CN101226591A - Personal identification method based on mobile phone pick-up head combining with human face recognition technique - Google Patents

Abstract

Disclosed is an identification method based on the combination of cell phone cameras and face recognition technique in the field of wireless network technique and pattern recognition. A mobile network is used to transfer easily cognizable information of to-be-identified persons and face image information collected by the cell phone cameras to a server which combines the data meeting the specifications of sex, age and the like in a data bank to a candidate data set according to the easily cognizable information. The sever identifies the face data in the candidate data set by employing principal component analysis and linear discrimination analysis for training, identifies image-transferred face areas and selects the most similar results, the number of which is set by users, from the identification results to transfer back to cell phones for final judgment by the users. The invention realizes the purposes of largely reducing cost and difficulty and increasing speed and accuracy of identification.

Description

Personal identification method based on mobile phone pick-up head combining with human face recognition technique

Technical field

What the present invention relates to is the method in a kind of image recognition technology field, specifically is a kind of personal identification method based on mobile phone pick-up head combining with human face recognition technique.

Background technology

The biometric identity recognition methods is the biological characteristic that utilizes human body, the method that people's identity is discerned.Traditional biological characteristic comprises people's face, fingerprint, human eye iris, vocal print, smell or the like.Biometric identity is identified in different fields such as public security, bank, personal information preservation and has obtained to use widely.People's face detects to refer to and detect human face region in piece image, and recognition of face then refers to utilize people's face vision biological information to carry out authentication.These two technology all are the important technologies in the area of pattern recognition.People's face detects and recognition technology is widely applied to many fields such as identification, man-machine interface, content-based retrieval, Digital Video Processing, visual monitoring.Traditional biometric identity recognition methods for example fingerprint recognition and human eye iris recognition all needs special checkout equipment, and this kind equipment is expensive and heavy, only can use in the minority occasion, be unfavorable for widespread use.

Find through retrieval prior art, Chinese patent application number is 200410097054.0, patent name is " utilizing the camera that is provided with on the mobile phone to confirm the device and method of identity ", this invention utilization is arranged on camera on the mobile phone and takes face and import after the characteristic information of health, from the image that comprises background, extract face's field profile and extract face feature and be sent to identity information database and retrieve, and image as a result re-send on the mobile phone, thereby confirm identity quickly and accurately.But this method need be too much in the information that mobile phone end is handled, and the mobile phone end performance expends too high, makes system accuracy and real-time all descend.The described extraction of this method people face eyebrow on the other hand, features such as bone do not have the method for practical value at present, so its operability and realizability are not strong.

Summary of the invention

The present invention is directed to above-mentioned the deficiencies in the prior art, a kind of personal identification method based on mobile phone pick-up head combining with human face recognition technique has been proposed, make it utilize the mobile network that person's to be identified easy identification information and the human face photo image information by the mobile phone cam collection are sent to server end, the inquiry decision system of identification server is formed a candidate data collection according to easy identification information will satisfy easy identification information in the database data, after people's face data that identification server by utilizing principal component analysis (PCA) (PCA) and linear discriminant analysis (LDA) are concentrated candidate data are carried out recognition training, human face region to input picture is discerned, and the result who selects the user to set number from recognition result sends back mobile phone end confession user by MMS (Multimedia Message Service) (MMS) and finally judges.

The present invention is achieved by the following technical solutions, the present invention includes following concrete steps:

Step 1, the user obtains person's to be identified easy identification information, and with easy identification information input handset end;

Step 2 utilizes mobile phone cam to gather person's facial image to be identified;

Step 3, easy identification information and facial image that step 1 and step 2 are obtained respectively are sent to server end by mobile Internet, and send the identification request to server end;

Step 4, utilization is selected the data the identity database of server end from the person's to be identified that mobile phone end is uploaded easy identification information, and the data that will conform to the easy identification information of person to be identified are formed a candidate data collection D _New

Step 5, server end utilize principal component analysis (PCA) and linear discriminant analysis to candidate data collection D _NewIn people's face data carry out recognition of face training, generate human face recognition transformation matrix H;

Step 6 utilizes oval complexion model to handle the facial image that mobile phone end transmits, and obtains area of skin color wherein;

Step 7 utilizes people's face cascade sort detecting device to detect area of skin color, obtains human face region wherein and generates people's face F to be identified _Input

Step 8 is from candidate data collection D _NewThe middle width of cloth facial image of selecting is as comparison people face F _i, the human face recognition transformation matrix H that generates with step 5 is to comparison people face F _iAnd people's face F to be identified _InputCarry out linear transformation, generate the human face recognition vector v _iAnd v _Input, v wherein _iBe comparison people face F _iThe human face recognition transformation results, v _InputBe people's face F to be identified _InputThe human face recognition transformation results;

Step 9 is calculated the human face recognition vector v _iAnd v _InputEuclidean distance d _{I, input}, i=1,2...i are candidate data collection D _NewMiddle identity information quantity;

Step 10 is to candidate data collection D _NewEveryone face data repeating step eight and step 9, and the Euclidean distance d that is generating _{I, input}In, according to Euclidean distance order from small to large choose the user set number identity data return mobile phone end as recognition result, the mobile phone holder will carry out final identification according to the identity data that this user sets number.

