CN101526997A - Embedded infrared face image identifying method and identifying device - Google Patents

Abstract

The invention relates to an embedded infrared face image identifying method and an identifying device. The method comprises: acquiring, displaying and identifying infrared face images, and outputting results; positioning face in the images; automatically extracting face characteristic information; and automatically training a face characteristic information base and outputting automatic face identifying algorithm and results. In the method, the face information can be utilized to achieve automatic acquisition and identification of the face images under all weather conditions and achieve the judgment of personal identity so as to meet the application requirements of commercial/domestic entrance guard, attendance record, identity authentication and the like. The method can be mutually fused with the prior other biological characteristic identifying methods and security systems to form personal identity identification and safeguard systems with completer functions. The method can also support the research on face identifying application systems in other fields such as face identifying case solution in the criminal investigation, face analytical diagnosis in the medical and health field, anti-fatigue driving in the traffic field and the like.

Description

Embedded infrared face image-recognizing method and recognition device

[technical field]:

The invention belongs to the biometrics identification technology field, relate generally to a plurality of sub-fields such as Digital Image Processing, infrared face identification, pattern recognition system optimization, Embedded System Design.

[background technology]:

Recognition of face is the biometrics identification technology that rises gradually in recent years, can be used for personal identification evaluation/identification, in a lot of applications very extensive market prospects is arranged.Present face recognition technology mainly comprises the gray scale/coloured image method of gathering based on the visible light sensing, and based on the face recognition technology of infrared imaging.The face recognition technology of infrared imaging has been represented the emerging direction in this field, compares with visual light imaging, and the infrared imaging recognition of face has following characteristics:

1. illumination invariant feature: the variation of light intensity and the variation of light-source angle do not influence the result of infrared imaging.This makes recognition of face that higher stability be arranged, and shows that according to foreign scholar's contrast experiment under same illumination model, the recognition of face precision of infrared imaging is higher by 30% than the precision of visual light imaging.

2. anti-camouflage characteristic: the vascular distribution of people's face is the main thermal source of infrared imaging, and this vascular distribution can not have more stability because of thermal imaging mechanism makes the face recognition technology of infrared imaging compare with the visible light technology because of cosmetic, lift face and camouflage change.

3. orientation independent: based on the face recognition technology of visible light position, the angle of people's face imaging had very high requirement, but for infrared imaging, this factor can not influence the precision of infrared imaging face recognition technology.

4. temperature sensitivity: thermal imaging mechanism makes infrared imaging be subjected to Temperature Influence easily, for recognition of face, different moods, phychology and the different fluctuations that all can cause infrared imaging season, the main method that overcomes this fluctuation is the training sample that upgrades in time, suitably takes image fusion technology simultaneously.

The face recognition technology of infrared imaging is compared with the face recognition technology based on visible light, has higher accuracy and processing speed faster.Application is comparatively widely arranged in fields such as gate inhibition's security protection, video monitoring, personal biology characteristics identifications.

[summary of the invention]:

The present invention seeks to solve the integrated infrared face recognition technology application problem of soft or hard, a kind of embedded infrared face image-recognizing method and recognition device are provided.

Embedded infrared face image-recognizing method provided by the invention comprises:

1st, the infrared face image detects and the location automatically

1.1st, infrared face image acquisition: utilize the infrared thermal imaging device, adopt the passive infrared formation method, 256 gray scale infrared face images of wide * height=320 * 240 pixels that collection is of a size of should comprise imaging background and people's head complete image information in the image;

1.2nd, human face region is located automatically: at the infrared face image of being gathered, use the adaptive classification algorithm to set up the Detection of Existence method of people's face in the image and the automatic accurate positioning method of human face region, this method can be rejected the image that does not comprise people's face automatically, and extracts the human face region image that only comprises face's information automatically from the infrared face image that comprises background;

1.3rd, positioning result normalization: the human face region image transitions is become to be of a size of 256 gray level images of wide * height=25 * 20 pixels, be used for follow-up recognition of face;

The step of the automatic accurate positioning method of the Detection of Existence method of people's face and human face region is as follows in the image described in the 1.2nd step:

