CN1971630A - Access control device and check on work attendance tool based on human face identification technique - Google Patents

Abstract

A gate inhibition and work attendance device based on face discrimination technique are disclosed those include: cam used to read the face images, embedded system used to transmit the video information, input device used to input the identification information and computer used to distinguish the images and manage the work attendance. Said cam is connected with the embedded system, said embedded system is connected with the input device, and said embedded system is connected with the computer data. The gate inhibition and work attendance device also include mirror face used to confirm the side position of face, the cam is mounted on the top of mirror face, the optimal visual range of the cam is matched with the mirror face; said computer includes: image acquisition module, face image library, face detecting position module, image preprocessing module, face exercising module and face discrimination module. The invention provides the gate inhibition and work attendance device based on face discrimination technique with characterized that high discrimination success rate, easy to realize, high-speed to discriminate and judge, convenient to install and low implement cost.

Description

Gate inhibition and Work attendance device based on human face identification technique

(1) technical field

The present invention relates to a kind of gate inhibition and Work attendance device.

(2) background technology

Enterprise settles gate control system to become requisite ladder of management to employee's attendance management and important place, the check card work attendance or the work attendance of swiping the card commonly used at present, contact attendance recorder and radio-frequency card attendance recorder are arranged in the work attendance of swiping the card, this class Work attendance method, happen occasionally on behalf of the diligent phenomenon of praticing fraud of checking card or swiping the card by other staff, influenced the accuracy of work attendance.In order to prevent the generation of the above-mentioned diligent phenomenon of praticing fraud, often occur for the phenomenon of checking card, so the insider has invented fingerprint attendance machine in conjunction with technique of fingerprint indentification, invented the face picture in conjunction with human face identification technique and differentiated attendance recorder.For fingerprint attendance machine, because present technique of fingerprint indentification fingerprint resolution can't reach well-content effect, because there is not positive evidence to prove still employee's the own problem that the fingerprint resolution causes, so tend to produce occurrences in human life timekeeper and employee to the attendance record different opinions, produced dispute.China's utility model patent 200520061451.2 has proposed a kind of straight card fingerprint attendance machine, fingerprint is differentiated combine with the work attendance of checking card, so that can also stay the inquiry afterwards that employee and occurrences in human life timekeeper be convenient in record at card when fingerprint is differentiated the generation problem.Differentiate attendance recorder for the face picture, Chinese patent 99117360.0 has proposed a kind of picture and has differentiated gate inhibition and attendance checking system, by the camera motion capture work attendance personnel head portrait that links to each other with computing machine, compare, judge with employee's head portrait of typing in advance in the database, make the decision of whether passing through, and make work attendance and work attendance statistics, carry out attendance management.Can both prevent from effectively the to pratice fraud generation of diligent phenomenon of above-mentioned two kinds of attendance recorders.

But above-mentioned attendance recorder also exists some problems, because enterprise often is placed in attendance recorder the porch of enterprise, the work attendance machine that is placed on the outside can take place stolen and originally artificially damage to happen occasionally, and uses above-mentioned attendance recorder that bigger difficulty is arranged for mobile bigger units such as mine, building ground; Because staff number is many, and work attendance is to concentrate on the commuter time, the problem of work attendance can occur waiting in line, occur to occur in the time of to differentiate particularly that bigger to wait work attendance crowded for some bigger enterprises at fingerprint or people's face; Differentiate attendance recorder for the face picture, require camera motion capture work attendance personnel head portrait, then this head portrait and the head portrait that leaves the Computer Storage unit in are mated and differentiate calculating, because at present the human face identification technique that adopts does not add any innovation means and directly is not used in the face picture and differentiates in gate inhibition and the attendance checking system, thus realization difficulty height has appearred, differentiate that success ratio is low, differentiate judgement time long, to problems such as equipment requirements height.

(3) summary of the invention

Differentiate that success ratio is low, realize the difficulty height, differentiate that judgement time is long, the deficiency inconvenient, that implementation cost is high is installed for the people's face that overcomes existing gate inhibition and Work attendance device, the invention provides and a kind ofly differentiate the success ratio height, be easy to gate inhibition and the Work attendance device of realizing, differentiate that judgement speed is fast, easy for installation and implementation cost is low based on human face identification technique.

The technical solution adopted for the present invention to solve the technical problems is:

A kind of gate inhibition and Work attendance device based on human face identification technique, comprise the camera that is used to read facial image, be used to gather the embedded system that transmits video information, be used to import the input equipment of identity information and be used to carry out that image is differentiated and the computing machine of staff attendance management, described camera connects embedded system, described embedded system connects input equipment, described embedded system is connected with computer data, described computing machine comprises: image collection module is used to obtain the video image that embedded system is gathered; Facial image database is used for the facial image of storing sample;

Described gate inhibition and Work attendance device also comprise the minute surface that is used to confirm people's face front position, and described camera is positioned at the top of minute surface, and the optimum visual scope and the minute surface of described camera are complementary;

Described computing machine also comprises:

People's face detection and location module is used for determining the definite position of people's face people's face being separated from background, is output as and cuts apart good people's face topography; The rgb color format conversion is become YCbCr color form, obtain YCb ' Cr ' color space through non-linear segmented color color conversion then; The RGB coordinate space is as follows to the transformation for mula of YCb ' Cr ' coordinate space to YCbCr coordinate space and YCbCr coordinate space:

Y＝0.29990*R+0.5870*G+0.1140*B

Cr(Y)＝0.5000*R-0.4187*G-0.0813*B+128

Cb(Y)＝-0.1787*R-0.3313*G+0.5000*B+128 (11)

$C_i^{\&prime;}\left(Y\right)=\left\{\begin{array}{cc}{C}_i\left(Y\right)& \mathrm{if\; Y}\&Element;[{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1,K_h]\\ (C_i\left(Y\right)-\stackrel{\&OverBar;}{C_i}\left(Y\right))\frac{W_{C_i}}{W_{C_i}\left(Y\right)}+\stackrel{\&OverBar;}{C_i}\left(Y\right)& \mathrm{if\; Y}<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\mathrm{or\; Y}>K_h\end{array}\right.---\left(12\right)$

Wherein i represents b or r,

With The axis of expression area of skin color, calculate with following formula:

$\stackrel{\&OverBar;}{C_b}\left(Y\right)=\left\{\begin{array}{ccc}108+\frac{(K_l-Y)(118-108)}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ 108+\frac{(Y-K_h)(118-108)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

$\stackrel{\&OverBar;}{C_r}\left(Y\right)=\left\{\begin{array}{ccc}154-\frac{(K_l-Y)(154-144)}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ 154+\frac{(Y-K_h)(154-132)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

K _lAnd K _hThe versicolor segmentation thresholding of the linear segmentation of right and wrong; Y _MinAnd Y _MaxBe the minimum and the maximal value of Y component in the colour of skin cluster areas that obtains by people's face experimental data according to the somewhere;

W _Cb(Y) and W _Cr(Y) width of expression area of skin color:

$W_{C_i}\left(Y\right)=\left\{\begin{array}{ccc}{\mathrm{WL}}_{C_i}+\frac{(Y-Y_{\min })(W_{C_i}-{\mathrm{WL}}_{C_i})}{K_l-Y_{\min }}& \mathrm{if}& Y<K_l\\ {\mathrm{WH}}_{C_i}+\frac{(Y_{\max }-Y)(W_{C_i}-{\mathrm{WH}}_{C_i})}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

W _Ci, WL _Ci, WH _CiBe constant, be respectively: $W_{C_b}=46.97,$ ${\mathrm{WL}}_{C_b}=23,$ ${\mathrm{WH}}_{C_b}=14,$ $W_{C_r}=38.76,$ ${\mathrm{WL}}_{C_r}=20,$ ${\mathrm{WH}}_{C_r}=10;$

The image pretreatment module is used for the facial image that extracts is carried out geometrical normalization, eliminates noise and gray scale normalization processing, and unified people is bold little and brightness;

People's face training module is used for the facial image of image data base is projected to the subspace by the PCA method, makes image space be converted into best description feature space, and the projection matrix of Fisher face algorithm is:

${W_{\mathrm{opt}}}^T=W_{\mathrm{fld}}^T{W}_{\mathrm{pca}}^T---\left(1\right)$

In the formula:

W _pca＝argmax|W ^TS _TW|

$W_{\mathrm{fld}}=\arg \max \left|\frac{W^T{W}_{\mathrm{pca\&alpha;}}^T{S}_B{W}_{\mathrm{pca}}W}{W^T{W}_{\mathrm{pca}}^T{S}_w{W}_{\mathrm{pca}}W}\right|$

Column vector is orthogonal vector, is called the Fisher face, and face images projects to W in the training set _OptColumn vector on obtain a stack features of every width of cloth facial image;

Adopt the Fisher linear discriminant analysis that the MEF space is converted into best diagnostic characteristics space again, similar sample is intensive as far as possible, different class samples as far as possible separately, and used people's face parameter when obtaining differentiating;

People's face identification module, the people's face parameter that is used for the people's face parameter that will extract and current personnel's image training compares, and confidence level surpasses assign thresholds M, then compares successfully, and identity obtains affirmation.

