CN103400122A - Method for recognizing faces of living bodies rapidly - Google Patents

Abstract

The invention discloses a method for recognizing faces of living bodies rapidly. The method comprises the steps as follows: a), a plurality of continuous facial images are input; b), pupil positions are determined for each facial image, and a pupil zone is cut; and c), an eyeball opening or closing state is judged, and if an eye twinkling process exists, activity discrimination is performed. In the step a), according to an input video flow, the plurality of the continuous facial images are acquired, and if two adjacent facial images are in the same state, the images are discarded and a plurality of facial images are acquired again; and in the step c), eye-opening samples and eye-closing samples are trained with a support vector machine training method and an AdaBoost training method. According to the method for recognizing the faces of the living bodies rapidly, in a recognition process, the activity discrimination is performed through simple eye-twinkling action recognition, so that a user is not required to make a matching action when a face database is established, the matching action of the user is greatly simplified, and the recognition efficiency is improved.

Description

A kind of method for quickly identifying of living body faces

Technical field

The present invention relates to a kind of biometric discrimination method, relate in particular to a kind of method for quickly identifying of living body faces, belong to image processing, pattern-recognition and computer vision field.

Background technology

Biometrics identification technology take people's face as feature development in recent years is very rapid, and some gyp practical face identification systems have appeared on market.Safety problem has also obtained the very large attention of people in the practical process of face identification system.Active discrimination, namely how to distinguish the photo of validated user and it is lived I, most important to the safety of face identification system.

At present the active method of differentiating mainly contains two kinds from whether needing the user to coordinate to distinguish, and a kind of user of needs coordinates, and needs the user to do the variation of expression or attitude.Also has a kind of cooperation that does not need the user.All only has first method at present seen system.Existing system General Requirements user blinks and smiles in the process of identification action cooperation.Perhaps require the user to record when setting up face database in short, the moving information of user's lip and acoustic information join in storehouse simultaneously, in identification, need the user to say again with a word, whether system is analyzed the information of sound and two passages of video simultaneously, thereby draw, be he or she's conclusion.How simplifying user's action cooperation and improving recognition efficiency is the technical research focus of this area always.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind of method for quickly identifying of living body faces, does not need the user to do interoperation when setting up face database, and the action that can simplify the user coordinates and the raising recognition efficiency.

The present invention solves the problems of the technologies described above the technical scheme that adopts to be to provide a kind of method for quickly identifying of living body faces, comprises the steps: a) to input several continuous facial images; B) every width facial image is determined pupil position and determined human eye area; C) the judgement eyeball is opened the state of closing, if there is process by activity, differentiate nictation.

The method for quickly identifying of above-mentioned living body faces, wherein, described step a) is obtained several continuous facial images according to the video flowing of input,, if adjacent two width facial images are not abandoned for same state, again obtains several continuous facial images.

The method for quickly identifying of above-mentioned living body faces, wherein, described step b) comprise following process: b1) initialisation image size and video frame rate parameter; B2) determine the size of current eye areas; B3) the current scanning window of location 40 pixel x20 pixels; B4) eliminate overlapping eyeball window.

The method for quickly identifying of above-mentioned living body faces, wherein, described step c) by training method of support vector machine and AdaBoost training method to opening eyes and close one's eyes the sample training.

The method for quickly identifying of above-mentioned living body faces, wherein, described step c) the sample training algorithm of opening eyes and close one's eyes in is as follows:

C1): input training sample set (x ₁, y ₁) ..., (x _n, y _n), x _i∈ X, y _i∈ Y={0,1}; Initializes weights, establish positive example sample and negative data and be respectively m, and 1, $w_{1,i}=\left\{\begin{array}{cc}\frac{1}{2m}& y_i=-1\\ \frac{1}{2l}& y_i=0\end{array}\right.;$

C2): the normalization weight,

Make w _tBe a probability distribution, t=1 ..., T, T are the feature sum of choosing;

C3): to Weak Classifier h of each feature j training _jAnd calculating has added the error of weight:

ε _j＝∑ _iw _i|h _j(x _i)-y _i|；

C4): Select Error ε _iMinimum Weak Classifier h _t

C5): upgrade weight If x _iBy correct Klassifizierung _i=0, otherwise e _i=1; Wherein, ${\mathrm{\β}}_i=\frac{{\mathrm{\ϵ}}_i}{1-{\mathrm{\ϵ}}_i};$

C6): the sorter that finally obtains is:

$h\left(x\right)=\mathrm{sign}(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_i{h}_t\left(x\right)-\frac{1}{2}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t);$ Wherein, ${\mathrm{\α}}_i=\log \frac{1}{{\mathrm{\β}}_i}.$

The method for quickly identifying of above-mentioned living body faces, wherein, described step continuous facial image a) is 15～25 two field pictures adjacent in input video stream.