1.2.1, facial image detect the training process with the location sorter

1.2.1.1, images acquired sample set (x _i, y _i), x wherein _iRepresent an image pattern, y _iRepresent whether this image pattern is facial image, being then is 1, is not to be 0 then, makes that n is the sum of image pattern.It is 5000 that the present invention selects n, and wherein the facial image sample is 3000,2000 of non-face image patterns; Human face region position in the manual sign facial image sample, definition human face region be R (W, H), wherein W is the width of human face region, H is the height of human face region, selection W is 80 among the present invention, H is 100;

1.2.1.2, initial weight W is set for each width of cloth image pattern _{1, i}, wherein subscript 1 is represented the number of times that weights are provided with, the numbering of i presentation video sample in sample set.Suppose that the concentrated facial image of image pattern adds up to L, non-face total number of images is M.Then the initial weight of face images is 1/2L, and the initial weight of non-face image is 1/2M.Set L=3000 among the present invention, M=2000;

1.2.1.3, use Haar feature are carried out the sorter training, Haar is characterized as rectangular characteristic, have 14 kinds of three classes, its definition as shown in Figure 3, to each Haar feature, calculate the difference of deceiving pixel value sum and white pixel value sum in its corresponding region, this difference is the value of this Haar feature.To each pixel in the human face region, all add up 14 kinds of Haar features of its three class, thereby obtain all Haar eigenwerts of whole human face region.

1.2.1.4 is based on the sorter training of Adaboost method.Adopt the method for iteration training to carry out detection of people's face and location sorter training, set up two layers of classified body system, wherein ground floor is a strong classifier, form the cascade pattern between a plurality of strong classifiers, each strong classifier is divided into 14 Weak Classifiers, each Weak Classifier comprises 1 Haar feature, and it is 10 that the present invention uses the strong classifier number, carries out following iteration training flow process:

To t=1,2,3 ... 10,

1.2.1.4.1, find the solution the normalization weights q of every width of cloth image pattern _{T, i}, wherein t is an iterations, the numbering of i presentation video sample in sample set;

$q_{t,i}=\frac{w_{t,i}}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{tj}}}$

1.2.1.4.2, (p Θ), makes it can distinguish facial image feature and non-face characteristics of image for x, f, calculates the weighting (q of corresponding all Weak Classifiers to train each Weak Classifier h _i) error rate ε _f

${\mathrm{\ϵ}}_f=\underset{i}{\mathrm{\Σ}}{q}_i|h(x_i,f,p,\mathrm{\θ})-y_i|$

1.2.1.4.3, the Weak Classifier of weighting error rate minimum is chosen for best Weak Classifier, the minimum weighting error rate of order is ε _t, pairing Weak Classifier is h _t

1.2.1.4.4, upgrade the weights of each image pattern

$w_{t+1,i}=w_{t,i}{\mathrm{\β}}_t^{1-\mathrm{ei}}$

Wherein, if current Weak Classifier can correctly be discerned x _i, ei=0 then; Otherwise, ei=1 and ${\mathrm{\β}}_{\mathrm{\τ}}=\frac{{\mathrm{\Σ}}_{\mathrm{\τ}}}{1-{\mathrm{\ϵ}}_{\mathrm{\τ}}};$

1.2.1.4.5, the strong classifier that finally obtains are as follows

Wherein ${\mathrm{\α}}_{\mathrm{\τ}}=\log \frac{1}{{\mathrm{\β}}_{\mathrm{\τ}}};$

1.2.1.5, through the iteration training in 1.2.1.4 step, can obtain to comprise the two-stage classification body system of 10 strong classifiers and 140 Weak Classifiers, based on this sorter system, can realize the people's face Detection of Existence and the location of the image gathered at the present invention.

1.2.2, facial image detect and location algorithm

1.2.2.1, to pending facial image, according to each pixel in the sequential scanning image of Row Column,, calculate the pairing Haar eigenwert of current pixel according to the Haar feature of using in 1.2.1.4 step.