Further, in described people's face training module, the resolution of everyone face training sample is m*n, and it is the column vector of a mn, all training samples add up to N, the total number of persons that training sample is concentrated is P; Individual human face training sample x _iTo expression, i represents i people, and k represents i people's k sample; I the total n of people _iIndividual training sample, visible N=n ₁+ n ₂+ ... + n _pPeople's face sample to be identified is represented with y, and the resolution of sample is identical with training sample; The algorithm of training is:

Step1: it is right to import n people's face training sample, i=1, and 2 ..., p; K=1,2 ... n;

Step2: calculate the mean vector m of all samples and the mean vector (i people's the average image vector) of i class people face sample

$m=\frac{1}{N}\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{x}_i^k\&Element;R^{\mathrm{mn}\&times;1}---\left(2\right)$

$m_i=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{x}_i^k\&Element;R^{\mathrm{mn}\&times;1}---\left(3\right)$

Step3: the generation matrix of asking Karhunen-Loeve transformation

$S_b=\underset{i=0}{\overset{p-1_i}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)(m_i-m){(m_i-m)}^T---\left(4\right)$

P (ω in the following formula _i) be ω _i(i=1,2 ..., the c) prior probability of quasi-mode;

Step4: according to the theory of SVD, scatter matrix S between compute classes _bEigenvalue _iCharacteristic of correspondence vector u with it _i

Step5: get maximum preceding a (the individual eigenwert characteristic of correspondence vector of a≤p), structural matrix

U＝[u ₁，u ₂，…，u _a]∈R ^mn×a (5)

Step6: calculate everyone face training sample at u ₁, u ₂..., u _aProjection under the formed subspace, note x _i ^kAfter projection z _i ^k∈ R ^{A * 1}, formula (6) is then arranged

$Z=|{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">z</figure-callout>}}_1^1,{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">z</figure-callout>}}_1^2,...,z_1^{{\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">n</figure-callout>}}_1},{\mathrm{<figure-callout\; id="2"\; label="z"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">z</figure-callout>}}_2^1,...{,z}_P^{n_p}|\&Element;R^{a\&times;N}---\left(6\right)$

Step7: calculate the mean vector m of all kinds of people's face samples under the subspace _PcaiMean vector m with all classes _Pca, computing formula is shown in (7), (8):

$m_{\mathrm{pcai}}=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{z}_i^k\&Element;R^{a\&times;1},i=\mathrm{1,2},...,n_c---\left(7\right)$

$m_{\mathrm{pca}}=\underset{i=1}{\overset{p_i}{\mathrm{\&Sigma;}}}{m}_{\mathrm{pcai}}\&Element;R^{a\&times;1}---\left(8\right)$

Step8: calculate scatter matrix S in the total class under the subspace _wAnd scatter matrix S between class _b, computing formula is shown in (9), (10):

$S_w=\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}{S}_i=\underset{i=1}{\overset{p_l}{\mathrm{\&Sigma;}}}\left[\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}(z_i^k-m_{\mathrm{pcai}}){(z_i^k-m_{\mathrm{pcai}})}^T\right]\&Element;R^{a\&times;a}---\left(9\right)$

$S_b=\underset{i=1}{\overset{p_i}{\mathrm{\&Sigma;}}}{(m_{\mathrm{pcai}}-m_{\mathrm{pca}})(m_{\mathrm{pcai}}-m_{\mathrm{pca}})}^T\&Element;R^{a\&times;a}---\left(10\right)$

Step9: know S by inference according to the Fisher criterion function _bw ^*=λ S _ww ^*, it is a matrix S _bWith respect to matrix S _wGeneralized eigenvalue problem, find the solution S _bWith respect to matrix S _wGeneralized eigenvector, w ^*, i=1,2 ..., p

Step10: get generalized eigenvector and form matrix W and matrix U and multiply each other and obtain new matrix T, i.e. T=UW, the column vector of T is the result of training, keeps T and uses when to be identified;

Step11: calculate everyone the projection meanS of average face under the T of subspace, keep the projection coordinate coefficient of meanS, use when to be identified as this type of sample;

Step12: adopt the method for Euclidean distance, determine the similarity threshold value, use when the reservation threshold value is to be identified according to reject rate as similarity measurement;

In described people's face identification module, the algorithm of discriminating is:

Step1: calculate sample y to be identified at t ₁, t ₂..., t _C-1Projection coordinate's coefficient under the subspace that is generated

Step2: calculate The projection coordinate coefficient corresponding with this type of sample

Between similarity measurement value d;

Step3: compare the size of the similarity threshold value M that obtains in d and the training process,, then declare people's face sample to be identified and pass through, otherwise refuse to know if d is littler than threshold value M.

Described people's face training module also comprises: the value determining unit of projection coordinate's coefficient is used for determining the dimension of the subspace of projection coordinate.

Described people's face training module also comprises: the similarity threshold setup unit, be used for the selected threshold size, and determine similarity threshold M according to wrong percent of pass and false rejection rate.

Described people's face training module also comprises: adaptive updates face characteristic parameter unit, be used for utilizing each to the obtained facial image of people's face discrimination process, facial image in the facial image database is upgraded, face characteristic parameter in the face characteristic parameter library is upgraded, removed to replace the most outmoded facial image and the characteristic parameter of facial image with the characteristic parameter of up-to-date facial image of obtaining and facial image;

Described people's face identification module also comprises: repeatedly gather facial image pedestrian's face discriminating unit of going forward side by side, be used for when input equipment has input, beginning to gather facial image and it being kept at the storage unit of embedded system, when having imported identity information and pressed the facial image of gathering facial image again when sending key and will originally being kept at simultaneously in the storage unit of embedded system, the user sends to background computer together, after background computer obtains two width of cloth facial images, carry out people's face and differentiate calculating, simultaneously embedded system is gathered facial image, and facial image is sent to background computer once more carries out people's face and differentiate; At first differentiate,, finish if differentiate and pass through to press the facial image of being gathered when sending key the user; Do not have to pass through end if differentiate if differentiate the first time by then the facial image of being gathered for the second time being carried out people's face discriminating second time; If still not by then the facial image of being gathered for the third time being carried out people's face discriminating for the third time.

Described image pretreatment module comprises:

The denoising processing unit, be used for adopting filtering method to remove noise based on pixel " density ", the filtering method of similar " convolution " is added up the number of skin pixel in the 5*5 neighborhood that with each pixel is the center, central point is left the colour of skin when surpassing half, otherwise thinks it is the non-colour of skin;

Determine maximum territory element, be used for, the object in publishing picture along regular mark according to eight fillets the skin color segmentation figure after handling through denoising, calculate the area of each object, the object that keeps the area maximum, and it is filled, obtain filtering and filled later skin color segmentation image;

The vertical projection unit is used for according to formula (13) later skin color segmentation image being filled in filtering and carries out vertical projection, and calculates the average Mean of each row non-zero projection value;

$\mathrm{Mea\&alpha;n}=\frac{\underset{x=1}{\overset{M}{\mathrm{\&Sigma;}}}{P}_y\left(x\right)}{\mathrm{num}}---\left(13\right)$

Wherein num is P _yThe number of the non-zero points (x);

Selected threshold is half of Mean, if projection value greater than threshold value, then keeps; Otherwise put 0; The zone that keeps the area maximum is as last projection output;

Determine people's face width unit, be used for the vertical projection diagram that obtains above-mentioned, look for first non-0 point from left to right, its corresponding columns promptly is the left margin of people's face; Equally, turn left from the right side and to look for first non-0 point, the right margin of the columns behaviour face that it is corresponding, the distance between the border, the left and right sides is exactly the width of people's face;

Determine unit, the crown, be used for the vertical projection diagram that obtains above-mentioned, from top to bottom, seek first skin pixel and count out, determine that this journey is the crown greater than 15 row;

Obtain unit, people's face rectangular area, be used for according to the crown of determining in the top step and the corresponding former figure of people's face width, cutting out last people's face rectangular area by getting people's face height=people's face width * 1.4.

Described gate inhibition and Work attendance device also comprise the houselights unit, and described houselights unit is arranged on the top of minute surface, and described illuminating lamp unit is positioned at the top of camera.

Described embedded system comprises: the image recording module is used for person's to be measured video information recording is got off; Image processing module, the video data that is used for noting carry out compressed encoding, multiplexing and be modulated into compressed video data; First radio receiving transmitting module is used for according to communication standard, sends the video data of compression; Described computing machine comprises: second radio receiving transmitting module, be used for according to communication standard, and receive the video data of compression; The image decompression processing module, the data that are used for receiving decompress, demultiplexing and demodulation, revert to video data.

The described embedded system that is used for the video image processing is connected with computing machine by network.

