CN103914904A - Face identification numbering machine - Google Patents

Abstract

The invention relates to a face identification numbering machine. The numbering machine comprises a numbering business trigger module used for triggering a numbering process according to external input, a certificate information acquisition module used for acquiring the certificate information according to the trigger of the numbering business trigger module, a person image acquisition module used for acquiring a photograph of a certificate holder according to the trigger of the numbering business trigger module, a person image analysis module used for extracting corresponding face characteristic values from a certificate photograph and the certificate holder photograph, and a person image comparison module used for comparing the face characteristic value of the certificate photograph and the face characteristic value of the certificate holder photograph, and determining the comparison result is allowed and printing a number if the similarity between the face characteristic value of the certificate photograph and the face characteristic value of the certificate holder photograph reaches or exceeds a preset first threshold, and determining the comparison result is unallowed and starting a verification failure process if the similarity is less than a second threshold.

Description

Recognition of face queue machine

Technical field

Automatic face recognition of the present invention field, particularly a kind of recognition of face queue machine.

Background technology

Queue machine claims again queue machine, number calling machine, number ticket machine, the number of getting machine, sends out number machine.Queue machine full name is: queue management system, or claim electronic queuing system, intelligent queue system, computer calling system.The queue machine developing history of existing more than 30 year, is widely used in the service field that need to wait in line.What require along with social development with to public order is growing, and the business handling work of a lot of public industries has all adopted the queue machine combining with touch-screen, metal cabinet, printer one to carry out on-the-spot order management.Queue machine is generally applicable to service industry's working hall as unit Zero queuings such as finance, postal service, hospital, the tax, communication, visa, industry and commerce, social security center, insurances, so not only can effectively improve service environment, but also can increase work efficiency better.

Although queue machine can significantly be increased work efficiency, the on-the-spot order management work of applying unit has obtained very large improvement, but find in actual use, existing queue machine function singleness, only can play the function of order keeping, but have the industry of requirement but can not meet user's actual needs for finance, hospital, the governmental affairs etc. to the person's of handling identity.Particularly, for example, existing queue machine is difficult to prevent that same people from getting multiple queue numbers, and law-breaker is had an opportunity to take advantage of.Even if install authentication system additional on existing queue machine, for example install identity card reading device additional, to guarantee that same I.D. can only get a queue number, but still can not take precautions against the hand-held multiple I.D.s of same people to reach the way of getting multiple queue numbers.Therefore, existing queue machine is difficult to guarantee that " number (queue number), card (identity document), people's (holder) " is consistent, has left hidden danger to public order and public interest.

Summary of the invention

Technical matters to be solved by this invention is for a kind of recognition of face queue machine is provided.

For solving the problems of the technologies described above, the present invention realizes as follows:

A kind of recognition of face queue machine, comprising:

Numbering service trigger module, for according to outside input, triggers row number flow process;

Certificate information collection module, for according to the triggering of numbering service trigger module, gathers certificate information;

Human image collecting module, for according to the triggering of numbering service trigger module, gathers holder photo;

Portrait analysis module is for extracting corresponding face characteristic value from described certificate photograph and holder photo;

Portrait comparing module is for comparing the face characteristic value of the face characteristic value of described certificate photograph and holder photo, if the similarity of the face characteristic value of the face characteristic value of described certificate photograph and holder photo meets or exceeds default first threshold, judge that comparison result is as passing through; If lower than Second Threshold, judge that comparison result is as not passing through, and start authentication failed flow process;

Print module is for print queue's number in the time that portrait comparing module judges that contrast is passed through; With

The device system of calling out the numbers for judge in portrait comparing module contrast by time include this holder in row number queue, and call out the numbers in the time taking turns to the queue number of this holder.

Good effect of the present invention:

Recognition of face queue machine of the present invention, use image technique, card reader of ID card, database system the technical approach such as to transfer and obtain portrait and the identity information in all kinds of lawful documents that personnel provide, and compare with itself and external camera figure information dynamic or that static state is obtained, reach the target of " number witness is consistent ".

Accompanying drawing explanation

Fig. 1 is the structural representation block diagram of face identification queue machine of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is further detailed explanation.

Please refer to Fig. 1, Fig. 1 is the structural representation block diagram of face identification queue machine of the present invention, and this recognition of face queue machine comprises numbering service trigger module, certificate information collection module, human image collecting module, portrait analysis module, portrait comparing module, database storage backup module, print module and the device system of calling out the numbers.

This numbering service trigger module is connected to certificate information collection module and human image collecting module, certificate information collection module and human image collecting module are connected to portrait analysis module, portrait analysis module is connected to portrait comparing module, and portrait comparing module is connected respectively to database storage backup module, print module and the device system of calling out the numbers.

Numbering service trigger module, for according to outside input, triggers row number flow process.Numbering service trigger module can comprise the input equipments such as touch-screen, touch key-press, mechanical key, keyboard, mouse.

Certificate information collection module, for according to the triggering of numbering service trigger module, gathers certificate information.Certificate can be the effective identity certificates such as first generation I.D., second generation identity card, officer's identity card or passport.Certificate information comprises certificate photograph information, certificate Word message or certificate number information.Certificate information collection module can read kind equipment by certificates such as card reader of ID card, camera, video camera, optical character identification and form.

Human image collecting module, for according to the triggering of numbering service trigger module, gathers holder photo.Human image collecting module can be camera, video camera etc.

Portrait analysis module is for extracting corresponding face characteristic value from described certificate photograph and holder photo.

Portrait comparing module is for comparing the face characteristic value of the face characteristic value of described certificate photograph and holder photo, if the similarity of the face characteristic value of the face characteristic value of described certificate photograph and holder photo meets or exceeds default first threshold, judge that comparison result is as passing through; If lower than Second Threshold, judge that comparison result is as not passing through, and start authentication failed flow process.Authentication failed flow process can comprise re-execute row number flow process or prompting row number not by or adjust first threshold or adjust Second Threshold or trigger staff and get involved and carry out manual intervention.Wherein, first threshold and Second Threshold can be identical, also can be not identical.Described portrait comparing module can also be used for comparison record to be sent to described data memory module.