1.2.2.2, the eigenwert to obtaining in 1.2.2.1 step use 10 strong classifiers to differentiate successively, if the differentiation result of certain strong classifier is 0, illustrate that the current Haar characteristic area of selecting for use of this pixel is not a human face region.If the differentiation result of 10 strong classifiers is 1 all, illustrate that the current Haar characteristic area of selecting for use of this pixel is a human face region.

1.2.2.3, all pixels in the image have all been carried out going on foot described differentiation process as 1.2.2.2 after, can obtain that all are differentiated the pixel region for people's face zone in the image, these pixel regions are communicated with, can obtain final human face region.

2nd, the infrared face image is discerned comparison automatically

2.1st, face characteristic extracts: based on the human face region image after the normalization, the image with 25 * 20 is converted into the column vector of one 500 dimension according to line scanning; Adopt principal component analytical method, obtaining the space size is the eigentransformation matrixes of 500 row * 100 row, and the vector that calculates the space size on this basis and be 100 row * 1 row is as the human face region image feature information;

It is as follows to adopt the principal component analysis method to carry out the algorithm steps of feature dimensionality reduction:

2.1.1,500 dimensional vectors are transformed to the matrix of 500*500, computation of mean values is obtained covariance;

2.1.2, ask covariance characteristic of correspondence value and characteristic of correspondence vector;

2.1.3, eigenwert is sorted from big to small, choose the eigenwert of final employing according to the accumulation contribution rate;

2.1.4, the selected eigenwert characteristic of correspondence vector of selection constitute transformation matrix;

2.1.5, for the data of new typing, directly multiply each other with this transformation matrix, obtain the data behind the dimensionality reduction of correspondence;

Wherein, accumulation contribution rate η _mComputing formula as follows: η _m=λ ₁+ λ ₂+ ...+λ _m/ (λ ₁+ λ ₂+ ...+λ _p)

λ in the formula is a proper vector characteristic of correspondence value, and m is the number that is chosen for the proper vector of major component, and p is the sum of proper vector; It is 99% that the present invention selects the contribution rate of accumulative total threshold value, and major component proper vector number is 100.

2.2nd, face characteristic training: all can be all necessary through features training by the facial image that native system is differentiated automatically; Native system provides the semi-automatic mode of man-machine interaction to gather people's face training image, and uses the method in the 2.1st step to extract face characteristic, and this feature will be kept in the characteristic information storehouse as standard feature, for the use of the 2.3rd step.

The detailed process of features training is as follows:

Suppose that embedded infrared face recognition methods of the present invention can differentiate A people's facial image,, gather lineup's face image pattern, make it add up to P as undergoing training sample to wherein everyone; Gather one group of other people face image pattern that does not belong to this A people simultaneously as the refusal training sample, make it add up to Q;

To each width of cloth training image sample, calculate according to the computing formula of covariance matrix, generate a symmetric matrix of one 500 * 500 dimension; This symmetric matrix is carried out finding the solution of eigenwert and proper vector, obtain transformation matrix, generate one 500 * 100 transformation matrix, Here it is eigentransformation matrix remembers that this matrix is M; When carrying out features training or when identification, at first image is converted to 500 dimensional vector Y according to the mode of line scanning, the vectorial Y behind the dimensionality reduction can obtain according to following formula easily:

Y=M ^TX, T is that matrix changes the order sign in the formula, X is 500 dimensional vectors of indication in the 2.1st step,

Obtain one 100 * 1 vector, in the features training process, the proper vector of the sample of undergoing training is kept in the characteristic information storehouse as the features training result, aspect ratio to identifying in, use this vector to carry out object matching;

Adopt the K Mean Method to set up the cluster system, with A people's facial image poly-be the K class, K=A among the present invention;

Adopt the iteration training method to calculate the classification thresholds S of each class, this threshold value can realize undergoing training correct receptance of sample is 99%, and the correct rejection ratio of refusing training sample is 99.9%.

2.3rd, face characteristic comparison identification: the method for utilizing for the 2.1st step described is extracted the proper vector of pending facial image, each standard feature in the characteristic information storehouse that obtains in itself and the 2.2nd step is carried out distance comparison, and according to the threshold value S that calculates in the 2.2nd step with judge face characteristic whether with the characteristic information storehouse in certain proper vector than in, as than in then return by than the pairing personally identifiable information of middle facial image, as not than in then return the refusal identifying information.

