CN106096560A - A kind of face alignment method - Google Patents

Abstract

The open a kind of face alignment method of the present invention, including: collect a number of facial image as training sample and forecast sample, by image enhaucament, for training sample, demarcate face key point, and preserve key point positional information；By random forests algorithm, the training sample demarcated in S1 is learnt, obtain the Feature Mapping function of the key point demarcatedThus obtain the local binary feature of the key point demarcated；One overall binary feature of the composition that the local binary feature of the demarcation key point obtained in step S2 combined, utilizes this feature to obtain an overall linear regression model (LRM) W with the learning style of overall situation linear regression^t, and then realize the location of the face key point of sample to be tested.

Description

A kind of face alignment method

Technical field

The present invention relates to field of face identification, be specifically related to a kind of face alignment method.

Background technology

Face recognition technology is face feature based on people, facial image or the video flowing to input, it is judged that whether it There is face, if there is face, the most further provide the position of each face, size and the position of each major facial organ Information, and according to these information, extract the identity characteristic contained in each face further, and it is entered with known face Row contrast, thus identify the identity of each face.Whole process generally comprise Face datection, facial pretreatment, face alignment, with And the step such as recognition of face.Face recognition algorithms mainly have algorithm based on human face characteristic point, recognizer based on template with And recognizer of based on view picture face figure.

Face alignment is that facial image previous step detected all snaps on one group of datum mark, and being actually will Each datum mark (eyes, nose, the interior tail of the eye, face) of the face plucked out all snaps in one group of normal place.It is made For step most crucial in algorithm based on human face characteristic point, the feature location of mistake can cause the gross distortion of face characteristic, Even coarse alignment also can make recognition effect be greatly reduced.Therefore, it is possible to it is crucial special to extract face quickly and accurately Levy algorithm a little extremely important to the raising of follow-up discrimination.

Face alignment algorithm mainly has based on Statistical learning model Feature Points Extraction at present, wherein has active shape mould Type and actively presentation model.The extraction of active shape model ASM (active shape models), greatly effectively raises The detection of face characteristic key point, thus improve face identification rate to a certain extent.Its concrete scheme includes:

1. the foundation of statistical model

Use the artificial method demarcated, demarcate a number of characteristic point that can fully describe and characterize objective contour.Cause This, every facial image may be expressed as the two-dimensional coordinate dot information of N number of fixed point.

2. shape Statistics model

Fixed point quantity obtained above is very big, and its feature point set vector dimension formed is the highest, has between its key point Certain dependency, positional distance each other is all substantially constant, it is therefore desirable to carries out dimensionality reduction with PCA, extracts main one-tenth Point.According to the characteristic vector chosen in reduction process and characteristic of correspondence value, a linear variable Statistical Shape can be set up Model, expression formula is:

$x\≈\frac{-}{X}+pb$

Wherein P is the transition matrix that main constituent characteristic vector is constituted, and b is form parameter vector, and this form parameter represents Proportion shared by each shape components in model.Because this system is the model of a linear variable, and need to be limited in necessarily Variable range in, therefore shape vector b is often restricted to:

$-3\sqrt{{\mathrm{\λ}}_i}\≤b_i\≤3\sqrt{{\mathrm{\λ}}_i}$

3. model and new point set mate

First, calculate eyes, the position of face, do simple yardstick and rotationally-varying, wherein calculate eyes or face position The formula put is:

$X=T_{X_t,Y_t,S,\mathrm{\θ}}(\stackrel{\‾}{X}+Pb)$

Wherein T is for rotating scaling translation matrix

$T_{X_t,Y_t,S,\mathrm{\θ}\left(\begin{array}{c}x\\ y\end{array}\right)}=\left(\begin{array}{c}{X}_t\\ Y_t\end{array}\right)=\left(\begin{array}{cc}scos\mathrm{\θ}& -ssin\mathrm{\θ}\\ ssin\mathrm{\θ}& s\cos \mathrm{\θ}\end{array}\right)\left(\begin{array}{c}x\\ y\end{array}\right)$

Closest further according to model X and characteristics of image point set Y so that it is object function and each point Euclidean distance are minimum, target Function representation formula is:

${|Y-T_{X_t,Y_t,S,\mathrm{\θ}}(\stackrel{\‾}{X}+Pb)|}^2$

4. critical point detection

The face substantially target location detected with human-face detector and size and direction.Adjust above-mentioned 5 parameters, pass through Prior information characteristic point, it can be assumed that characteristic point is mainly the strong marginal point of image, and the gradient of gradation of image is obeyed Gauss and is divided Cloth, then can find the maximum of gradient near characteristic point by model, i.e. it is believed that be characterized a position.For each specific The candidate point of position, calculates the Euclidean distance between this location point in the partial model of each candidate point and training pattern respectively, Formula is:

F(b,X_t,Y_t, s, θ) and=| X'-X |²

When positioning key point coordinate on the basis of facial image, the accuracy of key point position depends on the manual mark of face Determining the accuracy of key point and the quantity of sample, the algorithm effect of ASM (active shape model) this location key point is one Determine can effectively position in degree key point, but this method does not accounts for the gray scale letter of contoured interior target object Breath, therefore process imaging surface half-tone information than more rich object (such as face) or image by illumination effect bigger time, Just seem unable to do what one wishes.Image is carried out needing substantial amounts of sample set to facial image normalization and right during pretreatment by this method Training set data uses PCA (principal component analysis) method, and process is loaded down with trivial details, the most long, and locating human face's key point is less accurate, Relatively rough, do not possess practicality, the most time-consumingly, it is impossible to meet real-time requirement.

Summary of the invention

Present invention aim to address the defect of prior art, it is provided that a kind of face alignment method more quickly and effectively, real Existing face key point is accurately positioned, and is greatly shortened positioning time, and the technical scheme of employing is as follows:

A kind of face alignment method, comprises the following steps:

S1. collect a number of facial image as training sample and forecast sample, for training sample, image is increased By force, demarcate face key point, and preserve key point positional information；

S2. the feature learning to obtain demarcating key point to the training sample demarcated in S1 by random forests algorithm is reflected Penetrate functionAnd then obtain the local binarization feature demarcating key point；

S3. one overall two-value of the composition that the local binary feature of the demarcation key point obtained in step S2 combined is special Levy；

S4. the overall binary feature in step S3 is utilized to obtain overall linear regression with the learning style of overall situation linear regression Model W^t, and then realize the location of the face key point of sample to be tested.

The present invention uses local binary feature (LBF, the Local Binary of random forest study face characteristic key point Features), and learn the local criterion of these features, utilize this local criterion to independently be everyone as guidance Face labelling learns the local binary feature of one group of high identification, these local binary features are linked up as feature to Amount.This method can obtain good effect on facial image, simultaneously because extract and return local binary feature calculation Measuring the lowest, therefore the present invention is effectively increased the speed of Face datection, more a lot of than the face alignment method of prior art.

Further, image enhaucament being included, image carries out deblurring, mist elimination, Geometry rectification and contrast to be strengthened.

Further, described step S4 particularly as follows:

One overall portion binary feature of composition will be connected before and after the local binary feature of all random trees, utilize this overall situation Portion's binary feature does overall situation linear regression, deformation is given a forecast as prediction target, training overall situation linear regressor, and it is expressed Formula is:

$W^t=\underset{W^t}{\mathrm{argmin}}\underset{i=1}{\overset{N}{\Σ}}{||\mathrm{\Δ}S-W^t\·lbf||}_2^2+\mathrm{\λ}{||W^t||}_2^2$

Wherein, Δ s represents that deformation, lbf represent LBF feature, w^tBeing the parameter of linear regression, λ is inhibition parameter, shape Become predictor formula into:

Δ s=w^t.lbf。

Every one-level of multi-stage cascade homing method can be divided into two parts by as above say, utilizes random forest to carry Taking local binary feature, then recycling local binary feature does overall situation linear regression prediction shape increment Delta S.