The present invention contrasts prior art following beneficial effect: the method for quickly identifying of living body faces provided by the invention, carry out activity judgment by the action recognition of simply blinking in identifying, make and do not need the user to do interoperation when setting up face database, thereby greatly simplify user's action cooperation and improve recognition efficiency.

Description of drawings

Fig. 1 is the quick identification process schematic diagram of living body faces of the present invention;

To be the present invention carry out eyeball for every two field picture to Fig. 2 opens and close the differentiation schematic flow sheet;

Fig. 3 is for adopting simple rectangular characteristic to characterize people's face schematic diagram;

Fig. 4 is the integrogram schematic diagram based on the rectangular area feature;

Fig. 5 be in rectangle D brightness value a little and quote the calculating schematic diagram.

Embodiment

The invention will be further described below in conjunction with drawings and Examples.

Fig. 1 is the quick identification process schematic diagram of living body faces of the present invention.

See also Fig. 1, the method for quickly identifying of living body faces provided by the invention comprises the steps:

A) input several continuous facial images; In order to eliminate the unexpected variation of illumination, people's face moves suddenly and waits interference, can introduce level and smooth concept in practice, namely according to the video flowing of input, obtain the adjacent facial image of adjacent 15～25 frames, be preferably 20 frames,, if adjacent two width facial images are not abandoned for same state, again obtain several continuous facial images;

B) every width facial image is determined pupil position and determined human eye area, as shown in Figure 2, comprising: initialisation image size and video frame rate parameter; Determine the size of current eye areas; Locate the current scanning window of 40 pixel x20 pixels; And eliminate overlapping eyeball window;

C) the judgement eyeball is opened the state of closing, if there is process by activity, differentiate nictation; Concrete according to the state from the 1st frame to i frame human eye (S1, S2 ..., Si) judge whether to have blinked,, if blinked, by activity, differentiate; If to the N frame all do not have the nictation, output failure information, system turns back to original state.

The method for quickly identifying of living body faces provided by the invention, sample training by a large amount of people's eyeballs is based on supporting the eyeball motility model that mixes with AdaBoost to machine, then in activity differentiation process, by the cooperation that the identified person blinks, complete the active process of differentiating.Can not make the interoperation of nictation due to human face photo, so photo just by the eliminating of success in the outside of face identification system, thereby improved the security of automatic human face recognition system, particularly at the scene camera collection to people's face and the standard photographs masterplate such as Certification of Second Generation to compare in situation practical value very high.

The present invention proposes a kind of based on supporting to machine and the eye movement model that mixes with AdaBoost, so as to nictation process fast and effeciently identify.One side and, support vector machine is a kind of learning art,, as a kind of sorting algorithm, is widely used in each branch of pattern-recognition.Support vector machine method is to be based upon on the theoretical and structure risk minimum principle basis of the VC dimension of Statistical Learning Theory, seeks optimal compromise according to limited sample information between the complicacy of model and learning ability, to obtaining best Generalization Ability.The classification based training method of traditional use experience risk minimization, neural network for example, they lack tight mathematic(al) treatment.And support vector machine is found the optimal classification face with structural risk minimization, on mathematics, has proved that this is equivalent to the minimum real risk of searching.So the Generalization Ability of support vector machine under the condition of finite sample is fine.

The size of the human eye area image after the yardstick normalizing is wide M pixel (as 40,30,20), high N pixel (as 20,15,10).The input of support vector machine is exactly the vector of M*N dimension like this.The present invention selects the kernel function of support vector machine by experiment, has compared 3 kinds of kernel functions:

Polynomial expression inner product approach 1:

K(x，x _i)＝(x·x _i+1) ^q，q＝1 (1)

Polynomial expression inner product approach 2:

K(x，x _i)＝(x·x _i+1) ^q，q＝2 (2)

The radial basis function inner product:

$K(x,x_i)=\exp \{-\frac{{|x-x_i|}^2}{{\mathrm{\σ}}^2}\}---\left(3\right)$

On the other hand, AdaBoost is a kind of learning art,, as a kind of sorting algorithm, is also each branch that is widely used in pattern-recognition.The present invention proposes based on the quick eyes of AdaBoost and open and close detection technique.