Database storage backup module is for storing the related data of holder.Concrete, if system without this holder database, system can be according to behind the certificate information building database newdata storehouse of holder, certificate information, certificate photograph, this comparison scene photograph, this comparison result being stored.If system has this holder database, this comparison result and holder scene photograph are stored.The data message that data memory module is stored can be inquired about and be transferred by manual intervention module, if authority allows (authority setting completes in manual intervention module), also can be in the time of comparison next time of this holder this scene photograph information of this holder be sent to people and compares module and compare.

Print module is for print queue's number in the time that portrait comparing module judges that contrast is passed through.

The device system of calling out the numbers for judge in portrait comparing module contrast by time include this holder in row number queue, and call out the numbers in the time taking turns to the queue number of this holder.

Described portrait analysis module, portrait comparing module can be made up of computer system.Or formed by the processor such as FPGA, DSP or electronic circuit.Described database storage backup module can be made up of hard disk, CD or other storeies.

In an optional implementation, described certificate information collection module comprises identity card reader identification module, OCR module and certificate information identification module.Described identity card reader identification module and optical character identification OCR module are connected respectively to certificate information identification module, and certificate information identification module is connected with numbering service trigger module and portrait analysis module respectively.

Described identity card reader identification module is for reading the information that in certificate, chip comprises.

Described OCR module is for obtaining the Word message on certificate.

Described certificate information identification module is used for according to the triggering of described numbering service trigger module, control the collection of identity card reader identification module and OCR module, and be sent to portrait analysis module by controlling the data that identity card reader identification module and OCR module collect.

In an optional implementation, described human image collecting module comprises shooting taking module and dynamic human image collecting module.Described shooting taking module is for taking to obtain holder photo to holder.Described shooting taking module can be camera, video camera etc.Described dynamic human image collecting module, for according to the triggering of described numbering service trigger module, is controlled the collection of described shooting taking module, and the described holder photo collecting is sent to portrait analysis module.Described dynamic human image collecting module can be by realizations such as computer, processor, single-chip microcomputer, FPGA, logical circuits.

In an optional implementation, described in the device system of calling out the numbers comprise numbering service storehouse module, the management of calling out the numbers module, loudspeaker, display screen and the device of calling out the numbers.Numbering service storehouse module is connected with portrait comparing module, and the management of calling out the numbers module is connected with numbering service storehouse module, and loudspeaker, display screen and the device of calling out the numbers are connected with the management module of calling out the numbers respectively.

Numbering service storehouse module for portrait comparing module judge contrast by time include this holder in row number queue, with to row number queue manage.

Call out the numbers device for receiving operating personnel's control, call out the numbers or queue management arranges instruction to the management module input of calling out the numbers;

The management of calling out the numbers module is for according to the row number information of numbering service storehouse module output, or according to the described device output of calling out the numbers call out the numbers or queue management arranges instruction, control loudspeaker and display screen and call out the numbers.

In an optional implementation, described in the management module of calling out the numbers may further include voice module, LTE control module, word processing module and hardware driving.

Structure, function and the principle of work of face identification queue machine of the present invention are more than introduced.Below in conjunction with another concrete implementation, recognition of face queue machine of the present invention is carried out to concrete introduction.But, below concrete implementation be a kind of exemplary introduction, be optional for recognition of face queue machine of the present invention.

1, business trigger module: recognition of face queue machine homepage is such as, according to business function kind (individual business, business event, the type of credential etc. of the customization of practical business demand, business trigger module herein can be self-defined according to demand) and the touch business trigger module of setting, Service Trigger Information can be sent to " certificate information identification " module by this module, thereby make work flow arrive next functional module place.

Hardware is realized: touch-screen+computer

2, certificate information identification module: its major function is to receive after the triggering command that " business trigger module " send, according to the certificate information mark in instruction, judgement and processing send to " identity card reader identification module " still " optical character identification module ", if be designated China second-generation identity card mark and Service Trigger Information will be sent to " identity card reader identification module " and carry out the acquisition process of certificate information, for example, if being designated is other class certificate (passports, the certificates such as officer's identity card) Service Trigger Information will be sent to " optical character identification module " and carry out the acquisition process of certificate information.And the information flow that " identity card reader identification module " and " optical character identification module " gathered is to next functional module place.

Hardware is realized: computer

3, identity card reader identification module: its major function is to utilize API/SOCKET/WEBSERVICE supervisor interactive mode, carry out the mutual of China second-generation identity card part information by usb data transmission mode and China second-generation identity card card reader, thereby obtain the essential information (as information such as information, name, sex, certificate number, native places) of China second-generation identity card

Hardware is realized: China second-generation identity card card reader+computer

4, optical character identification module (OCR identification module): its major function is by the character of printing on electronic equipment (digital camera) examination of document, by detecting its shape of dark, bright mode decision, then with character identifying method, shape is translated into the discernible information of program (comprising the information such as word and portrait).

Hardware is realized: digital camera (image capture device connecting by network/USB etc.)+computer

5, portrait analysis module: its major function receives certificate information and the business information that " certificate information identification module " transmits, the analysis of the information of carrying out, detach and arrange (comprising certificate essential information, business information, figure information etc.), and trigger " dynamically human image collecting module " and obtain after dynamic figure information, certificate figure information and dynamically figure information and satellite information circulate " portrait comparing module ".This module Core Feature is to use face recognition algorithms figure information to be calculated to and draw portrait eigenwert.

Hardware is realized: computer

6, dynamic human image collecting module: its major function is to receive " portrait analysis module " to call dynamic human image collecting instruction, and complete and " shooting taking module " make a video recording shooting instruction the shooting of carrying out alternately dynamic portrait, and receiving the dynamic portrait that " shooting taking module " taking module returns, " portrait analysis module " given in circulation simultaneously.