Described easy identification information, be meant person to be identified information external, that be easy to gather, comprise: sex, the range of age, height etc., utilize easy identification information the information filtering that obviously is not inconsistent in the database, reduce the scope of recognition of face, improve the accuracy rate of identification.

Person's facial image to be identified of described collection, be meant the facial image of gathering the RGB color space by mobile phone cam, images acquired resolution size is 160 * 120, in gathering the facial image process, make person's eyes to be identified be positioned at the position of human eye frame that acquisition software is set, and the looking natural of face, illumination are good, and image background is simple.

Described identity database, be meant the database of storage personally identifiable information, wherein each identity information essential information and a human face photo, essential information comprises: name, sex, age, height, body weight, blood group, human face photo detects the cascade classifier processing through oval complexion model and people's face, has obtained human face region and has removed background.

Described principal component analysis (PCA), be meant by extracting the main feature of data set and reduce the data set dimension, it is a linear transformation, in data projection to a new linear space, in new linear space according to how many sorting data proper vectors from high to low of information content, select the higher low volume data of information content, form the new data set of the low dimension that has kept the most information that former data set comprised.

Described principal component analysis (PCA), its linear transformation expression formula is as follows:

y＝Ψ ^Tx

Wherein: x is a vector before the conversion, and Ψ is the principal component analysis (PCA) transformation matrix, and T represents transposition, and y is the principal component analysis (PCA) result vector.

Described principal component analysis (PCA) transformation matrix Ψ, its method of determining is specific as follows:

If data set X, wherein total N M dimension data x _i, i=1...N, N＜＜M, data set X is expressed as the matrix of a M * N, be referred to as matrix X, establish Φ=[φ ₁, φ ₂... φ _M], φ wherein _i(i=1...M) be the M dimensional vector, so Φ is M * M square formation, then has:

∑ _Xφ _i＝λ _iφ _i

Wherein: λ _iBe eigenwert, ∑ _XBe the covariance matrix of matrix X, ∑ _X=E[(X-E[X]) (X-E[X]) ^T]

Be φ _i(i=1...M) be the covariance matrix ∑ of X _XCorresponding to eigenvalue _iProper vector;

With eigenvalue _i(i=1...M) be λ according to ordering from big to small ₍₁₎, λ ₍₂₎... λ _(M), its characteristic of correspondence vector is respectively φ ₍₁₎, φ ₍₂₎... φ _(M), then

${\mathrm{\Ψ}}^T=\left\{\begin{array}{c}{\mathrm{\φ}}_{\left(1\right)}^T\\ {\mathrm{\φ}}_{\left(2\right)}^T\\ \·\\ \·\\ \·\\ {\mathrm{\φ}}_{\left(300\right)}^T\end{array}\right\}$

Ψ=(Ψ ^T) ^TBe the principal component analysis (PCA) transformation matrix.

Described linear appreciation analysis, being meant by extracting the categories of datasets distinguishing characteristic makes the data set dimension reduce, the same with principal component analysis (PCA), it is a kind of linear transformation, linear appreciation is analyzed in N class data projection to the new linear space, make in this new linear space that Euclidean distance is nearest between the different sample datas in the same classification, Euclidean distance between the inhomogeneity farthest simultaneously, projection has not only reduced the dimension of data like this, it also handles what is more important at classification information, has improved the accuracy of classification.

Described linear appreciation analysis, its linear transformation expression formula is as follows:

z＝W ^Ty

Wherein: z is the human face recognition vector, and y is the result vector of principal component analysis (PCA), is in the human face recognition linear space, and W is linear appreciation analytic transformation matrix.

