x-1, y-1) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at the gray-scale value that (x, y) puts lower-left angle point, g ₃(x-1, y+1) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at the gray-scale value that (x, y) puts upper left angle point, g ₃(x+1, y-1) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at the gray-scale value that (x, y) puts bottom right angle point, g ₃(x+1, y+1) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at the gray-scale value that (x, y) puts upper right angle point, g ₃(x-1, y) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at the gray-scale value that (x, y) puts horizontal left direction point, g ₃(x, y-1) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at (x, y) point immediately below point gray-scale value, g ₃(x+1, y) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at the gray-scale value that (x, y) puts horizontal right direction point, g ₃(x, y+1) is image I ₃in with (x, y) point centered by 3 × 3 grids in be positioned at (x, y) point directly over point gray-scale value, g ₄(x, y) is the gray-scale value that (x, y) puts after gaussian filtering, obtains image I after process ₄, its size is still W × H, the image I of traversal W × H ₃in all non-border pixel,

Step 4: to image I ₄carry out integrate levels projection, projection computing formula is as follows:

$1\left(r\right)=\underset{s=1}{\overset{H}{\Σ}}{g}_4(s,r),1\≤r\≤W$

In formula, l (r) represents the integrate levels projection result that r is capable, g ₄(s, r) is image I ₄the gray-scale value at (s, r) some place, W is picturedeep, and H is picturewide,

From the integrate levels projection result that W is capable, choose the minimum integrate levels projection result of gray-scale value, make integrate levels line number r corresponding to of computing formula l (r) when obtaining minimum value that project be c, from image I ₄in select vertical coordinate scope for (c-δ, c+ δ) region, as the region to be detected that human eye may exist, wherein, represent the value of δ be H/12 round value downwards, δ is the vertical coordinate interval estimation parameter that human eye may exist,

Step 5: to image I ₄carry out edge extracting, specific implementation process is as follows:

Step 5.1: first, carries out the gradient magnitude matrix correspondence image I after gradient calculation ₅the grey scale pixel value of each frontier point be image I before carrying out gradient calculation ₄the grey scale pixel value of frontier point, secondly, for image I ₄middle non-border pixel point (i, j), chooses $S_x=\left[\begin{array}{cc}-1& 1\\ -1& 1\end{array}\right],$ $S_y=\left[\begin{array}{cc}1& 1\\ -1& -1\end{array}\right],$ S _xfor horizontal direction edge detection operator, S _yfor vertical direction edge detection operator, calculate the horizontal direction at non-border pixel point (i, j) place, the first-order partial derivative of vertical direction, and the gradient magnitude at non-border pixel point (i, j) place and gradient direction, computing formula is as follows:

P(i,j)＝[f(i,j+1)-f(i,j)+f(i+1,j+1)-f(i+1,j)]/2

Q(i,j)＝[f(i,j)-f(i+1,j)+f(i,j+1)-f(i+1,j+1)]/2

$M(i,j)=\sqrt{P{(i,j)}^2+Q{(i,j)}^2}$

θ(i，j)＝arctan[Q(i，j)/P(i，j)]

F (i, j) representative image I in formula ₄the gray-scale value of middle non-border pixel point (i, j), i is image I ₄in the horizontal coordinate of non-frontier point, j is image I ₄in the vertical coordinate of non-frontier point, f (i+1, j) is image I ₄in be the gray-scale value being positioned at the pixel in the grid lower left corner in the grid of 2 × 2 of top left hand element with non-border pixel point (i, j) point, f (i, j+1) is image I ₄in be the gray-scale value being positioned at the pixel in the grid upper right corner in the grid of 2 × 2 of top left hand element with non-border pixel point (i, j) point, f (i+1, j+1) is image I ₄in with non-border pixel point (i, j) for top left hand element 2 × 2 grid in be positioned at the gray-scale value of the pixel in the grid lower right corner, P (i, j), Q (i, j) non-border pixel point (i is represented respectively, j) horizontal direction at place, the first-order partial derivative of vertical direction, M (i, j) non-border pixel point (i is represented, j) gradient magnitude, θ (i, j) represents non-border pixel point (i, j) gradient direction

Step 5.2: by the θ (i obtained in step 5.1, j) carry out discretize and obtain new gradient direction value θ ' (i, j), choose with non-border pixel point (i, j) 3 × 3 windows centered by, process the value of M (i, j), computing formula is as follows:

$M^{\&prime;}(i,j)=\left\{\begin{array}{cc}0,& M(i,j)<{\mathrm{\&omega;}}_{1}orM(i,j)<{\mathrm{\&omega;}}_2\\ M(i,j),& else\end{array}\right.$