Hardware is realized: computer

7, shooting taking module: its major function receives " dynamically human image collecting module " dynamically human image collecting instruction, utilize API/SOCKET/WEBSERVICE supervisor interactive mode and digital camera to carry out human image collecting information interaction simultaneously, when getting after the dynamic figure information of digital camera collection, this module is responsible for circulating dynamic figure information to " dynamically human image collecting module ".

Hardware is realized: digital camera (image capture device connecting by network/USB etc.)+computer

8, portrait comparing module: its major function receives portrait comparison instruction and the satellite information (comprising certificate essential information, business information, certificate figure information, dynamic figure information etc.) of " portrait analysis module " circulation, complete and must compare certificate figure information and dynamic figure information eigenwert, associated corresponding satellite information simultaneously, if compared successfully, send print command to print module print queue strip, if compared unsuccessfully, send circulation take instruction to " dynamically human image collecting module " from newly obtaining dynamic figure information.

Hardware is realized: computer

9, print module: its major function receives print command and the type information that " portrait comparing module " sends, prints row number strip by the print protocol of standard.

Hardware is realized: printer (heat-sensitive type, ink jet type, laser type, pin type etc.)+computer

10, numbering service storehouse: the row number storehouse that its major function receives " portrait comparing module " circulation inserts instruction and satellite information (comprising business information, certificate figure information, dynamic figure information, row number information etc.), and increase with " administration module of calling out the numbers ", delete, change, look into, information transfer instruction mutual, realize the row number information in storehouse increasing, delete, change, look into, the management function such as transfer.

Hardware is realized: computer

11, the management module of calling out the numbers: it comprises (" voice module ", " LED control module ", " word processing module " and " Hardware drive module "), its major function receives " numbering service storehouse " information interaction command information (comprising that information mind immediate skip, information detect instruction etc.), coordinate the realization of " voice module ", " LED control module ", " word processing module " and " Hardware drive module " each functions of modules, finally complete station-to-station service flow process.

Hardware is realized: device+LED screen+speech ciphering equipment+computer of calling out the numbers

12, the device of calling out the numbers: its major function is to operate by operating personnel the device function key of calling out the numbers, to " management of calling out the numbers module " send call out the numbers or queue management arrange instruction (comprise exit, along exhaling, the instruction such as recall, transition window, Priority Call, callback, the number of abandoning, inquiry) realize the actual function of calling out the numbers.

Hardware is realized: the device+computer of calling out the numbers

13, voice module: the row number information that instruction that what its major function received that " management of calling out the numbers module " coordinated flow turns over call out the numbers and " management of calling out the numbers module " are come alternately from numbering service storehouse, by speech conversion program, some attribute (for example row number order, sales counter number etc.) is realized to speech conversion, and this voice messaging is passed to speech ciphering equipment and carry out voice broadcast.

Hardware is realized: loudspeaker (loudspeaker, power amplifier etc.)+computer

14, word processing module: the row number information that instruction that what its major function received that " management of calling out the numbers module " coordinated flow turns over call out the numbers and " management of calling out the numbers module " are come alternately from numbering service storehouse, by copy editor's program, some attribute (for example row number order, sales counter number etc.) is realized to text conversion, and this Word message is coordinated to be given to " LED control module " by " management of calling out the numbers module " and carry out the word of LED screen and show.

Hardware is realized: computer

15, LED control module: its major function receives the word demonstration steering order that " management of calling out the numbers module " coordinated flow turns over, and by " Hardware drive module " bottom hardware control protocol information interaction, realizes the demonstration of Word message on LED screen.

Hardware is realized: computer

16, Hardware drive module: the functional module in its major function realization " management of calling out the numbers module " and mutual (external equipment includes LED screen controller, power amplifier, the device etc. of calling out the numbers) of the control of external hardware equipment bottom hardware and transfer instruction, with the communication of practical function module and bottom physical equipment, the realization of the final function of supporting to call out the numbers.

Hardware is realized: computer

Below in conjunction with the signaling/data interaction flow process between each module in Fig. 1, introduce the present embodiment recognition of face queue machine principle of work.

1. involved data and signaling in:

1, the mutual signaling of touch-screen and system program: while being arranged on the touch-screen of display front end with finger or the touch of other objects, the position (with coordinate form) touching is detected by touch screen controller, and deliver to CPU by interface (as RS-232 serial port), thereby determine the information of input

2, internal module calling data: " numbering service trigger module " is transferred to " certificate information identification module " by the interface interchange instruction with business information by modes such as XML/JSON/ binary streams and carries out certificate identification, and the rreturn value of reception " certificate information identification module " module.

2. involved data and signaling in:

1, internal module interaction data: mainly contain data interaction instruction, the data of transmitting by modes such as XML/JSON/ binary streams comprise certificate essential information, business information, figure information, Service Trigger Information etc.

3. involved data and signaling in:

1, the internal mode call instruction of determining: " portrait analysis module " transmits dynamic acquisition triggering command to " dynamically human image collecting module " by modes such as XML/JSON/ binary streams, and wait for and receive rreturn value.

4. involved data and signaling in:

1, internal module interactive instruction: " dynamically human image collecting module " transmits dynamic acquisition triggering command to " shooting taking module " by modes such as XML/JSON/ binary streams, and wait for reception rreturn value.

5. involved data and signaling in:

1, the mutual signaling of video camera and system program: mainly contain " shooting taking module " by connected modes such as USB/RJ45/COM mouths, the agreements such as the USB/TCP/IP/ serial ports by standard, realize the instruction (comprising the photographing instruction of bottom etc.) and data (data comprise picture etc.) of interactive information between external camera.

2, internal module calling data: mainly contain " shooting taking module " and return to the rreturn value of " dynamically human image collecting module ", the data that data interaction is transmitted by modes such as XML/JSON/ binary streams comprise certificate essential information, business information, figure information etc.