Described linear appreciation analytic transformation matrix W, it determines that method is as follows:

If data set Y has N kind C _i, i=1...N, data are the M dimension, and each class data comprises 1 or a plurality of sample y _i, i=1,2.. has for this N class problem,

$S_w^{-1}{S}_{b}W={\mathrm{\Λ}}_{w}W$

Wherein: Λ _wBe eigenvalue matrix, W is linear appreciation analytic transformation matrix, is matrix S _w ^-1S _bCorresponding to Λ _wEigenvectors matrix, S _wBe class inscattering matrix, characterize the situation of data variance in the same class,

$S_w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left\{\mathrm{Pr}\left(C_i\right)E\right[{(y-E\left(y_i\right))(y-E\left(y_i\right))}^T|C=C_i]\}$

And S _bBe scattering matrix between class, characterize data variance distribution situation between inhomogeneity,

$S_b=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left\{\mathrm{Pr}\left(C_i\right)(E\left(y_i\right)-E\left(y\right)){(E\left(y_i\right)-E\left(y\right))}^T\right\}$

Pr (C in wherein above-mentioned two formulas _i) be class C _iPrior probability, expression formula is:

$\mathrm{Pr}\left(C_i\right)=\frac{1}{N},(i=1...N)$

The prior probability of each class is equal,

By obtaining matrix S _w ^-1S _bEigenvectors matrix and eigenvalue matrix, just can unique definite linear appreciation analytic transformation matrix W.

Described human face recognition transformation matrix H is specially: H ^T=W ^TΨ ^T

Human face recognition is transformed to: z=H ^Tx

Wherein: Ψ and W are respectively above-mentioned principal component analysis (PCA) transformation matrix and linear Identification analytic transformation matrixes, and x is input facial image vector, and z is the human face recognition vector.

The oval complexion model of described utilization is handled the facial image that mobile phone end transmits, and comprises following concrete steps:

1. the rgb value of each pixel in the facial image that mobile phone end is collected (R, G B) convert to (Y, Cb, Cr) value, wherein, Y represents that monochrome information, Cb represent that blue colour difference information, Cr represent red colour difference information;

2. the parameter of oval complexion model is set;

3. to each pixel in the image, carry out skin color segmentation, obtain black and white two-value broca scale picture;

4. be communicated with in the border of area of skin color at four of black and white two-value broca scale picture, obtain the area of skin color positional information.

Described people's face cascade sort detecting device, be that linear combination by Weak Classifier obtains a strong classifier, again the strong classifier combination is obtained a complete cascade classifier, utilize the repeatedly classification of cascade to improve the accuracy that detects, the Weak Classifier that is used to detect, and utilize Adaboost (self-adaptation adjustment) method training to obtain by the strong classifier that weak typing linear combination obtains, people's face cascade sort detecting device that final training is come out comprises 20 strong classifiers, people's face cascade sort detecting device constructs Weak Classifier successively by reading the classifier data file that trains, strong classifier and cascade classifier.Whether Weak Classifier directly provides according to the input rectangular area is the classification results of people's face, strong classifier is determined classification results according to a plurality of Weak Classifier results' weighting, cascade classifier is obtained by the cascade of multilayer strong classifier, if certain strong classifier is non-face with this territorial classification, then cascade classifier thinks that detected zone is non-face, direct detection of end, otherwise, cascade classifier is sent data into next strong classifier classification, when having only all strong classifiers to think that all this zone is for human face region, people's face cascade sort detecting device just provides the zone and is the classification results of people's face.

The described human face recognition transformation matrix H that generates that uses is respectively to comparison people face F _iAnd people's face F to be identified _InputCarry out the human face recognition linear transformation, and generate corresponding human face recognition vector v _iAnd v _Input, specific as follows:

v _i＝H ^TF _i

v _input＝H ^TF _input

Wherein, F _iBe the people's face that from database, takes out, F _InputBe people's face to be identified that mobile phone end is uploaded, above-mentioned two formulas are respectively to F _iAnd F _InputCarried out the human face recognition conversion, v _iAnd v _InputBe corresponding human face recognition vector, final recognition of face will be undertaken by comparing these two vectors.

Described calculating human face recognition vector v _iAnd v _InputEuclidean distance d _{I, input}, specific as follows:

$d_{i,\mathrm{input}}=\sqrt{\underset{j=1}{\overset{300}{\mathrm{\Σ}}}{(v_{i,j}-v_{\mathrm{input},j})}^2}$

Wherein: v _{I, j}And v _{Input, j}It is respectively the human face recognition vector v _iAnd v _InputJ component, d _{I, input}It is vector v _iAnd v _InputEuclidean distance, this is apart from characterizing vector v _iAnd v _InputSimilarity degree, d _{I, input}More little then similarity degree is high more.