Technical conceive of the present invention is: human face identification technique is a kind of man-to-man comparison, it is a kind of affirmation of identity, under this pattern, face characteristic information is stored in facial image database and the face characteristic parameter library, human face identification technique only needs simply the face characteristic parameter of real-time people's face data and storage is compared, if confidence level surpasses the M value of an appointment, then to compare successfully, identity obtains confirming.Application scenarios in this patent is that the visitor provides an ID (member's job number), from the face characteristic parameter library of storage, obtain this visitor's face characteristic parameter simultaneously automatically according to this ID (member's job number), and compare, thereby judge whether the individuality corresponding with the ID that is provided is same individual to this visitor with the facial image of real-time collection and through the face characteristic parameter after the feature extraction.The computing that people's face is differentiated is carried out on background computer, the foreground embedded system sends the facial image of camera acquisition and the ID of visitor's input (member's job number) to background computer by wireless network, and ID that the background computer basis sends and the image that is comprising people's face carry out people's face and differentiate;

Accompanying drawing 2 is main functional modules that people's face of realizing on background computer is differentiated: 1) image collection module, this module are finished and are obtained image, and image comes from embedded system; 2) people's face detection and location module, this module finds the definite position of people's face, and further people's face is separated from background, be output as and cut apart good people's face topography, the whole discrimination process that is operated in of this part is very important, the work of this part can become fairly simple under specific circumstances, among the present invention by the concrete imaging conditions of control (the people's face that guarantees shooting be positive, illumination condition is constant, image background is simple as far as possible), it is oversimplified, make the location ratio of people's face be easier to; 3) image pretreatment module, the main effect of this module are to remove as much as possible or reduce illumination, imaging system, external environment condition or the like interference for pending image, for subsequent treatment provides high-quality image.This part is finished the geometrical normalization of the facial image that extracts, eliminates processing such as noise and gray scale normalization, make in the different images people be bold little and brightness unified so that under identical conditions, finish training and differentiate; 4) feature extraction and select module, this module be by extracting the feature that is used for differentiating according to certain strategy in the pretreated facial image of above-mentioned image, with the data mapping in original people's face space to feature space.Because original image data amount is sizable, differentiate in order to realize classification effectively, will carry out conversion to raw data, obtain reflecting the feature of classification essence.How extracting stable and effective feature is the identification system key of success, and this patent adopts statistical nature face discrimination method, and this eigenface discrimination method is the proper vector structural attitude face that utilizes the image correlation matrix; 5) training module, the parameter that after this module is carried out and finished generation be can be used for differentiating, the discriminating of people's face can be regarded as people's face object to be identified is grouped in a certain class, therefore the effect of training module is exactly the sorter of classifying and differentiating, on people's face sample training collection basis, determine certain decision rule, make the loss minimum that the wrong resolution that is caused being classified by people's face discriminating object by this decision rule is minimum or cause; 6) identification module is finished the differentiation work of people's face according to the parameter of training gained, provides last identification result, and makes corresponding judgment.People's face identification system generally can be divided into two processes substantially, and as shown in Figure 2, dotted line the first half is a training process, and this process has been finished the design of sorter; Dotted line the latter half is people's face discrimination process.

Wireless video transmission is with the video information digitizing after the transmission of delivering letters after the compressed encoding, multiplexing, modulation, then by carrying out Flame Image Process in the computing machine of delivering to human resource administration after decompression, demultiplexing, the demodulation, what whole process adopted is the full digital processing technology, owing to adopted Digital Signal Processing and forward error correction technique (FEC), can make its reliable transmission that has the higher signal receiving sensitivity and can guarantee signal, the advantage with aspects such as video transmission interference free performance, installation and maintenance is more obvious.

The real-time radio video transmission except have superior interference free performance, keep image information steady and audible, have also that equipment is small and exquisite need not to attach advantages such as other facilities, moderate cost, can realize point-to-point, point-to-multipoint, unidirectional and two-way real-time multimedia communication by suitable geocoding control.Because it is good that digital real-time radio Video transmission system need not attach other facilities and independence,, so adopt the real-time radio video transmission technologies can realize the gate inhibition and the work attendance of human face identification technique easily without any need for installation processes such as wirings.

Beneficial effect of the present invention mainly shows: 1, people's face is differentiated the success ratio height, is implemented convenient, people's face discriminating time weak point; 2, easy for installation; 3, implementation cost is low; 4, the IEEE802.11b radio communication of Cai Yonging is as the gate inhibition of human face identification technique and the video and audio communication of Work attendance device, and applicability is good; 5, along with facial image database and face characteristic database, accuracy height are upgraded in the growth meeting at user's age automatically.

(4) description of drawings

Fig. 1 is based on the gate inhibition of wireless video and human face identification technique and the structured flowchart of Work attendance device.

Fig. 2 is that people's face is differentiated training and identification flow figure.

Fig. 3 is a Fisher face algorithm flow block diagram.

Fig. 4 is cut apart the human face region process flow block diagram.

Fig. 5 is the synoptic diagram after human face region is cut apart.

Fig. 6 is the wireless communication configuration figure of foreground embedded system and background computer.

(5) embodiment

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

With reference to Fig. 1～Fig. 6, a kind of gate inhibition and Work attendance device based on human face identification technique, comprise the camera 6 that is used to read facial image, be used to gather the embedded system 7 that transmits video information, be used to import the input equipment 3 of identity information and be used to carry out the computing machine 9 that image is differentiated and staff attendance is managed, described camera 6 connects embedded system 7, described embedded system 7 connects input equipment 3, described embedded system 3 is connected with computing machine 9 data by communication network 8, described gate inhibition and Work attendance device also comprise the minute surface 1 that is used to confirm people's face front position, described camera 6 is positioned at the top of minute surface 1, and the optimum visual scope and the minute surface 1 of described camera 6 are complementary; Described computing machine 9 comprises: image collection module is used to obtain the video image that embedded system is gathered; Facial image database is used for the facial image of storing sample; People's face detection and location module is used for determining the definite position of people's face people's face being separated from background, is output as and cuts apart good people's face topography; The rgb color format conversion is become YCbCr color form, obtain YCb ' Cr ' color space through non-linear segmented color color conversion then; The RGB coordinate space is as follows to the transformation for mula of YCb ' Cr ' coordinate space to YCbCr coordinate space and YCbCr coordinate space:

Y＝0.29990*R+0.5870*G+0.1140*B

Cr(Y)＝0.5000*R-0.4187*G-0.0813*B+128

Cb(Y)＝-0.1787*R-0.3313*G+0.5000*B+128 (11)

$C_i^{\&prime;}\left(Y\right)=\left\{\begin{array}{ccc}{C}_i\left(Y\right)& \mathrm{if}& Y\&Element;[{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1,K_h]\\ (C_i\left(Y\right)-\stackrel{\&OverBar;}{C_i}\left(Y\right))\frac{W_{C_i}}{W_{C_i}\left(Y\right)}+\stackrel{\&OverBar;}{C_i}\left(Y\right)& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\mathrm{orY}>K_h\end{array}\right.---\left(12\right)$

Wherein i represents b or r, With The axis of expression area of skin color, calculate with following formula:

$\stackrel{\&OverBar;}{C_b}\left(Y\right)=\left\{\begin{array}{ccc}108+\frac{(K_l-Y)(118-108)}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ 108+\frac{(Y-K_h)(118-108)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

$\stackrel{\&OverBar;}{C_r}\left(Y\right)=\left\{\begin{array}{ccc}154-\frac{(K_l-Y)(154-144)}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ 154+\frac{(Y-K_h)(154-132)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

K _lAnd K _hThe versicolor segmentation thresholding of the linear segmentation of right and wrong; Y _MinAnd Y _MaxBe the minimum and the maximal value of Y component in the colour of skin cluster areas that obtains by people's face experimental data according to the somewhere;

W _Cb(Y) and W _Cr(Y) width of expression area of skin color:

$W_{C_i}\left(Y\right)=\left\{\begin{array}{ccc}{\mathrm{WL}}_{C_i}+\frac{(Y-Y_{\min })(W_{C_i}-{\mathrm{WL}}_{C_i})}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ {\mathrm{WH}}_{C_i}+\frac{(Y_{\max }-Y)(W_{C_i}-{\mathrm{WH}}_{C_I})}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

W _Ci, WL _Ci, WH _CiBe constant, be respectively: $W_{C_b}=46.97,$ ${\mathrm{WL}}_{C_b}=23,$ ${\mathrm{WH}}_{C_b}=14,$ $W_{C_r}=38.76,$ ${\mathrm{WL}}_{C_r}=20,$ ${\mathrm{WH}}_{C_r}=10;$

The image pretreatment module is used for the facial image that extracts is carried out geometrical normalization, eliminates noise and gray scale normalization processing, and unified people is bold little and brightness;

People's face training module is used for the facial image of image data base is projected to the subspace by the PCA method, makes image space be converted into best description feature space, and the projection matrix of Fisher face algorithm is:

${W_{\mathrm{opt}}}^T=W_{\mathrm{fld}}^T{W}_{\mathrm{pca}}^T---\left(1\right)$

In the formula:

W _pca＝argmax|W ^TS _TW|

$W_{\mathrm{fld}}=\arg \max \left|\frac{W^T{W}_{\mathrm{pca}}^T{S}_B{W}_{\mathrm{pca}}W}{W^T{W}_{\mathrm{pca}}^T{S}_w{W}_{\mathrm{pca}}W}\right|$

Column vector is orthogonal vector, is called the Fisher face, and face images projects to W in the training set _OptColumn vector on obtain a stack features of every width of cloth facial image;

Adopt the Fisher linear discriminant analysis that the MEF space is converted into best diagnostic characteristics space again, similar sample is intensive as far as possible, different class samples as far as possible separately, and used people's face parameter when obtaining differentiating;

People's face identification module, the people's face parameter that is used for the people's face parameter that will extract and current personnel's image training compares, and confidence level surpasses assign thresholds M, then compares successfully, and identity obtains affirmation.