6. involved data and signaling in:

1, internal module interaction data: mainly contain the rreturn value that " dynamically human image collecting module " returns to " portrait analysis module ", the data that data interaction is transmitted by modes such as XML/JSON/ binary streams comprise involved data and signaling in certificate essential information, business information, figure information etc.:

1, internal module interaction data: the main portrait comparison instruction of transmitting by modes such as XML/JSON/ binary streams containing " portrait analysis module " and satellite information (comprising certificate essential information, business information, certificate figure information, dynamic figure information etc.).

8. involved data and signaling in:

1, the mutual signaling of printer and system program: mainly contain " portrait comparing module " by connected modes such as USB/RJ45/COM mouths, the agreements such as the USB/TCP/IP/ serial ports by standard, the instruction (the printing interactive instruction of standard) that realizes interactive information between external printer and data (data comprise picture, arrange in numerical order sequence number, business information etc.).

Internal module calling data:

1, " portrait comparing module " circulates the row number storehouse insertion instruction of circulation and satellite information (comprising business information, certificate figure information, dynamic figure information, row number information etc.) to " numbering service storehouse " by modes such as XML/JSON/ binary streams.

2, " portrait comparing module " database interactive interface by standard databases such as () SqlServer, ORALE, carries out storage and management by the data of all formation.

9. and in involved data and signaling:

Internal module interaction data: " numbering service storehouse " and " administration module of calling out the numbers " band by the modes such as XML/JSON/ binary stream complete increasings, delete, change, look into, the instruction such as information transfer mutual, and the realization function of calling out the numbers.

10. and in involved data and signaling:

1, the mutual signaling of device and system program of calling out the numbers: mainly contain " administration module of calling out the numbers " by connected modes such as USB/RJ45/COM mouths, the agreements such as the USB/TCP/IP/ serial ports by standard, realize and the instruction of the interactive information between device of calling out the numbers (comprise exit, along exhaling, the instruction such as recall, transition window, Priority Call, callback, the number of abandoning, inquiry).

in involved data and signaling:

1, the mutual signaling of loudspeaker and system program: the phonetic control command of main transmission and speech data (row number order, sales counter number etc. speech data).

in involved data and signaling:

1, the mutual signaling of LED control card and system program: the word of main transmission shows steering order.

Portrait analysis module 14 in the present invention can utilize face recognition technology comparison film to identify, and obtains face characteristic value.Illustrate optional face recognition technology below.

1 face identification method based on geometric properties

Method based on geometric properties is one of early stage face identification method.The face that these class methods are utilized face are if the local shape feature of eyes, nose, face etc. and these face features are at the geometric properties distributing on the face.In the time cutting apart, obtain face feature, often to use some prioris of human face structure.It identifies required feature is generally as basic eigenvector, to be the coupling between eigenvector take shape and the geometric relationship (as indexs such as the Euclidean distance between face feature, curvature, angles) of human face in essence.

2 template matching methods

The related operation that face sample in the facial image of input and training set is normalized one by one, what have optimum matching is recognition result.

3 based on statistics method

Method based on statistics is generally made facial image as a whole, represents with a vector in higher dimensional space, and like this, recognition of face problem is converted into and in higher dimensional space, finds the problem of separating hypersurface (plane).If separation is linear method of lineoid, if separation is that hypersurface is called nonlinear method.By training sample is obtained with statistical technique and separate hypersurface (plane).The method of conventional some based on statistics comprises intrinsic face method (Eigenfaces), Fishe face method (Fisherfaces), independent component analysis (ICA), local retaining projection (LPP), hidden Markov model (HMM), support vector machine (SVM), nuclear technology etc.

3.1 intrinsic face methods

Supposing has N image in facial image database, with vector representation be X1, X2 ..., XN (vectorial dimension is made as L), its face the average image is the inequality that can obtain thus every width image is

X′ _i=X _i－X _ave；i=1，2，3…N； (1)

Can calculate like this covariance matrix:

$C=\frac{1}{N}{\mathrm{\Σ}}_{i=1}^k{X}_i^{\′}{\left(X_i^{\′}\right)}^T.---\left(2\right)$

The eigenvalue K of compute matrix C _kwith corresponding latent vector U _k. the vector space that these latent vectors of obtaining form, just can represent the principal character information of facial image. all N image in facial image database, all to this space projection, is obtained to projection vector separately

$\begin{array}{c}{Y}_1,Y_2,\·\·\·,Y_N,{\left(Y_i\right)}^T=[y^{i1}{y}^{i2}\·\·\·y^{\mathrm{iL}}],i=\mathrm{1,2},\·\·\·,N,y^{\mathrm{ij}}={\left(\mathrm{uj}\right)}^T{X}_i^{\′},j=1,\\ 2,\·\·\·,L.---\left(3\right)\end{array}$

For facial image X to be identified, by calculate its with Xave differ from projection vector Y:

$y^{\mathrm{ij}}={\left(\mathrm{uj}\right)}^T{X}_i^{\′},j=\mathrm{1,2},\·\·\·,L.---\left(4\right)$

The projection vector Y corresponding with the facial image of N in facial image database again ₁, Y ₂..., Y _nrelatively, complete identification according to certain distance criterion. as adopted Euclidian distance, calculate e _i=|| Y-Y _i||, i=1,2 ..., it is n pattern that N. identifies facial image, $n=\underset{i}{\arg }$ $\left(\underset{i=\mathrm{1,2},\·\·\·N^{\left(e^{\′}\right)}}{\min }\right).---\left(5\right)$

In actual computation, the size of Matrix C is L × L, even also very large to less its value of image of size. for example image is 24 × 8 sizes, and the large young pathbreaker of Matrix C is that the inequality of every width image is formed a matrix by (24 × 28) 2 ≈ 4.5 × 105.:

X′=[X ¹，X ²，…，X ⁿ]， (6)

