CN105426882A - Method for rapidly positioning human eyes in human face image - Google Patents

Abstract

The invention belongs to the field of machine learning, and particularly provides a method for rapidly positioning human eyes in a human face image. Human eye positioning is performed by adopting a random forest algorithm based on a self-adaptive gradient promotion decision tree, and multi-posture and diversified processing is performed on samples so that the method is enabled to have better robustness. Meanwhile, a multistage positioning structure is adopted, i.e. the position of previous positioning acts as a center to narrow searching range so that convergence speed and accuracy can be enhanced. Finally weighted averaging is performed on the result of the random forest so that the coordinate positions of the human eyes are determined and real-time performance is great. Positioning precision is further enhanced, and time consumption is reduced by a low-complexity algorithm based on a binary tree. Learning from a large number of diversified human eye samples is performed based on a CART model to obtain a regression tree so that robustness is greatly enhanced, substantial performance gain is realized and the method can be applied to the field of human eye positioning.

Description

A kind of method of quick position human eye in facial image

Technical field

The invention belongs to computer vision and image processing field, specifically a kind of method of quick position human eye in facial image, is applied in eye recognition field.

Background technology

No matter be interpersonal interchange, or in the process of man-machine interaction, human eye, as one of most important feature of face, have very important effect to the location of human eye.Internet era at the beginning of, the protection of privacy information and its security just start to become hot issue, and iris recognition is as a kind of method higher than finger print safety property, enters the visual field of people.The space encoder of iris recognition enriches, and everyone iris lines is unique and easily distinguish.Moreover, iris in human body can show different proterties under different illumination conditions, therefore only have living body iris can pass through to detect, this has just stopped the similar potential safety hazard copying fingerprint and so on, makes iris recognition become the highest identity recognizing technology of current security.But people wish to obtain security high as far as possible under the Consumer's Experience of close friend.A user-friendly example is exactly recognition of face, and in open occasion (as safety check, gate inhibition), face can be caught by hardware device in unware situation user, therefore can not bring added burden to user.

In field of human-computer interaction, human eye location also has a wide range of applications.Along with the appearance of the wear-type wearable devices such as GoogleGlass, people start to seek the novel interactive mode except keyboard and mouse combination, and the hot topic that the human eye nearest apart from equipment just becomes wherein is alternative.Had at present some in-plant hardware devices to come out, such as PupilLabs company develops the equipment that one adopts the similar glasses of wear-type assembly and some cameras composition, is used for catching the eye movement of user.These human eye tracer techniques can allow people not make to use eyes to control to browse webpage in bimanual situation, player can be assisted to control the personage's motion in game, can also apply in the application scenarioss such as virtual reality.And the oculocentric location of people is the important ring in this kind of human-computer interaction technology.When human eye sight is mobile on screen time, the rotation of eyeball is very small, and therefore to realize the precise controlling such as the clicking operation of web page browsing even on screen, accurate human eye location technology is by indispensable.

Human eye is positioned at other field and is also widely used.Such as the pernicious traffic hazard that fatigue driving causes, the modes such as much research uses mobile unit to observe, and the eyeball of driver moves, eyelid closedown, facial expression, head move judge driver's whether fatigue driving, and take the necessary measures when there is fatigue driving.In recent ten years, human eye location receives the concern of more and more academia and industry member.

The method of many relative maturity has been had at present in human eye location.As: " Anintegrating; transformation-orientedapproachtoconcurrencycontrolandun doingroupeditors " that M.Ressel delivers on Proceedingsofthe1996ACMconferenceonComputersupportedcoop erativework utilizes the geometric object of Hough transformation to rule to identify, locates human eye in the drawings." Automaticadaptivecenterofpupildetectionusingfacedetectio nandcdfanalysis " that M.AsadifardandJ.Shanbezadeh delivers on ProceedingsoftheInternationalmultiConferenceofngineersan dComputerScientists utilizes active shape model (ASM) by face feature point modeling as a whole, utilizes unique point relational implementation location each other.Recent years, along with machine learning algorithm is in the increase of numerous areas depth & wideth, its application in human eye location also gets more and more.As:.According to Haar feature and the performance outstanding in Face datection of Adaboost sorter, Y.Maetal is at AutomaticFaceandGestureRecognition, " Robustpreciseeyelocationunderprobabilisticframework " that 2004.Proceedings.SixthIEEEInternationalConference delivers utilizes similar method human eye location, by the multiple simple sorter of cascade, mistake in sorter before emphasizing constantly trains new sorter, is finally combined into strong classifier.Although these method relative maturity, respectively have no advantages and disadvantages, shortcoming mainly concentrates on, for the method that complexity is not high, its degree of accuracy can not breakthrough bottleneck, and can meet the requirements of method for degree of accuracy, and the too high real-time that makes again of its complexity does not reach requirement.A lot of scenes of human eye location all have higher requirements to real-time and accuracy, and this just needs to find a kind of method that can all do well both ways.