The gate inhibition who is based on wireless video and human face identification technique shown in Figure 1 and the structured flowchart of Work attendance device, described gate inhibition and Work attendance device mainly are made up of two parts, the both video image acquisition on foreground and man-machine conversation part, people's face on backstage differentiates and the data storage part that the embedded system on foreground realizes communicating by letter by wireless network with the computing machine on backstage; The video image acquisition on described foreground and man-machine conversation partly comprise minute surface 1, sound playing unit 2, input keyboard 3, houselights unit 4, image unit 6, embedded system 7, sound playing unit 2, input keyboard 3, image unit 6 are connected with embedded system 7 by corresponding interface, and embedded system 7 communicates with background computer 9 by wireless communication unit 8 and is connected;

Location when described minute surface is used for the shooting of people's face, guarantee that captured facial image is positive and in the scope of minute surface, the user is carrying out using keyboard to import my id number when identity is differentiated, a projection is arranged on minute surface simultaneously, as 5 among Fig. 1, thereby guaranteed that image unit 6 faces the shooting of people's face, also confined the position of people's face in photographed images simultaneously, locating and differentiate for people's face provides good exterior condition;

Described sound playing unit is used for man-machine mutual, the prompting of giving some close friends of user;

Described input keyboard is used for id number and external personnel that the user imports the employee and imports visiting information, and the position of keyboard is placed in the middle and upper part, right side of minute surface, makes the user make a video recording when can use keyboard very naturally;

Described houselights unit is used to guarantee the consistance of illuminating ray when shooting, improves the resolution of people's face;

Described embedded system, the voice output of be used to gather people's face video image, transmission of video information and some other information, controlling sound playing unit;

Described wireless communication unit is used for the radio communication in the WLAN (wireless local area network), and wireless communication module is observed the IEEE802.11b wireless communication protocol;

Described background computer is used for the facial image that the foreground sends is differentiated, and the result that will differentiate sends the foreground embedded system to by the wireless communication unit;

Described embedded system, specifically selecting Samsung S3C2410X is embedded microprocessor, the combining wireless local area network technology, design realizes video data acquiring and wireless transmission based on the gate inhibition and the Work attendance device of wireless video and human face identification technique.Comprised software and hardware technology in the embedded system, wherein built-in Linux software is core technology, and it can realize the function of video server.

Described embedded microprocessor S3C2410X is a 16/32 RISC embedded microprocessor based on the ARM920T kernel, and this processor designs for handheld device and high performance-price ratio, low-power consumption microcontroller.It has adopted the new bus architecture of a kind of AMBA of being called (Advanced Microcontroller Bus Architecture).The main resource of S3C2410X inside has memory management unit MMU, system administration manager, respectively is the instruction and data buffer memory of 16KB, lcd controller (STN﹠amp; TFT), NAND FLASH Boot Loader, 3 passage UART, 4 passage DMA, 4 PWM clocks, 1 internal clocking, 8 path 10s are ADC, touch screen interface, multimedia card interface, I2C and I2S bus interface, 2 usb host interfaces, 1 USB device interface, SD main interface, 2SPI interface, pll clock generator and general purpose I/O port etc.

Described embedded microprocessor S3C2410X inside comprises a memory management unit that is MMU, can realize the mapping of virtual memory space to amount of physical memory.Usually the program of embedded system leaves among the ROM/FLASH, program can access preservation behind the system cut-off, but ROM/FLASH compares with SDRAM, it is slow many that speed is wanted, and usually the aborted vector table is left among the RAM in the embedded system, utilize memory-mapped mechanism can solve this needs.

Described ROM/FLASH adopts the K9S1208VOM of the 64MB of Samsung.It can carry out 100,000 times program/erase, and data are preserved and reached 10 years, are used to loading operation system image and large-capacity data.

Described SDRAM is the K4S561632C that adopts Samsung, is used for needed data in operation system and the stored programme operational process, and it is the synchronous dram of 4M*16bit*4bank, and capacity is 32MB.Realize the position expansion with two K4S561632C, making data-bus width is 32bit.

Described embedded software system mainly comprises writing of the installation of transplanting, driver of operating system, TCP/TP agreement and user application etc.

Adopted Linux as embedded OS among the present invention, Linux develops from UNIX, inherited the most advantage of UNIX, the disclosed kernel source code of Linux makes it become present most popular operating system, and Linux can be from its hardware-software of application cutting, this is concerning towards very necessary based on the gate inhibition of wireless video and human face identification technique and this special requirement of Work attendance device, here we are referred to as the customization operations system, and customization step is as follows: (1) writes plate base support package BSP; (2) each parts of cutting and configuration operation system, and revise corresponding configuration file; (3) compiling Kernel, assembly and BSP, generating run system image file; (4) image file is downloaded on the Target Board, debug.

Further, video information is to transmit in the mode of packing data, transmission through WLAN (wireless local area network) by ICP/IP protocol, therefore under the operating system support, realize ICP/IP protocol, just need carry out task division, the realization of TCP/IP can be divided into 4 tasks realizes: 1. IP task, the reorganization that is mainly used to solve IP fragmentation; 2. the TCP incoming task is mainly used to handle the TCP message segment that receives; 3. TCP output task is mainly used to packing data, the transmission that will export; 4. TCP task of timer, being mainly used to provides clock for various time delay incidents (as the repeating transmission incident).

Further, based on needing two USB interface in the gate inhibition of wireless video and human face identification technique and the Work attendance device, one of them USB interface is that camera is connected with S3C2410X, another USB interface is that wireless network card is connected with S3C2410X, because S3C2410X self-carried USB principal and subordinate interface, do not need special USB chip support, as long as can carry out USB transmission data to its install driver.

Described USB driver comprises following several sections: (1) establishment equipment, create two parameter calls of equipment function band, and a parameter is to point to the pointer of driver object, another parameter is to point to the pointer of physical device object; (2) closing device; (3) fetch equipment data, when client applications has requiring of fetch equipment data, system requires this to pass to function driver with the IRP form of IRP_MJ_READ, D12Meter_Read program by equipment is carried out, and then specifies the direct and equipment realization information interaction of usb bus driver by D12Meter_Read; (4) equipment is write data, when client applications has requiring of write device data, system requires this to pass to function driver with the IRP form of IRP_MJ_WRITE, and carry out by D12Meter_Write, and then by D12Meter_Write specify the usb bus driver directly with equipment realization information interaction.The USB driver is by PID in the installation file (.inf file) (product differentiate number) and VID (manufacturer differentiate number) discriminating USB device.

After the embedded OS loading is finished, driver and other corresponding application of wireless network card just can be installed.The driver of wireless network card is bundled in the operating system as a module, can avoids the WLAN Device Driver of at every turn all will resetting after system's power down.

As a kind of gate inhibition and Work attendance device based on human face identification technique, human face identification technique is a kind of man-to-man comparison, it is a kind of affirmation of identity, under this pattern, face characteristic information is stored in facial image database and the face characteristic parameter library, and human face identification technique only needs simply the face characteristic parameter of real-time people's face data and storage is compared, if confidence level surpasses the M value of an appointment, then compare successfully, identity obtains confirming.Application scenarios in this patent is that the visitor provides an ID (member's job number), from the face characteristic parameter library of storage, obtain this visitor's face characteristic parameter simultaneously automatically according to this ID (member's job number), and compare, thereby judge whether the individuality corresponding with the ID that is provided is same individual to this visitor with the facial image of real-time collection and through the face characteristic parameter after the feature extraction.The computing that people's face is differentiated is carried out on background computer 9, the foreground embedded system sends the facial image of camera acquisition and the ID of visitor's input (member's job number) to background computer by wireless network, and ID that the background computer basis sends and the image that is comprising people's face carry out people's face and differentiate;

Accompanying drawing 2 is main functional modules that people's face of realization on background computer 9 is differentiated: 1) image collection module, this module are finished and are obtained image, and image comes from embedded system; 2) people's face detection and location module, this module finds the definite position of people's face, and further people's face is separated from background, be output as and cut apart good people's face topography, the whole discrimination process that is operated in of this part is very important, the work of this part can become fairly simple under specific circumstances, among the present invention by the concrete imaging conditions of control (the people's face that guarantees shooting be positive, illumination condition is constant, image background is simple as far as possible), it is oversimplified, make the location ratio of people's face be easier to; 3) image pretreatment module, the main effect of this module are to remove as much as possible or reduce illumination, imaging system, external environment condition or the like interference for pending image, for subsequent treatment provides high-quality image.This part is finished the geometrical normalization of the facial image that extracts, eliminates processing such as noise and gray scale normalization, make in the different images people be bold little and brightness unified so that under identical conditions, finish training and differentiate; 4) feature extraction and select module, this module be by extracting the feature that is used for differentiating according to certain strategy in the pretreated facial image of above-mentioned image, with the data mapping in original people's face space to feature space.Because original image data amount is sizable, differentiate in order to realize classification effectively, will carry out conversion to raw data, obtain reflecting the feature of classification essence.How extracting stable and effective feature is the identification system key of success, and this patent adopts statistical nature face discrimination method, and this eigenface discrimination method is the proper vector structural attitude face that utilizes the image correlation matrix; 5) training module, the parameter that after this module is carried out and finished generation be can be used for differentiating, the discriminating of people's face can be regarded as people's face object to be identified is grouped in a certain class, therefore the effect of training module is exactly the sorter of classifying and differentiating, on people's face sample training collection basis, determine certain decision rule, make the loss minimum that the wrong resolution that is caused being classified by people's face discriminating object by this decision rule is minimum or cause; 6) identification module is finished the differentiation work of people's face according to the parameter of training gained, provides last identification result, and makes corresponding judgment.People's face identification system generally can be divided into two processes substantially, and as shown in Figure 2, dotted line the first half is a training process, and this process has been finished the design of sorter; Dotted line the latter half is people's face discrimination process.