Compared with prior art, the present invention has following beneficial effect: owing to adopt mature technology, and characteristics such as the movability of mobile phone, low cost and high popularity rate, the present invention realizes that cost and difficulty descend greatly.The mobile phone that can reach this system mobile phone end requirement is One's name is legion not only, and price is lower, and common hundreds of units get final product.The present invention uses user's input information that identity database is filtered before identification on the other hand, provides 10 data to select for the user when carrying out final identification, has improved the speed and the accuracy of identification.Show that after tested the identification accuracy of this system has reached more than 85%, make the reliability of this system reach the requirement of commercial application level.Above-mentioned 2 make the widespread deployment of this identification system become a reality, and also making the related scientific research achievements conversion is that the ability of commercial product is improved.The present invention all has very high application and commercial value in different field such as safety guarantee, traffic administration, population managements.

Description of drawings

Fig. 1 is a workflow synoptic diagram of the present invention;

Fig. 2 is inventor's face cascade sort detecting device testing process synoptic diagram;

Fig. 3 is inventor's face identification process synoptic diagram;

Fig. 4 (a) is a mobile phone end man face image acquiring synoptic diagram;

Fig. 4 (b) is people's face cascade sort detecting device testing result figure;

Fig. 5 is the mutual synoptic diagram of identification server-mobile phone end.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated: present embodiment has provided detailed embodiment and process being to implement under the prerequisite with the technical solution of the present invention, but protection scope of the present invention is not limited to following embodiment.

The server end basic condition that present embodiment adopted is as follows:

Server end has adopted Microsoft SQL Server 2005 that database service is provided, recognizer adopts C# language based on Microsoft.NET Framework 2.0 exploitations, this program is provided services on the Internet at front end, obtain the identification request from mobile phone end, rear end and SQL database carry out carrying out concrete identification alternately; Behind the end of identification result is sent back mobile phone by MMS (Multimedia Message Service).

Collected 353 people's identity information in the database of server end, human face region totally 5 information have only been kept in name, sex, age, human face photo, the photo for the sake of simplicity, wherein name is as unique definite these personnel of tables of data major key, everyone one of human face photo, amount to 353, every human face photo size is 160 * 120 pixels, these photos all pass through oval complexion model and people's face cascade sort detector processes, obtained human face region information, the distribution situation of these data is as shown in table 1:

Table 1 age-sex distribution table

Age sex	The man	The woman	Amount to
				10-20	54	46	100
21-30	118	87	205
				31-40	28	20	48
Amount to	200	153	353

The mobile phone end basic condition is as follows in the present embodiment: mobile phone end adopts the smart mobile phone that carries Symbian (letter is than peace) S60 operating system as the mobile phone end platform, the mobile phone end program adopts C Plus Plus to develop based on SymbianS60 SDK, need open GPRS online and two services of multimedia messages (MMS) that China Mobile provides when the mobile phone end program is used, mobile phone end can be carried out with server end alternately like this.

As shown in Figure 1, present embodiment comprises the steps:

Step 1, Information Monitoring: mobile phone end is imported person's to be identified easy identification information, for example: import following information: " man, 21-30 year ";

Gather person's to be identified facial image, the facial image size is 160 * 120 pixels, and shown in Fig. 4 (a), human eye is arranged in the position of human eye frame, and face is proper to look natural, and background is simple.

Step 2, upload information: mobile phone end is opened GPRS (GPS) network automatically and is connected, set up communication with server end and be connected, obtain server end connect set up response after, with the information easy to identify of step 1 collection and facial image by the GPRS network end that uploads onto the server.

Step 3, server end filtering information: after server end is received above-mentioned information, all information that satisfy " man, 21-30 year " in the identity database are selected, formed candidate data collection D _New, comprise 118 identity informations, corresponding 118 facial images.

Step 4, human face recognition transformation matrix H training: human face recognition transformation matrix H training comprises principal component analysis (PCA) (PCA) transformation matrix Ψ training and linear appreciation analysis (LDA) transformation matrix w training two parts.

Described principal component analysis (PCA) transformation matrix Ψ training, specific as follows:

1. according to aforesaid analysis, candidate data collection D _NewIn comprise 118 facial images, each image size is 160 * 120 pixels, be the data of 160 * 120=19200 dimension, so principal component analysis (PCA) conversion training dataset X is 19200 * 118 matrix.