Covariance matrix can be write as

$C=\frac{1}{N}{X}^{\′}{\left(X^{\′}\right)}^T.---\left(7\right)$

According to linear algebra theory, will calculate x ' (x ') ^teigenvalue K _kwith corresponding latent vector U _kproblem be converted into and ask (x ') ^tthe eigenvalue K of x _kwith corresponding latent vector V _k. (x ') ^tthe size of x is only N × N, generally all much smaller than L × L, therefore simplified calculating. obtaining V _kafter, U _kcan be obtained by following formula:

$U^k=\frac{1}{k^k}{X}^{\′}{V}_k---\left(8\right)$

3.2Fishe face method (Fisherfaces)

Suppose to have a set H to comprise N d dimension sample x ₁, x _z... x _n, wherein N1 sample that belongs to ω 1 class is designated as subset 1, and N2 sample that belongs to ω 2 classes is designated as H ₂.If to N _xcomponent do linear combination and can obtain scalar:

y _n=w ^Tx _n，n＝1，2…N _i. (1)

So just, obtain an one dimension sample y _nthe set of composition, and can be divided into two subset Y ₁with Y ₂.From set, if || w||=1, each y _nbe exactly corresponding x _nprojection on the straight line that is w to direction.In fact, the absolute value of w is inessential, and he only makes y _nbe multiplied by a scale factor, importantly select the direction of W.The direction difference of W, by the separable degree difference making after sample projection, thereby directly affects recognition effect.Therefore, the so-called problem of finding best projecting direction is exactly to find best conversion vector W on mathematics ^*problem.

First define the basic parameter of several necessity below to facilitate narration.

3.2.1 at d dimension space

(1) Different categories of samples mean vector M _i

$m_i=\frac{1}{N_i}{\mathrm{\Σ}}_{x\∈H_i}x,i=\mathrm{1,2}.---\left(2\right)$

(2) matrix within samples s. and total within class scatter matrix S _w

$s_i={\mathrm{\Σ}}_{x\∈H_i}(x-m_i){(x-m_i)}^T,i=\mathrm{1,2}.s_i=s_1+s_2---\left(3\right)$

(3) matrix between samples S _b

S _b＝(x－m _i)(x－m _i) ^T (4)

Wherein S _wsymmetrical positive semidefinite matrix, and normally nonsingular in the time of N>d.S _balso be symmetrical positive semidefinite matrix,

Under two class conditions, its value is more than or equal to 1.

3.2.2 in one dimension Y space

(1) Different categories of samples average

$\widetilde{m_i}=\frac{1}{N_i}{\mathrm{\Σ}}_{y\∈Y_i}y,i=\mathrm{1,2}---\left(5\right)$

(2) within-class scatter with dispersion in total class

$\stackrel{~}{s_i^2={\mathrm{\Σ}}_{v\∈Y_i}{(y-\widetilde{m_i})}^2},i=\mathrm{1,2}.\widetilde{s_w}=\widetilde{s_1^2}+\widetilde{s_2^2}---\left(6\right)$

Define now Fisher criterion function.After projection, in one dimension Y space, Different categories of samples is separated as far as possible, wishes the poor of two class averages) be the bigger the better; Wish that Different categories of samples inside is as far as possible intensive simultaneously.Wish that in class, dispersion is the smaller the better.Therefore, definition Fisher criterion function:

Should find J _f(w) molecule is large as far as possible and denominator is as far as possible little, is namely J _f(w) large as far as possible w is as projecting direction.But above formula is not aobvious containing w, therefore must manage J _f(w) become the explicit function of w, from definition can release:

$\widetilde{m_i}=\frac{1}{N_i}{\mathrm{\Σ}}_{y\∈H_i}y=\frac{1}{N_i}{\mathrm{\Σ}}_{y\∈H_i}{w}^{T}x=w^T\left(\frac{1}{N_i}{\mathrm{\Σ}}_{y\∈H_i}x\right)=w^T{m}_i---\left(8\right)$

Like this, J _f(w) molecule just becomes:

${(\widetilde{m_1}-\widetilde{m_2})}^2={(w^T{m}_1-w^T{m}_2)}^T=w^T(m_1-m_2){(m_1-m_2)}^{T}w=w^T{S}_{b}w---\left(9\right)$

Ask below and make J _f(w) get maximum value w ^*.Solve by Lagrange multiplier method.Another denominator equals non-zero constant, and definition lagrange function is: L (w, δ)=w ^ts _bw-δ (w ^ts _bw-c) in (10) formula, δ is Lagrange multiplier.Above formula is asked to local derviation: making partial derivative is zero: S _bw ^*=δ S _ww ^*(12)

Wherein w ^*make exactly J _f(w) maximum minimax solution.Because S _wnonsingular, both sides are premultiplication simultaneously can obtain:

$S_w^{-1}{S}_b{w}^{*}=\mathrm{\δ}{w}^{*}---\left(13\right)$

General matrix is asked in this actual listing eigenvalue problem.Utilize S _bdefinition, above formula can be rewritten as:

S _bw ^*＝(m ₁－m ₂)(m ₁－m ₂) ^Tw ^*＝(m ₁－m ₂) ^R (14)

3.3 support vector machine methods:

Support vector machine (Suppoa Vector Machine, SVM) method is a kind of new mode identification method growing up on the basis of Statistical Learning Theory, it is the method based on structural risk minimization principle, to the more insoluble problems of the artificial neural network based on empirical risk minimization, as: Model Selection and cross problem concerning study, non-linear and dimension disaster problem, local minimum point's problem etc. have all obtained solution to a great extent.But directly use SVM method to carry out recognition of face and have the difficulty of two aspects: the one, training SVM need to solve quadratic programming problem, and computing time, complexity and space complexity were all higher; The 2nd, in the time that non-face sample is unrestricted, need the training sample set of great scale, the support vector obtaining can be a lot, make the calculated amount of sorter too high.

