CN103761504A - Face recognition system - Google Patents

Abstract

The invention relates to a face recognition system. The face recognition system comprises a face detection and positioning module, a standardization module, a feature extraction module and a face recognition module in sequence. The face recognition system has the advantages that the recognition accuracy can reach above 90%, and recognition requirements are basically met. The face recognition system is good in real-time performance, convenient to carry, and capable of being popularized to the fields of dynamic image tracking, motion detection and the like through modification of programs.

Description

A kind of face identification system

Technical field

The present invention relates to a kind of face identification system.

Background technology

The travelling speed of the related algorithm of recognition of face is slow, PAL vision signal is carried out the operation of acquisition and processing and people's face location and generally all cannot independently be carried out by PC, face recognition device volume is large, heavy, power consumption is high, carry inconvenience, and these have all limited application and popularization that people's face sets system.

Summary of the invention

The technical problem to be solved in the present invention is: based on the problems referred to above, the invention provides a kind of face identification system.

The present invention solves the technical scheme that its technical matters adopts: a kind of face identification system, comprises the detection of people's face and location, standardization, feature extraction and four modules of recognition of face successively.

Further, people's face detects with locating module and is: the coordinate (x that determines human eye ₁, y ₁) and (x ₂, y ₂), can indirectly obtain thus the left upper apex of square people's face and the coordinate on summit, bottom right, establish it and be respectively (X ₁, Y ₁) and (X ₂, Y ₂), its detailed computing method are as follows:

Width _eyes=x ₂-x ₁；

Width _face=Width _eyes/R _H；

X ₁=x ₁-（Width _face-Width _eyes）/2；

X ₂=X ₁+Width _face；

Height _eyes=(y ₁+y ₂)/2；

Height _face=Width _face；

Y ₁=Height _eyes-Height _face/R _V；

Y ₂=Y ₁+Height _face；

In formula, R _hand R _vbe empirical constant, value is 2.0 and 3.5 respectively.

Further, by the pre-service of DSP image, reach image normalization module.

Further, characteristic extracting module adopts principal component analysis (PCA), comprises the following steps:

The first step, collects N sample as training set X, obtains sample mean m, is shown below

$m=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_i$

Wherein, xi ∈ sample training collection X=(x1, x2 ..., xN).

Second step, obtains scatter matrix S, is shown below

$S=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}(x_i-m){(x_i-m)}^t$

Obtain eigenvalue λ i and the characteristic of correspondence vector ei of scatter matrix.Wherein, ei is principal component, and eigenwert is arranged in order to λ 1 from big to small, and λ 2 ...

Suppose to take out p value, λ 1, and λ 2 ..., λ p can determine face space E=(e1, e2 ..., eP), on this face space, in training sample X, the point that each element projects to this space can be obtained by following formula

x' _i＝E ^tx _i，t＝1，2，…，N

What by above formula, obtained is the p dimensional vector after PCA dimensionality reduction by former vector.

Further, face recognition module is that KNN sorter is classified, and the realization of KNN divides training and identification two steps.

Further, during the training of KNN, the input using the result after every class sample dimensionality reduction as KNN; During the identification of KNN, k nearest neighbor algorithm by a test point x be categorized as with its immediate K neighbour in there is that maximum classifications, from test sample book point, start growth, continuous enlarged area, until comprise into K training sample point, and the classification of test sample book point is classified as to the classification of frequency of occurrences maximum in this nearest K training sample point.

The invention has the beneficial effects as follows: the accuracy of identification of this face identification system can reach more than 90%, has substantially met identification requirement.System real time is good, easy to carry, can be generalized to by the modification of program the fields such as dynamic image tracking, motion detection.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the present invention is further described.

Fig. 1 is system construction drawing of the present invention;

In Fig. 2, (a), for calculating the process flow diagram of PCA projection matrix, (b) is the workflow diagram of KNN sorter;

Fig. 3 is the mode classification that adds the judgement of whether refusing.

embodiment

The invention will be further described in conjunction with specific embodiments now, and following examples are intended to illustrate the present invention rather than limitation of the invention further.