According to the formula ∑ _X=E[(X-E[X]) (X-E[X]) ^T]

Ask the covariance matrix of X, after calculating finishes, ∑ _XIt is 19200 * 19200 square formation;

2. according to the formula ∑ _Xφ _i=λ _iφ _i, find the solution ∑ _XProper vector and eigenwert be respectively

φ _i，λi(i＝1...19200)

3. with λ _i(i=1...19200) be λ according to ordering from big to small ₍₁₎, λ ₍₂₎... λ _(19200)Its characteristic of correspondence vector is respectively φ ₍₁₎, φ ₍₂₎... φ _(19200), then

${\mathrm{\Ψ}}^T=\left\{\begin{array}{c}{\mathrm{\φ}}_{\left(1\right)}^T\\ {\mathrm{\φ}}_{\left(2\right)}^T\\ \·\\ \·\\ \·\\ {\mathrm{\φ}}_{\left(300\right)}^T\end{array}\right\}$

Be 300 * 19200 matrix, training dataset X is carried out principal component analysis (PCA) projective transformation Y=Ψ ^TX, Y are linear appreciation analyzing and training data set, and data set Y is 300 * 118 matrixes.

Described linear appreciation analytic transformation matrix W training, specific as follows:

According to formula: $S_w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left\{\mathrm{Pr}\left(C_i\right)E\right[{(y-E\left(y_i\right))(y-E\left(y_i\right))}^T|C=C_i]\}$

$S_b=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left\{\mathrm{Pr}\left(C_i\right)(E\left(y_i\right)-E\left(y\right)){(E\left(y_i\right)-E\left(y\right))}^T\right\}$

$\mathrm{Pr}\left(C_i\right)=\frac{1}{N},(i=...N)$

Find the solution S _wAnd S _b, find the solution the back matrix S _wAnd S _bBe 300 * 300 matrixes.

According to formula

$S_w^{-1}{S}_{b}W=W{\mathrm{\Λ}}_w$

Find the solution S _w ^-1S _bProper vector and eigenwert be respectively: w _i, λ _i(i=1...300) linear appreciation analytic transformation matrix W is

W＝[w ₍₁₎，w ₍₂₎...w ₍₃₀₀₎]

Linear appreciation analytic transformation matrix W is 300 * 300 matrixes.

Described human face recognition transformation matrix H ^T=W ^TΨ ^T, H ^TBe 300 * 19200 matrixes.

Step 5 is treated recognition image and is carried out people's face and detect, and people's face detects and comprises two parts: use that oval complexion model obtains the area of skin color positional information at the YCbCr color space, end user's face cascade sort detecting device detects human face region from area of skin color.

The oval complexion model of described use obtains the area of skin color positional information at the YCbCr color space, is specially:

1. with the rgb value of each pixel in the facial image that collects in the step 1 (R, G, B), by following formula convert to (Y, Cb, Cr) value:

$\left\{\begin{array}{c}Y=0.299R+0.587G+0.114B\\ \mathrm{Cb}=-\frac{R}{6}-\frac{G}{3}+\frac{B}{2}+128\\ \mathrm{Cr}=\frac{R}{2}-\frac{5G}{12}-\frac{B}{12}+128\end{array}\right.$

Wherein, Y represents that monochrome information, Cb represent that blue colour difference information, Cr represent red colour difference information.

Parameter and value thereof that 2. oval complexion model is set are shown below:

Cx＝109.38，Cy＝152.02，ecx＝1.60，ecy＝2.41

a＝25.29，b＝14.03，θ＝2.53

3. to each pixel in the image, carry out skin color segmentation, obtain two-value broca scale picture:

By the value of this pixel in the YCbCr space be (Y, Cb Cr) calculate the skin tone value of this pixel by following formula:

x＝(Cb-Cx)×cosθ+(Cr-Cy)×sinθ

y＝(Cb-Cx)×(-sinθ)+(Cr-Cy)×cosθ

If this skin tone value is less than 1, then this pixel belongs to area of skin color, is provided with that this pixel is a colour of skin point in the two-value broca scale picture; Otherwise do not belong to area of skin color, be provided with that this pixel is a background dot in the two-value broca scale picture.

As shown in Figure 2, described end user's face fruit cascade detectors detects human face region from area of skin color, be specially: whether Weak Classifier directly provides according to the input rectangular area is the classification results of people's face, strong classifier is determined classification results according to a plurality of Weak Classifier results' weighting, cascade classifier is obtained by the cascade of multilayer strong classifier, if certain strong classifier is non-face with this territorial classification, then cascade classifier thinks that detected zone is non-face, direct detection of end, otherwise, cascade classifier is sent data into next strong classifier classification, when having only all strong classifiers to think that all this zone is for human face region, cascade classifier just provides the zone and is the classification results of people's face.The testing result of people's face cascade sort detecting device is seen accompanying drawing 4 (b), and white box institute enclosing region is human face region among the figure.