For the research of these problems, many new methods are there are, SMO (Sequential MinimalOptimization) algorithm proposing as Platt has solved first problem effectively, the people such as Osuna have used a large amount of face samples in training, adopt the method for bootstrapping to collect " non-face sample; and adopt optimization method to reduce the quantity of support vector, solve to a certain extent Second Problem; People's face detection algorithm that the employing template matches such as Liang Luhong combine with SVM method, in the subspace limiting in template matches, adopt the method collection " non-face sample " of bootstrapping to train SVM, the difficulty of training and the support vector scale finally obtaining are reduced, make detection speed improve 20 times than simple SVM detecting device, obtained the comparable result of neural net method with CMU.Richman etc. propose with the nasal area training SVM in face, reduce training data, and need not consider the impact of SVM on the jewelry such as hair style, glasses, gather image and also do not require that realization positions and normalized facial image, the method have been applied in the Real time face detection system of Kodak.

1.3.4 the method based on nuclear technology

" core skill " (Kemeltrick) J likes that be to propose in the research of support vector machine morning.Principal component analysis (r based on core, PCA) method and Fisher discriminatory analysis method (KH) A based on core) be the core popularization of PeA and LDA, Baudat and Anouar have proposed the KFD method for many classification problems, and MingHuangYang discusses the eigenface method and the Fisher face method that have compared based on core skill.The people such as JianYang have proposed the application framework of KPCA+KFD, kernel discriminant analysis under this framework can utilize two class authentication informations, on the kernel of one class scatter matrix (referring to implement scatter matrix in the class after KPCA conversion) in class, obtain, another kind ofly in class, in the non-kernel of scatter matrix, obtain.Gao Xiumei proposes core Foley.Sammon discriminatory analysis (core F-S discriminatory analysis, KFSD anvil method.The people such as Xu Yong choose a small amount of " significantly " training sample set from all training samples, and the feature extraction efficiency of kernel method is improved a lot.

The basic thought of kernel method be by the sample in former feature space by the Nonlinear Mapping of certain form, transform to an even infinite dimensional space of higher-dimension, and in new space, apply linear analytical approach by means of " core skill " and solve.Because the linear direction in new space is corresponding to the non-linear direction of former feature space, so the discriminating direction that the discriminatory analysis based on core draws is the non-linear direction of corresponding former feature space also, the discriminatory analysis based on core is a kind of Nonlinear Discriminant Analysis method of luv space.With respect to other nonlinear method, unique and the crucial part of this method is the inner product operation that it carries out between sample by means of the kernel function one of holding withg both hands dexterously, subsequently the core sample vector generating is carried out to corresponding linear operation and ask for discriminant vectors collection, carry out the form after Nonlinear Mapping and do not need to obtain primitive character space sample, make it be better than common Nonlinear Discriminant Analysis method.

4 methods based on model

Flexible Model about Ecology comprises active shape model (ActiveShapeModels, ASMs) and active apparent model (ActiveAppearance Models, AAMs).ASMs/AAMs is described facial image respectively with shape and texture two parts with PCA, and then further by PCA, the two is merged face is carried out to statistical modeling.Flexible Model about Ecology has good face synthesis capability, is therefore widely used in face characteristic registration (FaceAlignment) and identification.

What the people such as Georghiades proposed has obtained good effect based on illumination cone (Illumination Cones) model aspect the impact that overcomes multi-pose in recognition of face, complex illumination condition.The people such as Georghiades find: all images of same face under same visual angle, different illumination conditions form a convex cone in image space---be illumination cone.Illumination cone model can be under Lambertian model, nonreentrant surface and far point light source assumed condition, recover the 3D shape of object and the surface reflectance of surface point according to 7 of unknown illumination condition same visual point images, and traditional photometric stereo vision can could be recovered according to the image of 3 given known illumination conditions the normal vector direction of body surface, like this, the image that just can be easy to any illumination condition under synthetic this visual angle, completes the calculating of illumination cone.Identification is to complete to the distance of each illumination cone by calculating input image.

The face identification method based on 3D deformation model that Blanz and Vetter propose is being set up on the basis of 3D shape and texture statistics distorted pattern, perspective projection and the illumination model parameter of the method that simultaneously also adopts graphics simulation to image acquisition process carried out modeling, thereby can make the face built-in attributes such as people's face shape and texture separate completely with the external parameter such as camera arrangement, light conditions, more be conducive to analysis and the identification of facial image.

5 methods based on artificial neural network

Artificial neural network is simulation people's neural Operational Mechanisms and a kind of nonlinear method of proposing.That the earliest artificial neural network is applied to recognition of face work is Kohonen, is characterized in utilizing the associative ability of network to recall face.Subsequently, many different network structures are suggested.Ranganath and Arun have proposed the radial primary function network for recognition of face, the people such as Lin have proposed the neural network based on Probabilistic Decision-making for face detection, eyes location and recognition of face, Lee etc. have proposed the Fuzzy BP network for recognition of face, and Lawrence has proposed to entangle for the convolution nerve net of recognition of face.

The advantage of neural network is to obtain this this rule and regular covert expression by the process of study, and its adaptability is stronger.

6 elastic graph matching process

6.1. this type of a kind of the most successful method of elastic bunch graph coupling (ElasticBunchGraphMatching, EBGM) dike.It is based on dynamic linking structure (DLA, DynamisLinkArchitecture), with a banded attributes figure, face is described, wherein the summit of banded attributes figure is defined facial key feature points, and its attribute is generally that multiresolution, the multi-direction local feature one at the individual features point place that obtains by Gabor wavelet transformation is called Jet and represents; The attribute on limit is the geometric relationship between different key points.Whole identifying comprises locates predefined some facial key feature points to input facial image by a kind of Optimizing Search strategy, and extracts their Jet feature, obtains the attributed graph of input picture; Then the similarity of calculating face character figure in itself and storehouse judges classification.