As preferably, described step S2 includes:

Picture initializes 1 shape, and wherein shape is the artificial key point demarcated, i.e. eyes nose face position letter Breath；

Calculate the surrounding pixel of each point, or the difference of pixel in corresponding point in two shapes based on this shape, make It is characterized value features；

Calculate current face's picture and discrepancy delta s of handmarking's shape, then one function y=f (x) of training, calculate Δ s=f (features)

In the present invention, random forest training method is directed to single key point, the characteristic point used in all random trees Will not be associated with in other key points.Two characteristic points are randomly selected in the appointment radius region centered by current key point, Then the pixel value difference of calculating the two point is as feature, and referred to as random tree feature, then by the random forest of this key point All random tree characteristic bindings get up to obtain the local binary feature (LBF) of this key point, finally by corresponding for all key points The local binary feature of random forest output is connected with each other and constitutes an overall binary feature, then makes of overall situation binary feature The overall situation returns, and is used for predicting key point.

Compared with prior art, beneficial effects of the present invention:

The present invention proposes random forest and characteristic point is positioned by the overall situation method that combines of linear regression, closes face Key spot placement accuracy is high, and the face picture detected can carry out key point prediction location more in real time.In face aligns The LBF feature used, through the training of random tree one-level level, it is possible to the shape and structure of memory face calibration, thus prediction people Face key point can be accurately positioned, and two key point margin of image element of LBF characteristic use, accelerate training and predetermined speed.Separately Outward, overall situation recurrence learning can carry out global shape constraint effectively, and reduces owing to the smudgy of local appearance is carried The error come, it is possible to be effectively used in patrol robot face identification system, improves face identification rate.

Accompanying drawing explanation

Fig. 1 is the schematic diagram that the present invention demarcates face key point；

Fig. 2 is the face alignment method schematic flow sheet of the present invention；

Fig. 3 is the structural representation of random forest of the present invention；

Fig. 4 is to use traditional ASM to carry out the prediction of face key point to position and use the present invention to predict that face key point is fixed Position Contrast on effect schematic diagram；

Fig. 5 is the experimental result schematic diagram of the present invention.

Detailed description of the invention

With embodiment, the present invention is described in further detail below in conjunction with the accompanying drawings.

Embodiment:

As it is shown in figure 1, a kind of face alignment method, comprise the following steps:

S1. collect a number of facial image as training sample and forecast sample, for training sample, image is increased By force, demarcate face key point, and preserve key point positional information, as shown in Figure 1；

S2. by random forests algorithm, the training sample demarcated in S1 is learnt, the key point feature that study is demarcated Mapping functionSecondary series as shown in Figure 2；

S3. the Feature Mapping function obtained in step S2 is utilizedObtain the local binary feature of key point；From Fig. 2 the 3rd Row can be seen that, this study based on local binary feature (LBF) be by " locally " principle make study well-regulated enter OK.This principle mainly has two aspects: for one landmark point determined in location in one-level, 1) texture of most identification Information distribution around the landmark point that upper level estimates, 2) information content of shape and the local grain of this landmark point carry Supply the information of abundance.Therefore we can be independent learns the feature of most intuition type for each landmark point and encodes local Textural characteristics, performs overall situation linear regression the most again and goes to merge the information content of shape, it is achieved the location of face alignment key point；

S4. overall situation linear regression method is utilized to realize the location of face key point of sample to be tested.

In the present embodiment, image enhaucament is included, and image carries out deblurring, mist elimination, Geometry rectification and contrast to be strengthened.

Described step S4 particularly as follows:

By connecting one overall binary feature of composition before and after the local binary feature of all random trees, utilize this overall situation two Value tag does linear regression, deformation is given a forecast as prediction target, training linear regressor, and wherein, linear regression is expressed Formula is:

$W^t=\underset{W^t}{\mathrm{argmin}}\underset{i=1}{\overset{N}{\Σ}}{||\mathrm{\Δ}S-W^t\·lbf||}_2^2+\mathrm{\λ}{||W^t||}_2^2$

Wherein, Δ s represents that deformation, lbf represent LBF feature, w^tBeing the parameter of linear regression, λ is inhibition parameter, shape Become predictor formula into:

Δ s=w^t.lbf。

Every one-level of multi-stage cascade homing method can be divided into two parts by as above say, utilizes random forest to carry Taking local binary feature, then recycling local binary feature does overall situation linear regression prediction shape increment Delta S.