Summary of the invention

In order to solve as above problem, the invention provides a kind of quick position human eye method in facial image.

Technical solution of the present invention is as follows:

A method for quick position human eye in facial image, its feature is, comprises the steps:

Step 1, based on random forest machine learning algorithm, training obtain some decision trees, form random forest by integrated technology;

Step 2, inputting facial image to be measured, and be converted into gray-scale map, is namely two-dimensional matrix;

Step 3, the random forest obtained utilized in step 1, step 2 is obtained to the multistage location structure of use of the two-dimensional matrix of gray-scale value, namely from a fixing human eye coarse positioning region, according to current positioning result, as the center of locating area next time, reduce hunting zone step by step, continuous iteration to the last one-level complete after restoring to normal position result, determine the position coordinates of human eye;

Step 4, will in step 3, there be many decision trees in random forest to obtain multiple people's eye coordinates weighted mean, obtain the result of final human eye location.

Training process in described step 1 based on random forest machine learning algorithm, concrete steps are as follows:

1.1), the face sample gray level image of input is done normalized, the coordinate in the upper left corner of picture, the upper right corner, the lower right corner is respectively (0,0), (1,0), (1,1).

1.2), as " seed " of random forest, variation process is done to sample, on the basis of master sample, carry out randomization, on horizontal ordinate, ordinate direction and dimension of picture, do the random offset in certain limit respectively, in certain angle, do random rotation simultaneously, generate the training sample of multi-pose;

1.3), by 1.2) two-dimensional matrix that obtains, get its most direct eigenwert, the gray scale coordinate figure of pixel in image, as input, the coordinate of human eye is as output, and decision tree is a full binary tree, i.e. grey scale difference I (I between root node and each intermediate node are trained and to be preserved based on difference coordinate ₁)-I (I ₂) and a threshold value T, our definition is tested (binarytest) to the two-value of image I and is:

$b\mathrm{int}est(I;I(I_1),I(I_2),T)=\left\{\begin{array}{cc}0,& I\left(I_1\right)-I\left(I_2\right)<T\\ 1,& otherwise\end{array}\right.$

Preferential, this binary tree is decision tree, and we promote decision Tree algorithms by self-adaption gradient and obtain, and concrete steps are as follows:

First, initialization error of fitting function:

$F_0\left(x\right)={\mathrm{argmin\&Sigma;}}_{i=1}^{N}L(y_{i,},F(x_i\left)\right)$

Wherein, F (x) is decision tree function, for error of fitting;

Further, calculate to the pseudo-residual error of negative gradient direction calculating, namely

$\widetilde{y_i}=-{\&lsqb;\&part;L(y_i,F(x_i\left)\right)/\&part;F\left(x_i\right)\&rsqb;}_{F\left(x\right)=F_{m-1}\left(x\right)},i=1...N$

Wherein, we define loss function and are,

L(y _i,F(x _i))＝1/2(y _i-F(x _i)) ²

Substitution obtains,

$\widetilde{y_i}=y_i-F\left(x_i\right)$

Further, the pseudo-residual error of matching is upgraded:

${\mathrm{\&alpha;}}_m=\arg {\min }_{\mathrm{\&alpha;},\mathrm{\&beta;}}{\mathrm{\&Sigma;}}_{i=1}^N\left|\right|\widetilde{y_i}-\mathrm{\&rho;}h(x_i;\mathrm{\&alpha;})|{|}^2$

Wherein, h (x _i; α) be fitting result, such as primary fitting result is h (x _i; α ₁)