Step 6 is carried out the human face recognition conversion to input people's face and comparison facial image.

Described human face recognition conversion is meant

v _i＝H ^TF _i

v _input＝H ^TF _input

Comparison people face F in above-mentioned two formulas _iReach people's face F to be identified that mobile phone end is uploaded _InputBe 19200 * 1 vector, and H ^TBe 300 * 19200 matrixes, so the human face recognition vector v after the conversion end _iAnd v _InputBe 300 * 1 vectors.

Step 7, the Euclidean distance between the human face recognition vector that calculation procedure six generates.

Described Euclidean distance is calculated according to following formula:

$d_{i,\mathrm{input}}=\sqrt{\underset{j=1}{\overset{300}{\mathrm{\Σ}}}{(v_{i,j}-v_{\mathrm{input},j})}^2}$

Wherein: v _{I, j}And v _Input, j is respectively the human face recognition vector v _iAnd v _InputJ component, d _{I, input}It is vector v _iAnd v _InputEuclidean distance, this is apart from characterizing vector v _iAnd v _InputSimilarity degree, d _{I, input}More little then similarity degree is high more.

The recognition of face flow process of above-mentioned steps six and step 7 as shown in Figure 3.

Step 8 to all data repeating steps six, seven of satisfy condition " man, 21-30 year ", is preserved all Euclidean distance d that generate _{I, input}, and from the data that satisfy above-mentioned condition, choose the identity data of 10 correspondences as recognition result according to Euclidean distance order from small to large.

Step 9 returns the recognition result of above-mentioned generation to mobile phone end.Recognition result will be divided into 10 multimedia messagess (MMS) and be sent to mobile phone end by the multimedia messages center, and the mobile phone holder will carry out final identification according to these 10 multimedia messagess (MMS).Described mobile phone end of step 1 and step 9 and server end reciprocal process are as shown in Figure 5.

After above-mentioned steps finishes, the cellphone subscriber will obtain enough information and carry out identification, and and then take further action.

Present embodiment combines mobile phone by utilizing mature technology (C# .NET technology, SQL Server, Symbian S60) with recognition of face, realized the reliable identity recognition technology.This technology has overcome limitation such as existing GPRS mobile Internet bandwidth is lower, and the smart mobile phone performance limitations is more, has realized a system that can carry out practical application.Its initial two ends participate in controlled characteristic by the user, make this system have the high accuracy that common face identification system did not have.The probability that comprises correct recognition result in its 10 width of cloth images that finally return of test shows to this embodiment is greater than 90%, and promptly the rate of accuracy reached of this system is to 90%.Show also that through test an identification request on average only needs 10 second time from issuing a request to return results on the other hand.

Claims (10)

1. the personal identification method based on mobile phone pick-up head combining with human face recognition technique is characterized in that, comprises the steps:

Step 1, the user obtains person's to be identified easy identification information, and with easy identification information input handset end;

Step 2 utilizes mobile phone cam to gather person's facial image to be identified;

Step 3, easy identification information and facial image that step 1 and step 2 are obtained respectively are sent to server end by mobile Internet, and send the identification request to server end;

Step 4, utilization is selected the data the identity database of server end from the person's to be identified that mobile phone end is uploaded easy identification information, and the data that will conform to the easy identification information of person to be identified are formed a candidate data collection D _New

Step 5, server end utilize principal component analysis (PCA) and linear discriminant analysis to candidate data collection D _NewIn people's face data carry out recognition of face training, generate human face recognition transformation matrix H;

Step 6 utilizes oval complexion model to handle the facial image that mobile phone end transmits, and obtains area of skin color wherein;

Step 7 utilizes people's face cascade sort detecting device to detect area of skin color, obtains human face region wherein and generates people's face F to be identified _Input

Step 8 is from candidate data collection D _NewThe middle width of cloth facial image of selecting is as comparison people face F _i, the human face recognition transformation matrix H that generates with step 5 is to comparison people face F _iAnd people's face F to be identified _InoutCarry out linear transformation, generate the human face recognition vector v _iAnd v _Input, v wherein _iBe comparison people face F _iThe human face recognition transformation results, v _InputBe people's face F to be identified _InputThe human face recognition transformation results;

Step 9 is calculated the human face recognition vector v _iAnd v _InputEuclidean distance d _{I, input}, i=1,2...i are candidate data collection D _NewMiddle identity information quantity;

Step 10 is to candidate data collection D _NewEveryone face data repeating step eight and step 9, and the Euclidean distance d that is generating _{I, input}In, according to Euclidean distance order from small to large choose the user set number identity data return mobile phone end as recognition result, the mobile phone holder will carry out final identification according to the identity data that this user sets number.