A kind of face identification system as shown in Figure 1, comprises the detection of people's face and location, standardization, feature extraction and four modules of recognition of face successively.

1. people's face detects and locating module

By acquired sample, judge the position of people's face, choose suitable people's face, intercepting out and making sample is important step.Face characteristic location has direct impact with the quality of feature extraction quality for facial image recognition effect.First determine the coordinate (x of human eye ₁, y ₁) and (x ₂, y ₂), can indirectly obtain thus the left upper apex of square people's face and the coordinate on summit, bottom right, establish it and be respectively (X ₁, Y ₁) and (X ₂, Y ₂), its detailed computing method are as follows:

Width _eyes=x ₂-x ₁；

Width _face=Width _eyes/R _H；

X ₁=x ₁-（Width _face-Width _eyes）/2；

X ₂=X ₁+Width _face；

Height _eyes=(y ₁+y ₂)/2；

Height _face=Width _face；

Y ₁=Height _eyes-Height _face/R _V；

Y ₂=Y ₁+Height _face；

In formula, R _hand R _vbe empirical constant, value is 2.0 and 3.5 respectively.So can in former figure, obtain the region coordinate of people's face, its size is with eye distance Width _eyessize and change, but as the input of PCA, require the dimension of input sample identical, so must be normalized picture.In design, income earner's face area sample is all zoomed to 24 * 24.Need in addition picture to be carried out to the operations such as contrast adjustment and histogram equalization, to improve the accuracy of identification.

Standardization module

The collection of image has adopted the common camera of pal mode output to add the Image Coding chip TVP5147 that TI company produces, and the view data of TVP5147 chip output is not rgb format, but export with yuv format.Need to convert rgb format to by dsp processor, just can carry out the pre-service of image, conversion formula as the formula (1)

R=Y+1.14V

G=Y-0.39U-0.58V （1）

B=Y+2.03U

DSP reads in memory headroom by view data, then it is carried out to computing, and the RGB obtaining is put into respectively to corresponding storage unit, and calculates gray-scale value Gray, and operational formula as the formula (2)

Gray=(R ³⁰+G ⁵⁹+B ¹¹+50)/100 （2）

The gray-scale value finally obtaining is stored in the middle of corresponding array.Every pictures consists of two field picture, so complete photo resolution is 720 * 576.But for system itself without its each pixel is changed, so intercepting wherein 320 * 240 is stored, the resolution of every is 320 * 120 like this, greatly reduces the time to gray level image pre-service and face locating by YUV, has improved the performance of system.

3. characteristic extracting module

When designer's face recognition classifier, conventionally regard a width picture as an one-dimensional vector.Although this and traditional regard picture as matrix form and have difference, but can extract and create favorable conditions for adopting principal component analysis (PCA) carry out eigenface.

The method of eigenface classification is piece image to be projected to a point in specific " face space ".This " face space " is comprised of one mutually orthogonal vector.These vectors are the important component parts that characterize each individual face cluster.The picture of different people face is far away differing of this space, and the projection of the different pictures of same person face on this space is at a distance of nearer.Therefore can use the method for PCA is that whole face identification system lays the first stone.

The first step, collects N sample as training set X, obtains sample mean m, as the formula (3)

$m=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_i---\left(3\right)$

Wherein, xi ∈ sample training collection X=(x1, x2 ..., xN).

Second step, obtains scatter matrix S, as the formula (4)

$S=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}(x_i-m){(x_i-m)}^t---\left(4\right)$

According to the ultimate principle of PCA, must obtain eigenvalue λ i and the characteristic of correspondence vector ei of scatter matrix.Wherein, ei is principal component, and the size of its characteristic of correspondence value represents the number of its inclusion information.So eigenwert need to be arranged in order to λ 1 from big to small, λ 2 ...