Further, weight and regression criterion multiplier is upgraded:

${\mathrm{\&omega;}}_i=\left|\right|y_i-F_M\left(x_i\right)|{|}^2/\left({\mathrm{\&Sigma;}}_{i=1}^N\right||y_i-F_M(x_i\left)\right|{|}^2)$

${\mathrm{\&rho;}}_m=\arg min\underset{i=1}{\overset{N}{\&Sigma;}}L(y_i,F_{m-1}(x_i)+\mathrm{\&rho;}h(x_i;{\mathrm{\&alpha;}}_m\left)\right)$

Further, Renewal model, m=1 → M, after M time, iteration terminates:

F _m(x)＝F _m-1(x)+γρ _mh(x；α _m)

Wherein, 1< γ≤0, namely γ is learning rate, and the size of γ determines the speed of convergence of iteration.

1.4), use 1.3) method, n pictures is divided into two classes, is mounted in two, the left and right child node of root node respectively, at ground floor subsequently, the test of with good grounds two-value is divided into four class pictures altogether, by that analogy, cut-off when arriving that suitably tree is dark;

1.5), for 1.4) in classifying quality, simultaneously also can with the aggregation extent estimated, namely the quadratic sum of error characterizes, and the object of therefore training is for minimizing following error:

$Q_{node}=\underset{p\&Element;S_l}{\mathrm{\&Sigma;}}\left|\right|p-{\mathrm{\&Sigma;}}_{q\&Element;S_l}q/\left|\right|S_l\left|\right||{|}^2+\underset{p\&Element;S_r}{\&Sigma;}||p-{\mathrm{\&Sigma;}}_{q\&Element;S_r}q/|\left|S_l\right|\left|\right|{|}^2$

Wherein S _land S _rbe respectively the set of coordinate composition in the left and right child node produced according to certain characteristic sum threshold classification.The condition that training stops is when arriving the degree of depth of specifying tree or specification error size.

Coordinate average of same class sample owing to needing to locate predicting the outcome of picture, so be an estimation to predicting the outcome, so this aggregation extent estimated is more, the confidence level carrying out predicting is also higher, proves that the effect of training is better, wherein, the aggregation extent estimated can represent by variance, but consider the quantity needing the picture ensureing carry on each node, namely ensure the average of classification, therefore the quadratic sum of use error is the most suitable as far as possible.Therefore the target that we obtain training minimizes following error:

$Q_{tree}=\underset{i}{\&Sigma;}\left\{\underset{p\&Element;S_l}{\&Sigma;}\left|\right|p-\frac{{\mathrm{\&Sigma;}}_{q\&Element;S_l}q}{\left|\right|S_1\left|\right|}|{|}^2\right\}$

Wherein, i=1,2 ..., 2 ^d, represent the i-th class, correspond to i-th leaf node, 2 ^dfor the sum of leaf node, S _irepresenting the set that the i-th class is corresponding, is exactly the set of the picture at i-th leaf node.

But in actual applications, binary tree is a nonlinear system inherently, consider that all influence factors are also very difficult and complicated simultaneously.Therefore decision tree training process herein adopts greedy method, determines characteristic parameter on each node and threshold value thereof successively, make formula (1) little from root node according to the degree of depth.Note now training sample not on leaf node, but along with the propelling of the degree of depth, we all using present node as leaf node, still use Q _treerepresent center estimated by all samples Euclidean distance quadratic sum to generic.Due to each only training node, so only have the training sample of this node to change.So when training each node, need travel through and select best feature and threshold value, two classes formed after making to be classified have minimum variance, and therefore objective function becomes:

$Q_{tree}=\underset{p\&Element;S_l}{\&Sigma;}\left|\right|p-\frac{{\mathrm{\&Sigma;}}_{q\&Element;S_l}q}{\left|\right|S_l\left|\right|}|{|}^2+\underset{p\&Element;S_r}{\&Sigma;}||p-\frac{{\mathrm{\&Sigma;}}_{q\&Element;S_r}q}{\left|\right|S_r\left|\right|}|{|}^2$

Wherein, S _lwith S _rthe set of the left and right child node represented respectively.