2. a kind of personal identification method according to claim 1 based on mobile phone pick-up head combining with human face recognition technique, it is characterized in that, person's facial image to be identified of described collection, be meant the facial image of gathering the RGB color space by mobile phone cam, images acquired resolution size is 160 * 120, in gathering the facial image process, makes person's eyes to be identified be positioned at the position of human eye frame that acquisition software is set, and the looking natural of face, illumination are good, and image background is simple.

3. a kind of personal identification method according to claim 1 based on mobile phone pick-up head combining with human face recognition technique, it is characterized in that, described identity database, be meant the database of storage personally identifiable information, wherein each identity information comprises essential information and a human face photo, essential information comprises: name, sex, age, height, body weight, blood group, human face photo detects the cascade classifier processing through oval complexion model and people's face, has obtained human face region and has removed background.

4. a kind of personal identification method according to claim 1 based on mobile phone pick-up head combining with human face recognition technique, it is characterized in that, described principal component analysis (PCA), be meant by extracting the main feature of data set and reduce the data set dimension, it is a linear transformation, in data projection to a new linear space, in new linear space according to how many sorting data proper vectors from high to low of information content, select the higher low volume data of information content, form the new data set of the low dimension that has kept the most information that former data set comprised;

The linear transformation expression formula of principal component analysis (PCA) is as follows:

y＝Ψ ^Tx

Wherein: x is a vector before the conversion, and Ψ is the principal component analysis (PCA) transformation matrix, and T represents transposition, and y is the principal component analysis (PCA) result vector;

Wherein, definite method of principal component analysis (PCA) transformation matrix Ψ is specific as follows:

If data set X, wherein total N M dimension data x _i, i=1...N, N＜＜M, data set X is expressed as the matrix of a M * N, be referred to as matrix X, establish Φ=[φ ₁, φ ₂... φ _M], φ wherein _i(i=1...M) be the M dimensional vector, so Φ is M * M square formation, then has:

∑ _Xφ _i＝λ _iφ _i

Wherein: λ _iBe eigenwert, Σ _XBe the covariance matrix of matrix X, ∑ _X=E[(X-E[X]) (X-E[X]) ^T] be φ _i(i=1...M) be the covariance matrix ∑ of X _XCorresponding to eigenvalue _iProper vector;

With eigenvalue _i(i=1...M) be λ according to ordering from big to small ₍₁₎, λ ₍₂₎... λ _(M), its characteristic of correspondence vector is respectively φ ₍₁₎, φ ₍₂₎... φ _(M), then

${\mathrm{\Ψ}}^T=\left\{\begin{array}{c}{\mathrm{\φ}}_{\left(1\right)}^T\\ {\mathrm{\φ}}_{\left(2\right)}^T\\ \·\\ \·\\ \·\\ {\mathrm{\φ}}_{\left(300\right)}^T\end{array}\right\}$

Ψ=(Ψ ^T) ^TBe the principal component analysis (PCA) transformation matrix.

5. the personal identification method based on mobile phone pick-up head combining with human face recognition technique according to claim 1, it is characterized in that, described linear appreciation analysis, being meant by extracting the categories of datasets distinguishing characteristic makes the data set dimension reduce, the same with principal component analysis (PCA), it is a kind of linear transformation, linear appreciation is analyzed in N class data projection to the new linear space, make in this new linear space that Euclidean distance is nearest between the different sample datas in the same classification, the Euclidean distance between the inhomogeneity farthest simultaneously;

The linear transformation expression formula that linear appreciation is analyzed is as follows:

z＝W ^Ty

Wherein: z is the human face recognition vector, and y is the result vector of principal component analysis (PCA), is in the human face recognition linear space, and W is linear appreciation analytic transformation matrix;

Wherein: definite method of linear appreciation analytic transformation matrix W is as follows:

If data set Y has N kind C _i, i=1...N, data are the M dimension, and each class data comprises 1 or a plurality of sample y _i, i=1,2.. has for this N class problem,

$S_w^{-1}{S}_{b}W={\mathrm{\Λ}}_{w}W$

Wherein: Λ _wBe eigenvalue matrix, W is linear appreciation analytic transformation matrix, is matrix S _w ^-1S _bCorresponding to Λ _wEigenvectors matrix, S _wBe class inscattering matrix, characterize the situation of data variance in the same class,

$S_w=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left\{\mathrm{Pr}\left(C_i\right)E\right[{(y-E\left(y_i\right))(y-E\left(y_i\right))}^T|C=C_i]\}$