Due to the dynamic perfromance of banded attributes figure, make this method there is higher robustness to attitude, expression shape change; And also there is certain general character with human visual system in the Jet feature of key point.But owing to needing the some facial key feature points of registration before identification, calculate relatively consuming time.

6.2. face location

We adopt the stacked detection of classifier face based on Adaboost statistical learning method face positioning stage.For the concrete condition in recognition of face, we select the maximum face detecting in image as face to be identified.

6.3. feature point extraction

In order to arrange the unique point in EGM, we need to extract 3 unique points, i.e. two eye center and a face center, and the eye center here not refers to pupil center, only refers to the center of eye areas, this is to consider that being difficult to robust is drawn into pupil center.We with reference to DAM (Direct Appearance Model) method 9], proposed a kind of Simple DAM algorithm and located these unique points.

In DAM method, mention between shape and texture, there is simple linear relationship: S=R*t+ ε

Wherein t is the projection in its principal component space through the face texture of certain correction, and s is the projection of shape in its principal component space.In our method, consider the simplest situation, only need 3 pairs of corresponding point, just can be by face proper non-front, be corrected to positive proper attitude.According to the method for DAM, we suppose, the face texture that face detection output is confined, and " between the vector of eyes and face " center " composition, there is the linear relationship of above formula in these three unique points on the face.Through training, we can find the mapping matrix of this linear relationship.Simple DAM arthmetic statement is as follows:

1. initialization current texture is the face texture that testing result is confined: t ← t ₀;

2. according to current texture, obtain the position of three unique points: S=R*t+ ε.If the position of three unique points and mean place are very approaching, finish;

3. according to the position of three unique points, whole picture (or comprising face and around an image-region) on apply affined transformation, by inclination face normalization; Again cut out a human face region according to the position of these three unique points and obtain new face texture, making current texture is the face texture after proofreading and correct; Forward 2 to.Because this method has considered the statistical relationship of unique point and texture to have very high robustness in itself, avoid method in the past only to process separately instability problem easily affected by noise according to piece image.

Because this method has considered the statistical relationship of unique point and texture to have very high robustness in itself, avoid method in the past only to process separately instability problem easily affected by noise according to piece image.

6.4. feature extraction

6.4.1.Gabor wave filter

In elastic graph matching algorithm, the unique point on face adopts Gabor wave filter to carry out feature extraction.Gabor kernel function

For:

(1)

Gabor wave filter is:

(2)

Wherein wave vector is:

wherein

(3)

Wherein coefficient of frequency V=0 ..., 4; Direction coefficient μ=0 ..., 7, form like this 40 related coefficients and describe in gray level image near the feature of neighborhood point.

Gabor small echo has following feature: the Section 2 of (1) bracket is removed DC component and made Gabor feature change and have robustness light intensity; The variation of contrast has robustness because small echo has carried out standardization; be Gauss function, this is actually the scope that has limited oscillating function by windowing, makes it effective in part, makes like this Gabor filtering can tolerate that image has certain distortion situation.

6.4.2. similar function

Gabor feature J to unique point:

J={J _jj wherein _j=a exp (D _j), j=0 ... 39

(4)

Consider how to measure the similarity between feature.

The similar function adopting at present has two kinds, and one is not consider angle, only considers amplitude, and relatively the inner product of two features, is called the irrelevant similar function of angle, is defined as follows

(5)

Another kind is the similar function of Angular correlation, is defined as follows

(6)

Wherein

(7)

Wherein

In our system, the similar function of Angular correlation has better performance.

6.5. face characteristic

In elastic graph matching process, there are three kinds of common face characteristic methods.The first is first to locate some human face characteristic points, then extracts the Gabor feature of these unique points, the face of limit common trait between these unique points and unique point, and wherein limit is used for carrying out topological constraints.The second is Wiskott[6 ］ propose first the structure of a similar storehouse of feature composition of the each unique point of same people in storehouse to be called to bundle (bunch), thereby the method that elastic graph coupling is developed into elastic bunch graph coupling (Elastic Bunch Graph Matching (EBGM)), the meaning of this method is to save system overhead.The third is because discovery does not need to locate especially accurately in recognition of face, even without topological constraints in the situation that, also can obtain the recognition effect of topological constraints, all right pick up speed ［ 5 ］ ［ 7 ］, thereby propose only to locate a small amount of unique point, such as only locating Liang Yanhezui center, generate on this basis the lattice of throwing the net, extract the Gabor character face of net point.It is good that experimental result in document ［ 7 ］ shows that the effect of the third method is come than the method with EBGM.Therefore, adopt herein document [7 ］ in method, face characteristic is as follows: adopt the grid of 10x10 as original mesh, first the 3rd row the 4th row of grid are decided to be to the position at left eye place, the 3rd row the 7th row are decided to be the position at right eye place, the position of mouth fixes on the 7th row, is then uniformly distributed on this basis other net point.

But the net point that can find out this 10x10 is not to be all distributed on the face.Have sub-fraction to be distributed in non-face region, some is distributed on facial contour, and along with the rotation of face can exceed human face region, some point is in human face region center.All inappropriate as unique point using these points, the point at least non-face region should foreclose, secondly the weight of each unique point should be different, likely can exceed human face region when the different attitude such as being distributed in point on facial contour, be also irrational if the point at they and human face region center has identical weight.Therefore to screen unique point, investigate their weight.We will screen lower joint and sequence 10x10 unique point.

6.6. characteristics of human body's sequence

Detect the class spacing of each unique point, with it measure the recognition capability of unique point.First we are using the primitive character net point of 10*10 as candidate feature point, each frame of video flowing is gathered to the Gabor feature of these 100 candidate feature points, each faceform in storehouse is calculated to similarity, the faceform who obtains highest similarity will obtain a ticket, and this result has comprised two aspects simultaneously.The one, Feature Selection, one is feature ordering.