People's face changes with age growth, and facial image is influenced by illumination, imaging angle and image-forming range etc., and this all multifactor human face identification technique that makes is complicated.Therefore differentiate success ratio for needing to adopt a kind of any special measures to improve in the gate inhibition of human face identification technique and the Work attendance device, reduce wrong resolution, specific practice is to adopt basic and the irrelevant Fisher face differentiation algorithm of human face expression on identification algorithm; In image data base and face characteristic database, need a kind of update strategy because therefore people's face changes with age growth, make that the face characteristic data are also upgraded in time with age; Eliminating aspect the influence that facial image is subjected to illumination, imaging angle and image-forming range, be provided with a minute surface in the gate inhibition of human face identification technique and the Work attendance device, it above the minute surface camera head, the right, middle part of minute surface is an input keyboard, the effect of input keyboard is that the employee is required to import ID number of employee, stranger person is required to import visiting key, adopt this mode can guarantee that the captured facial image of camera head is positive and image-forming range also can be controlled in certain scope; In order to reduce the influence of illumination to the facial image resolution, on described minute surface, be furnished with an illuminating device, make captured people face bit image be subjected to the influence of illumination to be reduced to bottom line.

Described people's face detection and location module, be used for the colour of skin of people's face is located face so that obtain the rectangular area, the colour of skin of people's face is people's face surface one of notable attribute the most, studies show that, people's skin color is distributed in the color space cluster in a relative small range, locate face with the colour of skin in this patent, this detection method has attitude unchangeability, simple and easy to do characteristics.Its gordian technique is to make up a kind of actual complexion model that utilizes, adopt the versicolor complexion model of non-linear segmentation in this patent by propositions such as Anil K.Jain, by the complexion model that this non-linear segmented color color conversion obtains, this model belongs to the Clustering Model in the color space.

YCbCr color form has and the similar principle of compositionality of human visual perception process, has the advantage that the luminance component in the color can be separated, but in the YCbCr color space, colour of skin cluster is to be biapiculate spindle shape, just in the bigger and less part of Y value, colour of skin cluster areas is reduction thereupon also, therefore must consider the different influences that cause of Y value, thereby YCbCr color form is carried out non-linear segmented color color conversion to improve the robustness of complexion model.What obtain at camera head is the rgb color form, at first the rgb color format conversion will be become YCbCr color form, obtains YCb ' Cr ' color space through non-linear segmented color color conversion then; The RGB coordinate space is as follows to the transformation for mula of YCb ' Cr ' coordinate space to YCbCr coordinate space and YCbCr coordinate space:

Y＝0.29990*R+0.5870*G+0.1140*B

Cr(Y)＝0.5000*R-0.4187*G-0.0813*B+128 (11)

Cb(Y)＝-0.1787*R-0.3313*G+0.5000*B+128

$C_i^{\&prime;}\left(Y\right)=\left\{\begin{array}{ccc}{C}_i\left(Y\right)& \mathrm{if}& Y\&Element;[{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1,K_h]\\ (C_i\left(Y\right)-\stackrel{\&OverBar;}{C_i}\left(Y\right))\frac{W_{C_i}}{W_{C_I}\left(Y\right)}+\stackrel{\&OverBar;}{C_i}\left(Y\right)& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\mathrm{orY}>K_h\end{array}\right.---\left(12\right)$

Wherein i represents b or r,

With The axis of expression area of skin color, can calculate with following formula:

$\stackrel{\&OverBar;}{C_b}\left(Y\right)=\left\{\begin{array}{ccc}108+\frac{(K_l-Y)(118-108)}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ 108+\frac{(Y-K_h)(118-108)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

$\stackrel{\&OverBar;}{C_r}\left(Y\right)=\left\{\begin{array}{ccc}154-\frac{(K_l-Y)(154-144)}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ 154+\frac{(Y-K_h)(154-132)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

K _lAnd K _hThe versicolor segmentation thresholding of the linear segmentation of right and wrong: K _l=125, K _h=188.Y _MinAnd Y _MaxBe the minimum and the maximal value of Y component in the colour of skin cluster areas that obtains by people's face experimental data: Y according to the somewhere _Min=16, Y _Max=235.

W _Cb(Y) and W _Cr(Y) width of expression area of skin color:

$W_{C_i}\left(Y\right)=\left\{\begin{array}{ccc}{\mathrm{WL}}_{C_i}+\frac{(Y-Y_{\min })(W_{C_i}-{\mathrm{WL}}_{C_i})}{K_l-Y_{\min }}& \mathrm{if}& Y<{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">K</figure-callout>}}_1\\ {\mathrm{WH}}_{C_i}+\frac{(Y_{\max }-Y)(W_{C_i}-{\mathrm{WH}}_{C_i})}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}\right.$

W _Ci, WL _Ci, WH _CiBe constant, be respectively: $W_{C_b}=46.97,$ ${\mathrm{WL}}_{C_b}=23,$ ${\mathrm{WH}}_{C_b}=14,$ $W_{C_r}=38.76,$ ${\mathrm{WL}}_{C_r}=20,$ ${\mathrm{WH}}_{C_r}=10.$

Described image pretreatment module is used to remove or reduces illumination, imaging system, external environment condition or the like interference for pending image, for subsequent treatment provides high-quality image; After the binary map after in above-mentioned people's face detection and location resume module, obtaining skin color segmentation, then to carry out denoising and handle, obtain maximum zone, vertical projection, determine people's face width, determine treatment scheme such as the crown, obtain people's face rectangular area at last; Treatment scheme as shown in Figure 4; Described denoising is handled the error in judgement that segmentation result brought that is used to reduce owing to based on complexion model, because can have some noise spots in the binary map after the skin color segmentation.

Described denoising is handled, and has adopted the filtering method based on pixel " density " to remove noise here.This is the filtering method of a kind of similar " convolution ".Say that intuitively add up the number of skin pixel exactly in the 5*5 neighborhood that with each pixel is the center, central point is left the colour of skin when surpassing half, otherwise thinks it is the non-colour of skin.Through removing the noise spot in the image after the denoising processing better;

The described zone that obtains maximum is to passing through the skin color segmentation figure after denoising is handled, the object in publishing picture along regular mark according to eight fillets, calculate the area of each object, the object that keeps the area maximum, and it is filled, obtain filtering and filled later skin color segmentation image;

Described vertical projection is according to formula (13) later skin color segmentation image to be filled in filtering to carry out vertical projection, and calculates the average Mean of each row non-zero projection value;

$\mathrm{Me\&alpha;n}=\frac{\underset{x=1}{\overset{M}{\mathrm{\&Sigma;}}}{P}_y\left(x\right)}{\mathrm{num}}---\left(13\right)$

Wherein num is P _yThe number of the non-zero points (x).

Selected threshold is half of Mean, if projection value greater than threshold value, then keeps; Otherwise put 0.Can remove the ear part substantially like this.The zone that keeps the area maximum is as last projection output;

Described definite people's face width for the above-mentioned vertical projection diagram that obtains, is looked for first non-0 point from left to right, and its corresponding columns promptly is the left margin of people's face; Equally, turn left from the right side and look for first non-0 point, the right margin of the columns behaviour face that it is corresponding.Distance between the border, the left and right sides is exactly the width of people's face;

Described definite crown for the above-mentioned vertical projection diagram that obtains, from top to bottom, is sought first skin pixel and is counted out greater than 15 row, determines that this journey is the crown;

Described people's face rectangular area that obtains is by getting people's face height=people's face width * 1.4, according to the crown of determining in the top step and the corresponding former figure of people's face width, cutting out last people's face rectangular area, as shown in Figure 5.

Described feature extraction and select module is used for the pretreated facial image of above-mentioned image is extracted the feature that is used for differentiating according to certain strategy, with the data mapping in original people's face space to feature space; Adopted Fisher face algorithm to carry out the feature extraction of people's face among the present invention;

Described Fisher face algorithm, it is a kind of algorithm that PCA combines with FLD, promptly at first utilizes PCA that the sample of higher dimensional space is projected to lower dimensional space guaranteeing that scatter matrix is nonsingular in the sample class, and then adopts Fisher linear discriminant analysis method.As shown in Figure 3.