About the AdaBoost technology, on the ICCV of calendar year 2001, the researcher Paul Viola of Compaq and Michael J.Jones have showed their Real time face detection system, and its speed is average per second 15 frames, and the image size is 384x288.They have proposed three contributions technically: be used as the facial image feature with simple rectangular characteristic 1.; 2. based on the sorter of AdaBoost; 3. adopted the Cascade technology to improve detection speed.This technology has the characteristics of real-time; The present invention follows the tracks of this method, has realized a quick eyeball activity detection system based on AdaBoost.

The quick eyes based on AdaBoost that below the present invention adopted are opened and are closed algorithm and describe in detail:

(1) rectangular characteristic

The present invention adopts simple rectangular characteristic to characterize people's face.Use characteristic rather than directly use the brightness value of picture point why mainly contains 2 following considerations:

1, can be counted as a summary to the complex distributions of positive example and negative data based on the method for expressing of feature., for the sorter that utilizes statistical learning method,, because the number of training sample is limited, so the sample of complexity is once summarized the training effect of obtaining, be very important.

2, utilize feature faster than the speed of the gray-scale value of direct use picture point.The feature class that uses is similar to the people such as M.Oren of MIT use in human detection Haar wavelet basis function.Specifically, use 3 kinds of features, be referred to as rectangular characteristic, as shown in Figure 3.Useful two rectangles, useful three, also have a kind of use 4 rectangles to carry out calculated characteristics.The value of the feature of two rectangles be in a rectangle brightness value a little and deduct in another rectangular area institute's brightness value a little with.These rectangular areas big or small identical.The value of using the feature of three rectangles be exactly the both sides rectangle point brightness value and deduct in the middle of rectangular area in brightness value a little with.Finally, use the poor of rectangular area brightness value on two diagonal line of value of feature of 4 rectangular areas.

For the image of 24 * 24, whole set of this feature based on rectangular block have 130,000.It is worth mentioning that, different with the Haar small echo, all rectangular characteristic are redundancies here; So need an algorithm of selecting feature.

(2) integrogram

Feature based on rectangular area has a kind of algorithm very fast.This algorithm has used a kind of indirectly image representation method, is referred to as integrogram.On an integrogram, some x, the value of y be the top of this point on original image and left side institute brightness value a little with, as shown in Figure 4.Formulate is exactly:

$\mathrm{ii}(x,y)=\underset{x^{\′}\≤x,y^{\′}\≤y}{\mathrm{\Σ}}i(x^{\′},y^{\′})---\left(3\right)$

Here ii (x, y) is integrogram, and i (x, y) is original image, s (x, y) be accumulative total row and.Utilize following formula, integrogram can only a little just can calculate with original image of traversal.

s(x，y)＝s(x，y-1)+i(x，y) (4)

ii(x，y)＝ii(x-1，y)+s(x，y) (5)

Utilize integrogram, in rectangular area brightness a little and can quote and obtain by four times, as shown in Figure 5.Because the rectangle that uses in the middle of these features is all adjacent, thus use the feature of 2 rectangles only need to quote 6 times, 8 times of 3 rectangles, 9 times of 4 rectangles.Sorter training algorithm based on AdaBoost

Determine the feature of using, after positive example sample set and negative data set, next will train a sorter to classify to the image pattern of input, to judge that whether it is for opening eyes and the pattern of closing one's eyes.There is the method for a lot of machine learning can be used for training a sorter.

Due to 130,000 rectangular characteristic just being arranged in the window of 24 * 24, considerably beyond the number of picture elements of image, although it is very fast to calculate their speed, all using them is also a unpractical idea.In fact,, in this characteristic set the inside of big figure very, only need use the seldom feature of number just can form an effective sorter.So maximum problem is exactly how to go for effective feature.

The method that the present invention uses is a distortion of Adaboost learning algorithm.This algorithm has been completed selection feature and two tasks of training classifier simultaneously.Adaboost obtains a strong sorter by the training process of an iteration.After training for the first time a Weak Classifier, the weight of training sample is adjusted, thereby the weight of the sample of the correct classification of Weak Classifier that is not trained is for the first time increased.So iteration is gone down, and the sorter that finally obtains is a linear combination of Weak Classifier that each training is obtained.