With regard to Feature Selection, first in the process that a lot of unique points are rotated at face, the most of the time is in outside the scope of face, this must screen out, even the point secondly within the scope of face neither be used for eigenface, they are all counted to similarity and only can bring interference to Gabor characteristic, dwindle interior spacing, and even put upside down recognition result.Therefore must carry out Feature Selection, inapplicable feature is rejected, this will expand class spacing effectively, the recognition of face ability of strengthening system, the robustness of raising system.Another benefit of Feature Selection is apparent: the speed that can improve system.Identify with several points that screen, in improving recognition capability, also improved the recognition speed of system.

6.7. similarity comparison

The result of Feature Selection and sequence has improved in light application ratio more even, unobstructed, and the not too large situation human face of face local deformation is identified the robustness to attitude, and has improved speed.This is a kind of more satisfactory situation, not too evenly as excessively strong in illumination in illumination so, in hypographous situation, in the situation that blocking, or there is larger local deformation, as eyes closed, face magnify etc. in more common situation by how to process following we such situation is discussed.First the performance characteristic of investigating characteristic similarity in these three kinds of situations, then redefines similarity function according to this feature, the unique point under these three kinds of situation impacts is got rid of outside similarity measurement, thereby improved the robustness of recognition of face.Characteristic of correspondence point similarity in the unique point of face and storehouse on correct faceform is very little, and we are referred to as, and feature lost efficacy or feature failure.These three kinds of common performances of situation are: the characteristic of correspondence point on any faceform in the unique point in specific region and storehouse is all dissimilar.This has just determined the characteristic of this Regional Similarity.It is random in experiment, can observing similarity in these regions, compares fluctuation very large with the unique point that does not have to lose efficacy, and similarity not necessarily obtains maximal value on which faceform, and the position of these feature failed areas is unforeseen.It is all not many that the upper feature of any like this faceform obtains peaked number of times, and face database capacity is larger, and the peaked chance of the upper acquisition of each faceform is fewer.This is that on each faceform, equiprobability obtains maximal value in theory because the characteristic of correspondence point on any faceform in unique point and storehouse is all dissimilar.Our solution is to improve similar function dynamically to screen feature to improve the robustness of recognition of face.

The module that the present invention is designed, except the implementation that can be given an example by above-described embodiment realize unexpected, all can also be by simulating and/or the hardware circuit such as digital circuit forms.Those skilled in the art, in the design basis of conventional simulation/digital circuit, according to content disclosed by the invention, can realize corresponding function.

Recognition of face queue machine of the present invention, use image technique, card reader of ID card, database system the technical approach such as to transfer and obtain portrait and the identity information in all kinds of lawful documents that personnel provide, and with itself and external camera dynamically or the static figure information obtaining compare, to reach the target of " authentication unification ".Preferably resolve and rely on strength or force when buying goods and materials and the ticketing service voucher for queuing by face recognition technology, the problem of the generation number of getting.On-the-spot order management and the discriminating to queuing people identity are improved.

In addition, can also all acquisitions in system comparison process with the information generating, be saved in system master host data base, in order to follow-up system application.

Claims (7)

1. a recognition of face queue machine, is characterized in that: comprising:

Numbering service trigger module, for according to outside input, triggers row number flow process;

Certificate information collection module, for according to the triggering of numbering service trigger module, gathers certificate information;

Human image collecting module, for according to the triggering of numbering service trigger module, gathers holder photo;

Portrait analysis module is for extracting corresponding face characteristic value from described certificate photograph and holder photo;

Portrait comparing module is for comparing the face characteristic value of the face characteristic value of described certificate photograph and holder photo, if the similarity of the face characteristic value of the face characteristic value of described certificate photograph and holder photo meets or exceeds default first threshold, judge that comparison result is as passing through; If lower than Second Threshold, judge that comparison result is as not passing through, and start authentication failed flow process;

Print module is for print queue's number in the time that portrait comparing module judges that contrast is passed through; With

The device system of calling out the numbers for judge in portrait comparing module contrast by time include this holder in row number queue, and call out the numbers in the time taking turns to the queue number of this holder.

2. recognition of face queue machine according to claim 1, is characterized in that: also comprise:

Database storage backup module is for storing the related data of holder.

3. recognition of face queue machine according to claim 2, it is characterized in that: this numbering service trigger module is connected to certificate information collection module and human image collecting module, certificate information collection module and human image collecting module are connected to portrait analysis module, portrait analysis module is connected to portrait comparing module, and portrait comparing module is connected respectively to database storage backup module, print module and the device system of calling out the numbers.

4. recognition of face queue machine according to claim 1, is characterized in that: described certificate information collection module comprises:

Identity card reader identification module is for reading the information that in certificate, chip comprises;

OCR module is for obtaining the Word message on certificate; With

Certificate information identification module for according to the triggering of described numbering service trigger module, is controlled the collections of identity card reader identification module and OCR module, and is sent to portrait analysis module by controlling the data that identity card reader identification module and OCR module collect.

5. recognition of face queue machine according to claim 1, is characterized in that: described human image collecting module comprises:

Shooting taking module is for taking to obtain holder photo to holder; With

Dynamically human image collecting module, for according to the triggering of described numbering service trigger module, is controlled the collection of described shooting taking module, and the described holder photo collecting is sent to portrait analysis module.

6. according to the recognition of face queue machine described in claim 1,4 or 5, it is characterized in that: described in the device system of calling out the numbers comprise:

Numbering service storehouse module for portrait comparing module judge contrast by time include this holder in row number queue, with to row number queue manage;

Call out the numbers device for receiving operating personnel's control, call out the numbers or queue management arranges instruction to the management module input of calling out the numbers;

The management of calling out the numbers module is for according to the row number information of numbering service storehouse module output, or according to the described device output of calling out the numbers call out the numbers or queue management arranges instruction, control loudspeaker and display screen and call out the numbers.

7. recognition of face queue machine according to claim 6, is characterized in that: described in the management module of calling out the numbers comprise voice module, LTE control module, word processing module and hardware driving.