And S _bBe scattering matrix between class, characterize data variance distribution situation between inhomogeneity,

$S_b=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\left\{\mathrm{Pr}\left(C_i\right)(E\left(y_i\right)-E\left(y\right)){(E\left(y_i\right)-E\left(y\right))}^T\right\}$

Pr (C in wherein above-mentioned two formulas _i) be class C _iPrior probability, expression formula is:

$\mathrm{Pr}\left(C_i\right)=\frac{1}{N},(i=1...N)$

The prior probability of each class is equal,

By obtaining matrix S _w ^-1S _bEigenvectors matrix and eigenvalue matrix, uniquely determine linear appreciation analytic transformation matrix W.

6. the personal identification method based on mobile phone pick-up head combining with human face recognition technique according to claim 1 is characterized in that, described human face recognition transformation matrix H is specially: H ^T=W ^TΨ ^T,

Human face recognition is transformed to: z=H ^Tx

Wherein: Ψ and W are respectively principal component analysis (PCA) transformation matrix and linear Identification analytic transformation matrix, and x is input facial image vector, and z is the human face recognition vector.

7. a kind of personal identification method based on mobile phone pick-up head combining with human face recognition technique according to claim 1 is characterized in that, the oval complexion model of described utilization is handled the facial image that mobile phone end transmits, and comprises following concrete steps:

1. the rgb value of each pixel in the facial image that mobile phone end is collected (R, G B) convert to (Y, Cb, Cr) value, wherein, Y represents that monochrome information, Cb represent that blue colour difference information, Cr represent red colour difference information;

2. the parameter of oval complexion model is set;

3. to each pixel in the image, carry out skin color segmentation, obtain black and white two-value broca scale picture;

4. be communicated with in the border of area of skin color at four of black and white two-value broca scale picture, obtain the area of skin color positional information.

8. a kind of personal identification method according to claim 1 based on mobile phone pick-up head combining with human face recognition technique, it is characterized in that, described people's face cascade sort detecting device, be that linear combination by Weak Classifier obtains a strong classifier, again the strong classifier combination is obtained a cascade classifier, utilize the repeatedly classification of cascade to improve the accuracy that detects, the Weak Classifier that is used to detect, and utilize self-adapting regulation method training to obtain by the strong classifier that weak typing linear combination obtains, people's face cascade sort detecting device that final training is come out comprises 20 strong classifiers, people's face cascade sort detecting device is by reading the classifier data file that trains, construct Weak Classifier successively, strong classifier and cascade classifier, whether Weak Classifier directly provides according to the input rectangular area is the classification results of people's face, strong classifier is determined classification results according to a plurality of Weak Classifier results' weighting, cascade classifier is obtained by the cascade of multilayer strong classifier, if certain strong classifier is non-face with this territorial classification, then cascade classifier thinks that detected zone is non-face, direct detection of end, otherwise, cascade classifier is sent data into next strong classifier classification, when having only all strong classifiers to think that all this zone is for human face region, people's face cascade sort detecting device just provides the zone and is the classification results of people's face.

9. a kind of personal identification method based on mobile phone pick-up head combining with human face recognition technique according to claim 1 is characterized in that, the described human face recognition transformation matrix H that generates that uses is respectively to comparison people face F _iAnd people's face F to be identified _InputCarry out the human face recognition linear transformation, and generate corresponding human face recognition vector v _iAnd v _Input, specific as follows:

v _i＝H ^TF _i

v _input＝H ^TF _input

Wherein, F _iBe the people's face that from database, takes out, F _InputBe people's face to be identified that mobile phone end is uploaded, above-mentioned two formulas are respectively to F _iAnd F _InputCarried out the human face recognition conversion, v _iAnd v _InputBe corresponding human face recognition vector, final recognition of face will be undertaken by comparing these two vectors.

10. a kind of personal identification method based on mobile phone pick-up head combining with human face recognition technique according to claim 1 is characterized in that, described calculating human face recognition vector v _iAnd v _InputEuclidean distance d _{I, input}, specific as follows:

$d_{i,\mathrm{input}}=\sqrt{\underset{j=1}{\overset{300}{\mathrm{\Σ}}}{(v_{i,j}-v_{\mathrm{input},j})}^2}$

Wherein: v _{I, j}And v _{Input, j}It is respectively the human face recognition vector v _iAnd v _InputJ component, d _{I, input}It is vector v _iAnd v _InputEuclidean distance, this is apart from characterizing vector v _iAnd v _InputSimilarity degree, d _{I, input}More little then similarity degree is high more.

Images