The present invention opens and closes on this concrete application problem at human eye, utilizes the Adaboost algorithm of a modification.The fundamental purpose of this algorithm is to carry out feature selecting in training classifier.Its way is exactly in the training Weak Classifier, use be the Weak Classifier that only uses the training of rectangular characteristic to obtain.That is to say, it selects the Weak Classifier of a classification error minimum under current sample weights distribution situation in 130,000 Weak Classifiers.Through T iteration, the present invention selects T feature, and has obtained a strong sorter like this.Amended algorithm is as follows:

Known: (x _i, y _i) ..., (x _n, y _n) be training sample set, x _i∈ X, y _i∈ Y={0,1}

Initializes weights, establishing positive example sample and negative data has respectively m, 1. $w_{1,i}=\left\{\begin{array}{cc}\frac{1}{2m}& y_i=-1\\ \frac{1}{2l}& y_i=0\end{array}\right.$

For t＝1，…，T：

I) normalization weight,

Make w _tIt is a probability distribution.

Ii), to each feature j, train a Weak Classifier h _jOnly use this feature.Calculating has added the error of weight:

ε _j＝∑ _iw _i|h _j(x _i)-y _i|。

Iii) Select Error ε _tMinimum Weak Classifier h _t

Iv) upgrade weight If x _iBe classified correct e _i=0, otherwise e _i=1.Wherein ${\mathrm{\β}}_i=\frac{{\mathrm{\ϵ}}_i}{1-{\mathrm{\ϵ}}_i}.$

The sorter that output finally finally obtains:

$h\left(x\right)=\mathrm{sign}(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_i{h}_t\left(x\right)-\frac{1}{2}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t);$ Wherein, ${\mathrm{\α}}_i=\log \frac{1}{{\mathrm{\β}}_i};$ Sign () is sign function.

Although the present invention discloses as above with preferred embodiment; so it is not in order to limit the present invention, any those skilled in the art, without departing from the spirit and scope of the present invention; when doing a little modification and perfect, so protection scope of the present invention is worked as with being as the criterion that claims were defined.

Claims (6)

1. the method for quickly identifying of a living body faces, is characterized in that, comprises the steps:

A) input several continuous facial images;

B) every width facial image is determined pupil position and determined human eye area;

C) the judgement eyeball is opened the state of closing, if there is process by activity, differentiate nictation.

2. the method for quickly identifying of living body faces as claimed in claim 1, it is characterized in that, described step a) is obtained several continuous facial images according to the video flowing of input,, if adjacent two width facial images are not abandoned for same state, again obtains several continuous facial images.

3. the method for quickly identifying of living body faces as claimed in claim 1, is characterized in that, described step b) comprise following process:

B1) initialisation image size and video frame rate parameter;

B2) determine the size of current eye areas;

B3) the current scanning window of location 40 pixel x20 pixels;

B4) eliminate overlapping eyeball window.

4. the method for quickly identifying of living body faces as claimed in claim 1, is characterized in that, described step c) by training method of support vector machine and AdaBoost training method to opening eyes and close one's eyes the sample training.

5. the method for quickly identifying of living body faces as claimed in claim 4, is characterized in that, described step c) in the sample training algorithm of opening eyes and close one's eyes as follows:

C1): input training sample set (x ₁, y ₁) ..., (x _n, y _n), x _i∈ X, y _i∈ Y={0,1}; Initializes weights, establish positive example sample and negative data and be respectively m, and 1, $w_{1,i}=\left\{\begin{array}{cc}\frac{1}{2m}& y_i=-1\\ \frac{1}{2l}& y_i=0\end{array}\right.;$

C2): the normalization weight,

Make w _tBe a probability distribution, t=1 ..., T, T are the feature sum of choosing;

C3): to Weak Classifier h of each feature j training _jAnd calculating has added the error of weight:

ε _j＝∑ _iw _i|h _j(x _i)-y _i|；

C4): Select Error ε _iMinimum Weak Classifier h _t

C5): upgrade weight

If x _iBy correct Klassifizierung _i=0, otherwise e _i=1; Wherein, ${\mathrm{\β}}_i=\frac{{\mathrm{\ϵ}}_i}{1-{\mathrm{\ϵ}}_i};$

C6): the sorter that finally obtains is:

$h\left(x\right)=\mathrm{sign}(\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_i{h}_t\left(x\right)-\frac{1}{2}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\α}}_t);$ Wherein, ${\mathrm{\α}}_i=\log \frac{1}{{\mathrm{\β}}_i}.$

6., as the method for quickly identifying of the described living body faces of claim 1～5 any one, it is characterized in that, described step continuous facial image a) is 15～25 two field pictures adjacent in input video stream.

Images