Described step S2 includes:

Picture initializes 1 shape, and wherein shape is the artificial key point demarcated, i.e. eyes nose face position letter Breath；

Calculate the surrounding pixel of each point, or the difference of pixel in corresponding point in two shapes based on this shape, make It is characterized value features；

Calculate current face's picture and discrepancy delta s of handmarking's shape, then one function y=f (x) of training, calculate Δ s=f (features).

Random forest is a kind of well Multiple Classifier Fusion algorithm, it is possible to well solving multicategory classification problem, it is basic Thought is that many Weak Classifiers are integrated into a strong classifier, as it is shown on figure 3, a random forest is made up of N decision tree, often Decision tree is a grader.

In the present embodiment, random forest training method is directed to single key point, the feature used in all random trees Point will not be associated with in other key points.Two features are randomly selected in the appointment radius region centered by current key point Point, then the pixel value difference of calculating the two point is as feature, and referred to as random tree feature, then by the random forest of this key point All random tree characteristic bindings get up to obtain the local binary feature (LBF) of this key point, finally that all key points are corresponding Random forest output local binary feature be connected with each other composition one overall binary feature.Again with overall situation binary feature Do overall situation recurrence, be used for predicting key point.

In the present embodiment, the facial image sum C:12186 of collection, wherein, training sample 17160, it was predicted that sample 3436.Demarcate 13 key points, respectively left eye eyeball, right eye eyeball, left eye canthus, right eye canthus, nose both sides, nose Point, the up and down corners of the mouth.As shown in Figures 4 and 5, the image on Fig. 4 left side is to use ASM to carry out key point to record in advance to experimental result The result arrived, right image is to use the method for the present invention to carry out key point to predict the result obtained, from fig. 4, it can be seen that this Invention effect is better than ASM.From fig. 5, it can be seen that the present invention can orient the key point of face faster, recognition of face can be met The requirement that system is real-time.

Claims (4)

1. a face alignment method, it is characterised in that comprise the following steps:

S1. a number of facial image is collected as training sample and forecast sample, for training sample, by image enhaucament, Demarcate face key point, and preserve key point positional information；

S2. by random forests algorithm, the training sample demarcated in S1 is learnt the Feature Mapping letter obtaining demarcating key point NumberAnd then obtain the local binarization feature demarcating key point；

S3. one overall binary feature of the composition that the local binary feature of the demarcation key point obtained in step S2 combined；

S4. the overall binary feature in step S3 is utilized to obtain overall linear regression model (LRM) with the learning style of overall situation linear regression W^t, and then realize the location of the face key point of sample to be tested.

A kind of face alignment method the most according to claim 1, it is characterised in that include image enhaucament image is carried out Deblurring, mist elimination, Geometry rectification and contrast strengthen.

A kind of face alignment method the most according to claim 1, it is characterised in that described step S2 includes:

Picture initializes 1 shape, and wherein shape is the artificial key point demarcated；

The surrounding pixel of each point, or the difference of pixel in two corresponding point of two shapes is calculated based on this shape, as Eigenvalue features；

Calculate current face's picture and discrepancy delta s of handmarking's shape, training function y=f (x), calculate Δ s=f (features)。

A kind of face alignment method the most according to claim 1, it is characterised in that described step S4 particularly as follows:

One overall binary feature of composition will be connected before and after the local binary feature of all random trees, utilize this overall situation two-value special Levying and do overall situation linear regression, deformation given a forecast as prediction target, training overall situation linear regressor, its expression formula is:

$W^t=\underset{W^t}{\mathrm{argmin}}\underset{i=1}{\overset{N}{\Σ}}\left|\right|\mathrm{\Δ}S-W^t\·lbf|{|}_2^2+\mathrm{\λ}|\left|W^t\right|{|}_2^2$

Wherein, Δ s represents that deformation, lbf represent LBF feature, w^tBeing the parameter of linear regression, λ is inhibition parameter, and deformation is pre- Survey formula is:

Δ s=w^t.lbf。