Suppose to take out p value, λ 1, and λ 2 ..., λ p can determine face space E=(e1, e2 ..., eP), on this face space, in training sample X, the point that each element projects to this space can be obtained by formula (5)

x' _i＝E ^tx _i，t＝1，2，…，N （5）

What by above formula, obtained is the p dimensional vector after PCA dimensionality reduction by former vector, and next step is inputted KNN sorter to classify.

4. face recognition module

Face recognition module is that KNN sorter is classified.The realization of KNN divides training and identification two steps.During training, the input using the result after every class sample dimensionality reduction as KNN.K nearest neighbor algorithm by a test point x be categorized as with its immediate K neighbour in there is that maximum classifications, from test sample book point, start growth, continuous enlarged area, until comprise into K training sample point, and the classification of test sample book point is classified as to the classification of frequency of occurrences maximum in this nearest K training sample point.Choosing appropriate K value has a significant impact the result of classification.If K value is chosen when excessive, may be able to more correctly classify, but sacrifice performance simultaneously, improved computation complexity.If it is too small that K value is chosen, greatly reduce computation complexity, but may affect the accuracy of classification.

By 320 * 240 the picture obtaining after the detection of remarkable face, using the part of intercepting people face as people's face sample.During design, all samples of people's face all will show on display, reduce the possibility of people's face error-detecting, improve to a certain extent the accuracy of system.

The sample resolution of people's face is 24 * 24, as the one-dimensional vector of 576 dimensions, inputs to PCA.In Fig. 2, (a) is for calculating the process flow diagram of PCA projection matrix, (b) be the workflow diagram of KNN sorter, the numerical value of training sample after PCA projection wherein, need in each identification, not recalculate, calculating while can be used as initialization, also can be stored in power down non-volatile media, in Flash storer, can improve the operational efficiency of equipment, reduce operand.

As shown in Figure 2, KNN sorter can judge immediate classification, but can not refuse classification, so produced anyone face, all will be assigned in a class of built-in sample set.Such mode classification is worthless, so must add the judgement of whether refusing, process flow diagram as shown in Figure 3.

As shown in Fig. 3, when sample point is after PCA dimensionality reduction, be delivered to KNN sorter and classify, resulting result necessarily can be judged to be K class, now can not be eager to come to a conclusion, and first obtains the Euclidean distance sum sum of the sample point of tested point and K class label.Define two threshold value a and b, if sum<a value is judged to be the first kind; If sum>b value, is judged to be refusal class; If sum, between a and b value, introduces precision controlled quentity controlled variable accuracy, calculate the difference of sum and a, if be less than precision controlled quentity controlled variable accuracy, be judged to be K class, otherwise refusal classification.By such process, indirectly solved sample wrong minute and cannot sentence no problem.

In this experiment, the value that the value of selected a is 12400, b is 16200, and definite needs of these two values carry out a large amount of experiments, therefrom find out rule.The size of the value of x directly affects the effect of identification, chooses x=4 respectively and x=5 tests.

(1) during x=4: program test can identify in storehouse adhere to 12 people's 36 width facial image separately time, correctly identify 33 width wherein, all the other 3 width images are all judged to no, sentence mistake for 0.Program test can not identify in storehouse adhere to 3 people's 33 width facial image separately time, 22 width images are successfully sentenced no, 11 width are judged by accident;

(2) during x=5: program test can identify in storehouse adhere to 12 people's 36 width facial image separately time, correctly identify 25 width wherein, all the other ll width images are all judged to no, O width is sentenced mistake.Program test can not identify in storehouse adhere to 3 people's 33 width facial image separately time, 28 width images are successfully sentenced no, 5 width are judged by accident.