Specific practice is at first facial image to be projected to the subspace by the PCA method, makes image space be converted into best description feature space; And then adopt the Fisher linear discriminant analysis that the MEF space is converted into best diagnostic characteristics space.

The projection matrix of Fisher face algorithm is:

${W_{\mathrm{opt}}}^T=W_{\mathrm{fld}}^T{W}_{\mathrm{pca}}^T---\left(1\right)$

In the formula:

W _pca＝argmax|W ^TS _TW|

$W_{\mathrm{fld}}=\arg \max \left|\frac{W^T{W}_{\mathrm{pca}}^T{S}_B{W}_{\mathrm{pca}}W}{W^T{W}_{\mathrm{pca}}^T{S}_W{W}_{\mathrm{pca}}W}\right|$

Column vector be orthogonal vector, be called the Fisher face.Face images projects to W in the training set _OptThereby column vector on obtain a stack features of every width of cloth facial image, these proper vectors can be directly used in classification.

The specific implementation step of Fisher face algorithm in this patent is described below.Earlier the meaning of sample and parameter symbol is done schematic illustration, sets forth from the training of people's face and discriminating two parts again:

1) resolution of everyone face training sample (facial image) is m*n, and it can be regarded as the column vector of a mn.

2) all training samples add up to N, the total number of persons that training sample is concentrated is P.

3) individual human face training sample x _i ^kTo expression, i represents i people, and k represents i people's k sample.

4) I total n of people _iIndividual training sample, visible N=n ₁+ n ₂+ ... + n _p

5) people's face sample to be identified is represented with y, and the resolution of sample is identical with training sample.

Described training module is used for people's face is classified; The parameter that after this module is carried out and finished generation be can be used for differentiating, the discriminating of people's face can be regarded as people's face object to be identified is grouped in a certain class, so the effect of this module is exactly the sorter of classifying and differentiating; Training process in training module is divided into two stages.What the phase one (Step1-Step6) mainly finished is the PCA training, it makes sample be compressed to lower dimensional space (a dimension) from higher dimensional space (mn dimension), and the principal character that has kept higher dimensional space, subordinate phase (Step7-Step12) is the Fisher classification based training, it makes similar sample intensive as far as possible, different class samples as far as possible separately, and used people's face parameter when obtaining differentiating.

Step1: it is right to import n people's face training sample, i=1, and 2 ..., p; K=1,2 ... n;

Step2: calculate the mean vector m of all samples and the mean vector (i people's the average image vector) of i class people face sample

$m=\frac{1}{N}\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{x}_i^k\&Element;R^{\mathrm{mn}\&times;1}---\left(2\right)$

$m_i=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{x}_i^k\&Element;R^{\mathrm{mn}\&times;1}---\left(3\right)$

Step3: the generation matrix of asking Karhunen-Loeve transformation

$S_b=\underset{i=0}{\overset{P-1_i}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)(m_i-m){(m_i-m)}^T----\left(4\right)$

P (ω in the following formula _i) be ω _i(i=1,2 ..., the c) prior probability of quasi-mode;

Step4: according to the theory of SVD, scatter matrix S between compute classes _bEigenvalue _iCharacteristic of correspondence vector u with it _i

Step5: get maximum preceding a (the individual eigenwert characteristic of correspondence vector of a≤p), structural matrix

U＝[u ₁，u ₂，…，u _a]∈R ^mn×a (5)

Step6: calculate everyone face training sample at u ₁, u ₂..., u _aProjection under the formed subspace, note x _i ^kAfter projection z _i ^k∈ R ^{A * 1}, formula (6) is then arranged

$Z=|{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">z</figure-callout>}}_1^1,{\mathrm{<figure-callout\; id="1"\; label="z"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">z</figure-callout>}}_1^2,...,z_1^{{\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">n</figure-callout>}}_1},{\mathrm{<figure-callout\; id="2"\; label="z"\; filenames="A20061015499600351.png"\; state="\{\{state\}\}">z</figure-callout>}}_2^1,...,z_p^{n_p}|\&Element;R^{a\&times;N}---\left(6\right)$

Step7: calculate the mean vector m of all kinds of people's face samples under the subspace _PcaiMean vector m with all classes _Pca, computing formula is shown in (7), (8):

$m_{\mathrm{pcai}}=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{Z}_i^k\&Element;R^{a\&times;1},i=\mathrm{1,2},...n_c---\left(7\right)$

$m_{\mathrm{pca}}=\underset{i=1}{\overset{p_i}{\mathrm{\&Sigma;}}}{m}_{\mathrm{pcai}}\&Element;R^{a\&times;1}---\left(8\right)$

Step8: calculate scatter matrix S in the total class under the subspace _wAnd scatter matrix S between class _b, computing formula is shown in (9), (10):

$S_w=\underset{i=1}{\overset{p}{\mathrm{\&Sigma;}}}{S}_i=\underset{i=1}{\overset{p_i}{\mathrm{\&Sigma;}}}\left[\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{(z_i^k-m_{\mathrm{pcai}})(z_i^k-m_{\mathrm{pcai}})}^T\right]\&Element;R^{a\&times;a}---\left(9\right)$

$S_b=\underset{l=1}{\overset{p_i}{\mathrm{\&Sigma;}}}(m_{\mathrm{pcai}}-m_{\mathrm{pca}}){(m_{\mathrm{pcai}}-m_{\mathrm{pca}})}^T\&Element;R^{a\&times;a}---\left(10\right)$

Step9: know S by inference according to the Fisher criterion function _bw ^*=λ S _ww ^*, it is a matrix S _bWith respect to matrix S _wGeneralized eigenvalue problem, find the solution S _bWith respect to matrix S _wGeneralized eigenvector, w ^*, i=1,2 ..., p

Step10: get generalized eigenvector and form matrix W and matrix U and multiply each other and obtain new matrix T, i.e. T=UW, the column vector of T is the result of training, keeps T and uses when to be identified.

Step11: calculate everyone the projection meanS of average face under the T of subspace, keep the projection coordinate coefficient of meanS, use when to be identified as this type of sample.

Step12: adopt the method for Euclidean distance, determine the similarity threshold value, use when the reservation threshold value is to be identified according to reject rate as similarity measurement.

Described identification module, be used to differentiate people's face, finish the discriminating work of individual people's face by this module according to the parameter of above-mentioned training gained, main core content is to differentiate algorithm in the identification module, because discrimination process and training process are closely-related, on the basis of above-mentioned training, differentiating algorithm can be described as:

Step1: calculate sample y to be identified at t ₁, t ₂..., t _C-1Projection coordinate's coefficient under the subspace that is generated

Step2: calculate The projection coordinate coefficient corresponding with this type of sample

Between similarity measurement value d.

Step3: compare the size of the similarity threshold value that obtains in d and the training process,, then declare people's face sample to be identified and pass through, otherwise refuse to know if d is littler than threshold value.

From above-mentioned people's face discrimination method, the subject matter that influences the resolution of people's face is: the obtaining value method of projection coordinate's coefficient of people's face sample correspondence; The span of the size of similarity threshold value; Can reflect the means of people's face with the adaptive updates face characteristic parameter of variations such as age; Just knowledge rate that people's face is differentiated and reject rate are the leading indicators of estimating people's face identification system;

The obtaining value method of described projection coordinate coefficient, its key point are the subspace dimensions of projection coordinate, and resolution increases gradually along with the increase of subspace dimension, and the group space dimensionality is during to some, and resolution tends towards stability.The subspace dimension adopts can both obtain reasonable resolution below 17 more than 5, in this interval, increase along with the subspace dimension, resolution also increases, and therefore differentiates that for the backstage computer computation ability that is adopted than under the condition with higher, adopts higher subspace dimension to suggestion among the present invention.

Described similarity threshold, its threshold size of choosing has determined the security and the efficient of system; Similarity threshold is excessive, can make the personnel that had authority to pass through originally can not obtain authentication, can't pass through; And, security being reduced if similarity threshold is too small, the personnel that originally have no right to pass through also obtain authentication.Weigh similarity threshold quality whether pointer and mainly contain two: wrong percent of pass and false rejection rate on this threshold value.Generally both are mutual restriction, and false rejection rate can be high relatively when the mistake percent of pass was low relatively, and the mistake percent of pass can be high relatively when false rejection rate was low relatively.The user interface that designs in this patent can be determined similarity threshold neatly according to the height of reject rate.