In formula, ω ₁and ω ₂represent with non-border pixel point (i respectively, j) along θ ' (i in 3 × 3 windows centered by, j) gradient magnitude of two pixels in direction, M'(i, j) be gradient magnitude after above formula process, gradient magnitude M'(i, j after process) as central element, image I ₄frontier point grey scale pixel value as boundary element structural matrix, this matrix correspondence produces image I ₅,

Step 6: to the image I obtained in step 5.2 ₅carry out binaryzation, concrete grammar is as follows:

Step 6.1: adopt maximum variance between clusters definite threshold, the process of definite threshold is as follows:

Statistical picture I ₅in the gray-scale value of each pixel, be integer using the highest gray-scale value as most high grade grey level m, m, then image grayscale range is made up of each round values in interval [0, m], and gray level is that the number of pixels of t is set to N _t, then total number of pixels the probability of each gray-scale value is p _t=N _t/ N, if be divided into two groups of G by integer k (0≤k≤m) by gray scale ₁=0,1 ..., k}, G ₂=k+1 ..., m}, G ₁for gray level is less than or equal to the gray level group of k, G ₂for gray level is greater than the gray level group that k is less than or equal to m, utilize the variance between following formulae discovery two gray level groups:

${\mathrm{\&sigma;}}^2\left(k\right)=\frac{{\&lsqb;\mathrm{\&mu;}\mathrm{\&omega;}\left(k\right)-\mathrm{\&mu;}\left(k\right)\&rsqb;}^2}{\mathrm{\&omega;}\left(k\right)\&lsqb;1-\mathrm{\&omega;}\left(k\right)\&rsqb;}$

In formula, $\mathrm{\&mu;}=\underset{t=0}{\overset{m}{\&Sigma;}}{\mathrm{tp}}_t$ For general image mean value, $\mathrm{\&mu;}\left(k\right)=\underset{t=0}{\overset{k}{\mathrm{\&Sigma;}}}t{p}_t$ For G ₁the mean value of group, $\mathrm{\&omega;}\left(k\right)=\underset{t=0}{\overset{k}{\&Sigma;}}{p}_t$ For G ₁the probability of group, chooses the value of each round values in interval [0, m] as k respectively, then calculates each variance yields σ corresponding to k value respectively ²(k), and maximum variance yields is selected from m+1 variance yields, then using the k value corresponding to maximum variance yields as threshold value T,

Step 6.2: use the threshold value T obtained in step 6.1 to image I ₅carry out binary conversion treatment, image I ₅its gray-scale value is set to 255 by the pixel that middle gray-scale value is more than or equal to threshold value T, and the grey scale pixel value lower than threshold value T is set to 0, obtains bianry image I ₆,

Step 7: horizontal level and the upright position of accurately determining human eye, specific implementation process is as follows:

Step 7.1: for bianry image I ₆, calculate row, column complexity function, computing formula is as follows:

$f_H\left(p\right)=\underset{q=1}{\overset{H}{\&Sigma;}}{I}_6(p,q),c-\mathrm{\&delta;}\&le;p\&le;c+\mathrm{\&delta;}$

$f_L\left(q\right)=\underset{p=c-\mathrm{\&delta;}}{\overset{c+\mathrm{\&delta;}}{\&Sigma;}}{I}_6(p,q),1\&le;q\&le;H$

In formula, f _hp () represents row complexity function, f _lq () represents row complexity function, I ₆(p, q) representative image I ₆in be positioned at the pixel value of the pixel at (p, q) place, c, δ are the human face region parameter obtained in step 4,

Step 7.2: adopt mean filter, respectively to row complexity function f _h(p), row complexity function f _lq () is carried out one dimension low-pass filtering and is obtained new row complexity function f _h' (p), new row complexity function f _l' (q), find new row complexity function f by calculating _h' the maximum of points of (p) and new row complexity function f _l' the maximum point determination position of human eye coordinate of (q), through calculating function f _h' the maximum of points p of (p) ₁, function f _l' two maximum point q of (q) ₁, q ₂, obtain pixel (p ₁, q ₁), pixel (p ₁, q ₂),

Step 7.3: the new image I obtained in step 3.2 ₄in, select the pixel (p that step 7.2 obtains respectively ₁, q ₁), pixel (p ₁, q ₂) 6 × 6 neighborhoods, minimum two pixels of the gray-scale value in this neighborhood are set as human eye center, and the position coordinates of minimum for gray-scale value two pixels is defined as eyes coordinate (μ ₁, ν ₁), (μ ₂, ν ₂).