The experimental data of analyzing is above known, and during x=4, the discrimination that can identify storehouse is 91.6%, and can not identify the no rate of sentencing of storehouse is 66.7%.During x=5, the discrimination that can identify storehouse is 69.4%, and can not identify the no rate of sentencing of storehouse is 84.8%.Therefore, while being applied to different occasions, should select different x values, when requiring to refuse as far as possible stranger's face, optional x value is 5, and when requiring to identify known person face, optional x value is 4 as far as possible.

During system, select DSP6713, the floating point processor of the C6000 series that Zhe Shi TI company produces, it has adopted vliw architecture, and the equivalent period number of instruction operation is lower, and travelling speed is very fast.The collection of image has adopted the common camera of pal mode output to add the Image Coding chip TVP5147 that TI company produces, this chip is supported multiple types, multiple interfaces is inputted, and can export the video data of yuv format, and line synchronizing signal and vertical synchronizing signal etc. are provided simultaneously.The temporary CPLD of use of data and SRAM realize.Design system forms.

The above-mentioned foundation desirable embodiment of the present invention of take is enlightenment, and by above-mentioned description, relevant staff can, within not departing from the scope of this invention technological thought, carry out various change and modification completely.The technical scope of this invention is not limited to the content on instructions, must determine its technical scope according to claim scope.

Claims (6)

1. a face identification system, is characterized in that: comprise successively the detection of people's face and location, standardization, feature extraction and four modules of recognition of face.

2. a kind of face identification system according to claim 1, is characterized in that: described people's face detects with locating module and is: the coordinate (x that determines human eye ₁, y ₁) and (x ₂, y ₂), can indirectly obtain thus the left upper apex of square people's face and the coordinate on summit, bottom right, establish it and be respectively (X ₁, Y ₁) and (X ₂, Y ₂), its detailed computing method are as follows:

Width _eyes=x ₂-x ₁；

Width _face=Width _eyes/R _H；

X ₁=x ₁-（Width _face-Width _eyes）/2；

X ₂=X ₁+Width _face；

Height _eyes=(y ₁+y ₂)/2；

Height _face=Width _face；

Y ₁=Height _eyes-Height _face/R _V；

Y ₂=Y ₁+Height _face；

In formula, R _hand R _vbe empirical constant, value is 2.0 and 3.5 respectively.

3. a kind of face identification system according to claim 1, is characterized in that: by the pre-service of DSP image, reach image normalization module.

4. a kind of face identification system according to claim 1, is characterized in that: described characteristic extracting module adopts principal component analysis (PCA), comprises the following steps:

The first step, collects N sample as training set X, obtains sample mean m, is shown below

$m=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{x}_i$

Wherein, xi ∈ sample training collection X=(x1, x2 ..., xN).

Second step, obtains scatter matrix S, is shown below

$S=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}(x_i-m){(x_i-m)}^t$

Obtain eigenvalue λ i and the characteristic of correspondence vector ei of scatter matrix.Wherein, ei is principal component, and eigenwert is arranged in order to λ 1 from big to small, and λ 2 ...

Suppose to take out p value, λ 1, and λ 2 ..., λ p can determine face space E=(e1, e2 ..., eP), on this face space, in training sample X, the point that each element projects to this space can be obtained by following formula

x' _i＝E ^tx _i，t＝1，2，…，N

What by above formula, obtained is the p dimensional vector after PCA dimensionality reduction by former vector.

5. a kind of face identification system according to claim 1, is characterized in that: described face recognition module is that KNN sorter is classified, and the realization of KNN divides training and identification two steps.

6. a kind of face identification system according to claim 5, is characterized in that: during the training of described KNN, and the input using the result after every class sample dimensionality reduction as KNN; During the identification of KNN, k nearest neighbor algorithm by a test point x be categorized as with its immediate K neighbour in there is that maximum classifications, from test sample book point, start growth, continuous enlarged area, until comprise into K training sample point, and the classification of test sample book point is classified as to the classification of frequency of occurrences maximum in this nearest K training sample point.

Images