Compared with prior art, the invention has the beneficial effects as follows:

1) the present invention improves the precision of location further, and the low complexity algorithm based on binary tree makes time loss reduce, and the present invention is based on CART model and from a large amount of diversified human eye sample, carries out study obtain regression tree, the good lifting that robustness also obtains.

2) for the situation that single regression tree model station-keeping ability is more weak, multiple weak fix devices that repetition learning is obtained by random forest and these two kinds of integrated technologies of gradient lifting decision tree by the present invention are combined into a strong fix device.

3) according to the applied environment of human eye location, decision tree of the present invention constantly reduces the scope of locating area in position fixing process, and the sample weights introducing positioning result is to improve training process and the forecasting process that random forest and gradient promote decision tree further, finally achieves significant performance gain.

4) the human eye location based on CART model can when only there being low-resolution image with extremely low computing cost, realize the positioning precision and the robustness that are better than all the other models most, therefore the storage space that required cost is just extra is on a small quantity a well selection for fast human-eye location.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the quick position human eye method in facial image of the present invention;

Fig. 2 multi-pose training sample generates sample;

Fig. 3 is the process that gradient promotes that decision tree uses integrated technology human eye location;

The multistage location structure design sketch of Fig. 4.

Fig. 5 mono-decision tree structure.

Embodiment

The measure realized to make the technology of the present invention, creation characteristic, reaching object and effect is easy to understand, below in conjunction with concrete diagram, setting forth the present invention further.

Fig. 1 is the process flow diagram of the method for quick position human eye in facial image of the present invention, comprises the steps:

Step one, based on random forest machine learning algorithm, by training obtain 25 tree be deeply 11 decision tree;

First, as shown in Figure 2, do normalization and variation process to sample, the coordinate in the upper left corner of picture, the upper right corner, the lower right corner is respectively (0,0), (1,0), (1,1), on the basis of master sample, carry out randomization, on horizontal ordinate, ordinate direction and dimension of picture, do the random offset in certain limit respectively, in certain angle, do random rotation simultaneously, generate the training sample of multi-pose;

Promote decision tree by gradient, as shown in Figure 3, gradient during γ=0.4 promotes decision tree, constantly revises the direction of advance thus arrives human eye center more accurately, obtaining better effect.

Obvious single decision tree must be a Weak Classifier, uses integrated technology by these tree composition random forests, forms strong classifier.

Step 2, input face picture, and be converted into gray-scale value two bit matrix;

Step 3, as shown in Figure 5, utilize each decision-tree model in step one, decision tree is a binary tree, and be generally a full binary tree, from the root node of binary tree, each nonleaf node is asked a training in advance and the problem kept, select to proceed to the left son of root node or right son, until proceed to the leaf node of decision tree according to the answer result of this problem.Obtain altogether 25 people's eye coordinateses.

Use 5 grades of location structures, determine the position coordinates of human eye.As shown in Figure 2, the present invention is in order to improve accuracy rate, have employed multistage location structure, also referred to as pyramid structure, precision is improved: from a fixing larger ROI (RegionofInterest) by the scope constantly reducing location, according to the regional center of current ROI positioning result as the less ROI of next stage, then constantly iteration until return the positioning result of afterbody ROI.

Step 4,25 people's eye coordinateses step 3 obtained are averaged, and obtain the result of final human eye location.

Claims (4)

1. the method for quick position human eye in facial image, is characterized in that, comprise the steps:

Step 1, based on random forest machine learning algorithm, training obtain some decision trees, form random forest by integrated technology;

Step 2, inputting facial image to be measured, and be converted into gray-scale map, is namely two-dimensional matrix;

Step 3, the random forest obtained utilized in step 1, step 2 is obtained to the multistage location structure of use of the two-dimensional matrix of gray-scale value, namely from a fixing human eye coarse positioning region, according to current positioning result, as the center of locating area next time, reduce hunting zone step by step, continuous iteration to the last one-level complete after restoring to normal position result, determine the position coordinates of human eye;

Step 4, will in step 3, there be many decision trees in random forest to obtain multiple people's eye coordinates weighted mean, obtain the result of final human eye location.