The means of described adaptive updates face characteristic parameter, be to utilize at every turn to obtained facial image in people's face discrimination process, facial image in the facial image database is upgraded, face characteristic parameter in the face characteristic parameter library is upgraded, the mode of upgrading is that the characteristic parameter with up-to-date facial image of obtaining and facial image removes to replace the most outmoded facial image and the characteristic parameter of facial image, the time of upgrading is to utilize do not use the time period of gate inhibition and Work attendance device in night, such as beginning morning, the start time of upgrading can be set by user oneself, if in the absence about setting, system starts the adaptive updates module automatically since morning, makes facial image in the facial image database and the face characteristic parameter in the face characteristic parameter library remain on up-to-date state;

From daily work attendance is used, can obtain just knowledge rate, reject rate and the wrong percent of pass and the false rejection rate of people's face discriminating of certain employee, above-mentioned four indexs have only reject rate can obtain by the statistics of computing machine, and just knowledge rate, mistake percent of pass and the false rejection rate differentiated about people's face need lean on artificial mode to confirm, wherein false rejection rate is the easiest adds up; Adopt two kinds of methods to solve in order to reduce among false rejection rate the present invention: 1) repeatedly to gather facial image, repeatedly carry out the method that people's face is differentiated; 2) utilize the False Rejects facial image to come update image storehouse and characteristic parameter storehouse;

The described facial image of repeatedly gathering, repeatedly carry out the method that people's face is differentiated, device is when gathering facial image, the input keyboard of system discovery on the work attendance panel begins to gather facial image when input is arranged and it is kept in the storage unit of embedded system, when having imported oneself ID number and pressed the facial image of gathering facial image again when sending key and will originally being kept at simultaneously in the storage unit of embedded system, the user sends to background computer together, background computer obtains behind ID number of the employee of foreground input and two width of cloth facial images, background computer is retrieved the record in employee's table according to ID number, obtain this employee's name and this employee's face characteristic information, at this moment background computer carries out people's face discriminating calculating, send resulting employee's name to foreground by network simultaneously, the foreground embedded system is notified the user by man-machine interface, " ^{* *}Your identity is differentiated, please wait "; simultaneously the foreground embedded system is gathered facial image; and facial image is sent to background computer once more carry out people's face and differentiate; at this moment exist this employee's three width of cloth facial images on the background computer; background computer is at first differentiated press the facial image of being gathered when sending key the user; if differentiate by not carrying out following processing, if differentiate the first time not by then the facial image of being gathered the second time being carried out people's face discriminating second time, if still not by then the facial image of being gathered for the third time not being carried out people's face discriminating for the third time; Everyly run into people's face and differentiate by just thinking that identity differentiates that each like this facial image of gathering all is an independent event, can know that by calculating probability three reject rates that make a mistake simultaneously will be well below the reject rate that once makes a mistake;

To improve the just knowledge rate that people's face is differentiated equally, reduce wrong percent of pass and also can adopt and above-mentionedly repeatedly gather facial image, repeatedly carry out the method that people's face is differentiated;

The described False Rejects facial image that utilizes comes update image storehouse and characteristic parameter storehouse, specific practice is to adopt the artificial mode of confirming to find that certain discriminating is False Rejects by personnel department, and the facial image of this False Rejects replaced the most outmoded facial image in some facial image databases, and this employee's face characteristic parameter is upgraded.

Embodiment 2

Other modes are identical with embodiment 1, the situation when communication aspects also can adopt the mode of wired network communication, this mode to be suitable for enterprise's work attendance and gate control system relatively fixedly.

Embodiment 3

There is not the enterprise of internal lan network for some, also can only use the work attendance function on foreground, some information stores of employee are in embedded system, importing employee ID number the employee takes later on, with employee ID number that imports, time and captured facial image are stored in the mobile storage unit, enterprise's information copy of the employee who is stored in the mobile storage unit can being turned out for work is carried out work attendance to it to computing machine, though not carrying out people's face, this mode do not differentiate, but with employee's image and employee ID number, information such as attendance time are placed on record, can confirm afterwards that therefore goodish work attendance effect is also arranged.

Described microprocessor is an embedded system, and the user program module among the present invention is realized by C and Java language.

The invention effect that the above embodiments 1,2 and 3 are produced is that enforcement is convenient, investment is little, the work attendance stand-by period short, the correct resolution height of people's face, can merge with occurrences in human life attendance management software, gate control system more easily, be specially adapted to the work attendance of the bigger enterprise and establishment of flowability.

Claims (10)

1, a kind of gate inhibition and Work attendance device based on human face identification technique, comprise the camera that is used to read facial image, be used to gather the embedded system that transmits video information, be used to import the input equipment of identity information and be used to carry out that image is differentiated and the computing machine of staff attendance management, described camera connects embedded system, described embedded system connects input equipment, described embedded system is connected with computer data, described computing machine comprises: image collection module is used to obtain the video image that embedded system is gathered;

Facial image database is used for the facial image of storing sample;

It is characterized in that: described gate inhibition and Work attendance device also comprise the minute surface that is used to confirm people's face front position, and described camera is positioned at the top of minute surface, and the optimum visual scope and the minute surface of described camera are complementary;

Described computing machine also comprises:

People's face detection and location module is used for determining the definite position of people's face people's face being separated from background, is output as and cuts apart good people's face topography; The rgb color format conversion is become YCbCr color form, obtain YCb ' Cr ' color space through non-linear segmented color color conversion then; The RGB coordinate space is as follows to the transformation for mula of YCb ' Cr ' coordinate space to YCbCr coordinate space and YCbCr coordinate space:

Y＝0.29990*R+0.5870*G+0.1140*B

Cr(Y)＝0.5000*R-0.4187*G-0.0813*B+128

Cb(Y)＝-0.1787*R-0.3313*G+0.5000*B+128 (11)

$C_i^{\&prime;}\left(Y\right)=\{\begin{array}{ccc}{C}_i\left(Y\right)& \mathrm{if}& Y\&Element;[K_1,K_h]\\ (C_i\left(Y\right)-{\stackrel{\&OverBar;}{C}}_i\left(Y\right))\frac{W_{C_i}}{W_{C_i}\left(Y\right)}+{\stackrel{\&OverBar;}{C}}_i\left(Y\right)& \mathrm{if}& Y<K_1\mathrm{orY}>K_h\end{array}---\left(12\right)$

Wherein i represents b or r, With The axis of expression area of skin color, calculate with following formula:

${\stackrel{\&OverBar;}{C}}_b\left(Y\right)=\{\begin{array}{ccc}108+\frac{(K_l-Y)(118-108)}{K_l-Y_{\min }}& \mathrm{if}& Y<K_l\\ 108+\frac{(Y-K_h)(118-108)}{Y_{\max }-K_h}& \mathrm{if}& Y>k_h\end{array}$

$\stackrel{-}{C_r}\left(Y\right)=\{\begin{array}{ccc}154-\frac{(K_l-Y)(154-144)}{K_l-Y_{\min }}& \mathrm{if}& Y<K_l\\ 154+\frac{(Y-K_h)(154-132)}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}$

K _lAnd K _hThe versicolor segmentation thresholding of the linear segmentation of right and wrong; Y _MinAnd Y _MaxBe the minimum and the maximal value of Y component in the colour of skin cluster areas that obtains by people's face experimental data according to the somewhere;

W _Cb(Y) and W _Cr(Y) width of expression area of skin color:

$W_{C_i}\left(Y\right)=\{\begin{array}{ccc}{\mathrm{WL}}_{C_l}+\frac{(Y-Y_{\min })(W_{C_i}-{\mathrm{WL}}_{C_i})}{K_l-Y_{\min }}& \mathrm{if}& Y<K_l\\ {\mathrm{WH}}_{C_i}+\frac{(Y_{\max }-Y)(W_{C_i}-{\mathrm{WH}}_{C_i})}{Y_{\max }-K_h}& \mathrm{if}& Y>K_h\end{array}$

W _Ci, WL _Ci, WH _CiBe constant, be respectively: $W_{C_b}=46.97,$ ${\mathrm{WL}}_{C_b}=23,$ ${\mathrm{WH}}_{C_b}=14,$ $W_{C_r}=38.76,$ ${\mathrm{WL}}_{C_r}=20,$ ${\mathrm{WH}}_{C_r}=10;$

The image pretreatment module is used for the facial image that extracts is carried out geometrical normalization, eliminates noise and gray scale normalization processing, and unified people is bold little and brightness;

People's face training module is used for the facial image of image data base is projected to the subspace by the PCA method, makes image space be converted into best description feature space, and the projection matrix of Fisher face algorithm is:

${W_{\mathrm{opt}}}^T=W_{\mathrm{fld}}^T{W}_{\mathrm{pca}}^T---\left(1\right)$

In the formula:

W _pca＝arg?max|W ^TS _TW|

$W_{\mathrm{fld}}=\arg \max \left|\frac{W^T{W}_{\mathrm{pca}}^T{S}_B{W}_{\mathrm{pca}}W}{W^T{W}_{\mathrm{pca}}^T{S}_W{W}_{\mathrm{pca}}W}\right|$

Column vector is orthogonal vector, is called the Fisher face, and face images projects to W in the training set _OptColumn vector on obtain a stack features of every width of cloth facial image;

Adopt the Fisher linear discriminant analysis that the MEF space is converted into best diagnostic characteristics space again, similar sample is intensive as far as possible, different class samples as far as possible separately, and used people's face parameter when obtaining differentiating;

People's face identification module, the people's face parameter that is used for the people's face parameter that will extract and current personnel's image training compares, and confidence level surpasses assign thresholds M, then compares successfully, and identity obtains affirmation.

2, gate inhibition and Work attendance device based on human face identification technique as claimed in claim 1, it is characterized in that: in described people's face training module, the resolution of everyone face training sample is m*n, it is the column vector of a mn, all training samples add up to N, the total number of persons that training sample is concentrated is P; Individual human face training sample x _i ^kTo expression, i represents i people, and k represents i people's k sample; I the total n of people _iIndividual training sample, visible N=n ₁+ n ₂+ ... + n _pPeople's face sample to be identified is represented with y, and the resolution of sample is identical with training sample; The algorithm of training is:

Step1: it is right to import n people's face training sample, i=1,2 .., p; K=1,2 ..n;

Step2: calculate the mean vector m of all samples and the mean vector (i people's the average image vector) of i class people face sample

$m=\frac{1}{N}\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{x}_i^k$ $R^{\mathrm{mn}\&times;1}$ $---$ $\left(2\right)$

$m_i=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{x}_i^k\&Element;R^{\mathrm{mn}\&times;1}---\left(3\right)$

Step3: the generation matrix of asking Karhunen-Loeve transformation

$S_b=\underset{i=0}{\overset{P-1_i}{\mathrm{\&Sigma;}}}P\left({\mathrm{\&omega;}}_i\right)(m_i-m){(m_i-m)}^T---\left(4\right)$

P (ω in the following formula _i) be ω _i(i=1,2 ..., the c) prior probability of quasi-mode;

Step4: according to the theory of SVD, scatter matrix S between compute classes _bEigenvalue _iCharacteristic of correspondence vector u with it _i

Step5: get maximum preceding α (the individual eigenwert characteristic of correspondence vector of α≤p), structural matrix

U＝[u ₁，u ₂，…，u _a]∈R ^mn×a (5)

Step6: calculate everyone face training sample at u ₁, u ₂..., u _aProjection under the formed subspace, note x _i ^kAfter projection z _i ^k∈ R ^{A * 1}, formula (6) is then arranged

$Z=|z_1^1,z_1^2,...,z_1^{n_1},z_1^{n_1},...,z_p^{n_p}|\&Element;R^{a\&times;N}---\left(6\right)$

Step7: calculate the mean vector m of all kinds of people's face samples under the subspace _PcaiMean vector m with all classes _Pca, computing formula is shown in (7), (8):

$m_{\mathrm{pcai}}=\frac{1}{n_i}\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}{z}_i^k\&Element;R^{a\&times;1},i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;n_c---\left(7\right)$

$m_{\mathrm{pca}}=\underset{i=1}{\overset{P_i}{\mathrm{\&Sigma;}}}{m}_{\mathrm{pcai}}\&Element;R^{a\&times;1}---\left(8\right)$

Step8: calculate scatter matrix S in the total class under the subspace _wAnd scatter matrix S between class _b, computing formula is shown in (9), (10):

$S_W=\underset{i=1}{\overset{P}{\mathrm{\&Sigma;}}}{S}_i=\underset{i=1}{\overset{P_i}{\mathrm{\&Sigma;}}}\left[\underset{k=1}{\overset{n_i}{\mathrm{\&Sigma;}}}(z_i^k-m_{\mathrm{pcai}}){(z_i^k-m_{\mathrm{pcai}})}^T\right]\&Element;R^{a\&times;a}---\left(9\right)$

$S_b=\underset{i=1}{\overset{P_i}{\mathrm{\&Sigma;}}}(m_{\mathrm{pcai}}-m_{\mathrm{pca}}){(m_{\mathrm{pcai}}-m_{\mathrm{pca}})}^T\&Element;R^{a\&times;a}---\left(10\right)$

Step9: know S by inference according to the Fisher criterion function _bw ^*=λ S _ww ^*, it is a matrix S _bWith respect to matrix S _wGeneralized eigenvalue problem, find the solution S _bWith respect to matrix S _wGeneralized eigenvector, w ^*, i=1,2 ..., P

Step10: get generalized eigenvector and form matrix W and matrix U and multiply each other and obtain new matrix T, i.e. T=UW, the column vector of T is the result of training, keeps T and uses when to be identified;

Step11: calculate everyone the projection meanS of average face under the T of subspace, keep the projection coordinate coefficient of meanS, use when to be identified as this type of sample;

Step12: adopt the method for Euclidean distance, determine the similarity threshold value, use when the reservation threshold value is to be identified according to reject rate as similarity measurement;

In described people's face identification module, the algorithm of discriminating is:

Step1: calculate sample y to be identified at t ₁, t ₂..., t _C-1Projection coordinate's coefficient under the subspace that is generated

Step2: calculate

The projection coordinate coefficient corresponding with this type of sample

Between similarity measurement value d;

Step3: compare the size of the similarity threshold value M that obtains in d and the training process,, then declare people's face sample to be identified and pass through, otherwise refuse to know if d is littler than threshold value M.

3, gate inhibition and Work attendance device based on human face identification technique as claimed in claim 2 is characterized in that: described people's face training module also comprises: the value determining unit of projection coordinate's coefficient is used for determining the dimension of the subspace of projection coordinate.

4, gate inhibition and Work attendance device based on human face identification technique as claimed in claim 2, it is characterized in that: described people's face training module also comprises: the similarity threshold setup unit, be used for the selected threshold size, determine similarity threshold M according to wrong percent of pass and false rejection rate.

5, gate inhibition and Work attendance device based on human face identification technique as claimed in claim 2, it is characterized in that: described people's face training module also comprises: adaptive updates face characteristic parameter unit, be used for utilizing each to the obtained facial image of people's face discrimination process, facial image in the facial image database is upgraded, face characteristic parameter in the face characteristic parameter library is upgraded, removed to replace the most outmoded facial image and the characteristic parameter of facial image with the characteristic parameter of up-to-date facial image of obtaining and facial image;

6, gate inhibition and Work attendance device based on human face identification technique as claimed in claim 2, it is characterized in that: described people's face identification module also comprises: repeatedly gather facial image pedestrian's face discriminating unit of going forward side by side, be used for when input equipment has input, beginning to gather facial image and it being kept at the storage unit of embedded system, when having imported identity information and pressed the facial image of gathering facial image again when sending key and will originally being kept at simultaneously in the storage unit of embedded system, the user sends to background computer together, after background computer obtains two width of cloth facial images, carry out people's face and differentiate calculating, simultaneously embedded system is gathered facial image, and facial image is sent to background computer once more carries out people's face and differentiate; At first differentiate,, finish if differentiate and pass through to press the facial image of being gathered when sending key the user; Do not have to pass through end if differentiate if differentiate the first time by then the facial image of being gathered for the second time being carried out people's face discriminating second time; If still not by then the facial image of being gathered for the third time being carried out people's face discriminating for the third time.

7, as one of claim 1-6 described gate inhibition and Work attendance device based on human face identification technique, it is characterized in that: described image pretreatment module comprises:

The denoising processing unit, be used for adopting filtering method to remove noise based on pixel " density ", the filtering method of similar " convolution " is added up the number of skin pixel in the 5*5 neighborhood that with each pixel is the center, central point is left the colour of skin when surpassing half, otherwise thinks it is the non-colour of skin;

Determine maximum territory element, be used for, the object in publishing picture along regular mark according to eight fillets the skin color segmentation figure after handling through denoising, calculate the area of each object, the object that keeps the area maximum, and it is filled, obtain filtering and filled later skin color segmentation image;

The vertical projection unit is used for according to formula (13) later skin color segmentation image being filled in filtering and carries out vertical projection, and calculates the average Mean of each row non-zero projection value;

$\mathrm{Mean}=\frac{\underset{x=1}{\overset{M}{\mathrm{\&Sigma;}}}{P}_y\left(x\right)}{\mathrm{num}}---\left(13\right)$

Wherein num is P _yThe number of the non-zero points (x);

Selected threshold is half of Mean, if projection value greater than threshold value, then keeps; Otherwise put 0; The zone that keeps the area maximum is as last projection output;

Determine people's face width unit, be used for the vertical projection diagram that obtains above-mentioned, look for first non-0 point from left to right, its corresponding columns promptly is the left margin of people's face; Equally, turn left from the right side and to look for first non-0 point, the right margin of the columns behaviour face that it is corresponding, the distance between the border, the left and right sides is exactly the width of people's face;

Determine unit, the crown, be used for the vertical projection diagram that obtains above-mentioned, from top to bottom, seek first skin pixel and count out, determine that this journey is the crown greater than 15 row;

Obtain unit, people's face rectangular area, be used for according to the crown of determining in the top step and the corresponding former figure of people's face width, cutting out last people's face rectangular area by getting people's face height=people's face width * 1.4.

8, as one of claim 1-6 described gate inhibition and Work attendance device based on human face identification technique, it is characterized in that: described gate inhibition and Work attendance device also comprise the houselights unit, described houselights unit is arranged on the top of minute surface, and described illuminating lamp unit is positioned at the top of camera.

9, as one of claim 1-6 described gate inhibition and Work attendance device based on human face identification technique, it is characterized in that: described embedded system comprises: the image recording module is used for person's to be measured video information recording is got off; Image processing module, the video data that is used for noting carry out compressed encoding, multiplexing and be modulated into compressed video data; First radio receiving transmitting module is used for according to communication standard, sends the video data of compression; Described computing machine comprises: second radio receiving transmitting module, be used for according to communication standard, and receive the video data of compression; The image decompression processing module, the data that are used for receiving decompress, demultiplexing and demodulation, revert to video data.

10, as one of claim 1-6 described gate inhibition and Work attendance device based on human face identification technique, it is characterized in that: the described embedded system that is used for the video image processing is connected with computing machine by network.

Images