2. the method for quick position human eye in a kind of facial image according to claim 1, it is characterized in that, the training process in described step 1 based on random forest machine learning algorithm, concrete steps are as follows:

Step 1.1) the face sample gray level image of input is done normalized, obtain standard exercise sample, the coordinate in the upper left corner of face sample gray level image, the upper right corner, the lower right corner is respectively (0,0), (1,0), (1,1);

Step 1.2) variation process is done to standard exercise sample: on horizontal ordinate, ordinate direction and dimension of picture, do the random offset in certain limit respectively, in certain angle, do random rotation simultaneously, generate the training sample of multi-pose;

Step 1.3) according to the two-dimensional matrix of the training sample of multi-pose, get its most direct eigenwert, by the gray scale coordinate figure of pixel in image, as input, the coordinate of human eye is as output, obtain binary tree, i.e. grey scale difference I (I between root node and each intermediate node are trained and to be preserved based on difference coordinate ₁)-I (I ₂) an and threshold value T, define and (binarytest) is tested to the two-value of image I be:

$b\mathrm{int}est(I;I(I_1),I(I_2),T)=\left\{\begin{array}{cc}0,& I\left(I_1\right)-I\left(I_2\right)<T\\ 1,& otherwise\end{array}\right.$

Step 1.4) according to binary tree, the n of input is opened multi-pose training sample and is divided into two classes, be mounted in two, the left and right child node of root node respectively, at ground floor subsequently, with good grounds two-value test is divided into four class pictures altogether, by that analogy, arrive specify tree dark time or specification error size time end.

3. the method for quick position human eye in a kind of facial image according to claim 2, it is characterized in that, the training process in described step 1 based on random forest machine learning algorithm, also comprises:

Step 1.5) to step 1.4) training result that obtains carries out error validity, determine whether minimum error, formula is as follows:

$Q_{node}=\underset{p\&Element;S_l}{\mathrm{\&Sigma;}}\left|\right|p-{\mathrm{\&Sigma;}}_{q\&Element;S_l}q/\left|\right|S_l\left|\right||{|}^2+\underset{p\&Element;S_r}{\&Sigma;}||p-{\mathrm{\&Sigma;}}_{q\&Element;S_r}q/|\left|S_l\right|\left|\right|{|}^2$

Q in formula _nodebe exactly least error, represent center estimated by all samples Euclidean distance quadratic sum to generic, S _land S _rbe respectively the set being divided into coordinate composition in the left and right child node of two classes produced according to certain characteristic sum threshold classification.

4. the method for quick position human eye in a kind of facial image according to claim 2, is characterized in that, described step 1.3) concrete steps that obtain binary tree are as follows:

Step 1.31) initialization error of fitting function, formula is as follows:

$F_0\left(x\right)={\mathrm{argmin\&Sigma;}}_{i=1}^{N}L(y_{i,},F(x_i\left)\right)$

Wherein, F (x) is decision tree function, for error of fitting;

Step 1.32) calculate to the pseudo-residual error of negative gradient direction calculating, formula is as follows:

$\widetilde{y_i}=y_i-F\left(x_i\right)$

Step 1.33) upgrade the pseudo-residual error of matching:

${\mathrm{\&alpha;}}_m=\arg {\min }_{\mathrm{\&alpha;},\mathrm{\&beta;}}{\mathrm{\&Sigma;}}_{i=1}^N\left|\right|\widetilde{y_i}-\mathrm{\&rho;}h(x_i;\mathrm{\&alpha;})|{|}^2$

Wherein, h (x _i; α) be fitting result, primary fitting result is h (x _i; α ₁

Step 1.34) upgrade weight and regression criterion multiplier:

${\mathrm{\&omega;}}_i=\left|\right|y_i-F_M\left(x_i\right)|{|}^2/\left({\mathrm{\&Sigma;}}_{i=1}^N\right||y_i-F_M(x_i\left)\right|{|}^2)$

${\mathrm{\&rho;}}_m=\arg min\underset{i=1}{\overset{N}{\&Sigma;}}L(y_i,F_{m-1}(x_i)+\mathrm{\&rho;}h(x_i;{\mathrm{\&alpha;}}_m\left)\right)$

Step 1.34) Renewal model, m=1 → M, after M time, iteration terminates:

F _m(x)＝F _m-1(x)+γρ _mh(x；α _m)

Wherein, 1< γ≤0, namely γ is learning rate, and the size of γ determines the speed of convergence of iteration.