3rd, a kind of embedded infrared face pattern recognition device, this recognition device comprises:

Core processor assembly: comprise based on the S3c2440 primary processor of ARM920T kernel with based on the TMS320DM642 coprocessor of TMS320C6000DSP platform;

Wherein coprocessor TMS320DM642 is responsible for receiving and handling the video image of input, obtains the infrared face view data, and sends data to primary processor S3c2440, is finished the location and the discriminator of infrared face image by s3c2440;

Infrared image acquisition assembly: comprise infrared thermal imaging device and coding and decoding video chip; The infrared thermal imaging device is used for captured video image, and output simulating signal, the coding and decoding video chip is a digital signal with the analog signal conversion of infrared thermal imaging device output, and output format is that the digital video signal of YUV4:2:2 of ITU-R BY.656 is to the TMS320DM642 coprocessor;

Memory module:: use the SDRAM chip of 16 of two 32MB to be connected, be used for the primary processor data storage, use the NandFlash chip of a slice 64MB to be connected, be used for the procedure stores of primary processor with primary processor S3c2440 with primary processor S3c2440.Use the SDRAM chip of a slice 32MB16 position to be connected, be used for the coprocessor data storage, use the NandFlash chip of a slice 32MB to be connected, be used for the procedure stores of coprocessor with coprocessor TMS320DM642 with coprocessor TMS320DM642.

USB output interface assembly: the USB HOST controller that uses primary processor S3c2440 to carry is realized the USB output interface.

Power supply module: provide by 220V and exchange the direct power supply of the transformer that changes the 5V direct current and two kinds of patterns of powered battery, be used to drive the operation of hardware unit;

Miniature keyboard input module: be connected with primary processor, input password and steering order for the user;

LCD display unit: be connected with primary processor, be used to show facial image, identification comparison result;

Audio frequency output prompting assembly: be connected with primary processor, be used to provide voice suggestion, consumer-oriented operating process.

Advantage of the present invention and good effect:

The present invention is directed to recognition of face and application problem, designed and Implemented complete recognition methods system and hardware architecture, set up that the software and hardware height is integrated, robotization, intelligentized infrared embedded infrared face recognition system.

The recognition methods system of this system has covered complete process flow such as infrared face image detection, human face region framing, infrared face image characteristics extraction and dimensionality reduction, infrared face image recognition, and be cured in the embedded hardware structure this flow process is integrated, can satisfy the needs that the identification of all kinds infrared face is used, have the advantage that processing speed is fast, accuracy of identification is high, application scalability is strong.

The recognition methods of this system has realized the face characteristic dimensionality reduction algorithm based on principal component analytical method, uses the 10*1 dimensional feature to realize the high precision comparison of people's face, has improved the speed of recognition of face.

The hardware systems of this system covered analog video signal collection and transmission, phonetic entry output control, complicated power system design, USB interface design, diversiform data transmission mechanism design, based on the contents such as core controlling mechanism design of ARM+DSP.Have the advantages that volume is little, function is complete, versatility is good, integrated level is high.

This system can be used as flush bonding module and combines with other equipment, forms the equipment in integral biological characteristic identificating equipment, gate inhibition's security protection equipment and other recognition of face fields with infrared face recognition function, is with a wide range of applications.

[description of drawings]:

Fig. 1 is an infrared face localization method process flow diagram;

Fig. 2 is an infrared face image-recognizing method process flow diagram;

Fig. 3 is that facial image detects and the used Haar feature of the training synoptic diagram of locating sorter, and Fig. 3 (1) is 4 kinds of limit feature synoptic diagram, and Fig. 3 (2) is 8 kinds of line feature synoptic diagram, and Fig. 3 (3) is 2 kinds of central feature synoptic diagram;

Fig. 4 is to be images acquired (Fig. 4-a) and the people's face positioning result (comparison diagram of Fig. 4-b) of example with true man;

Fig. 5 is an embedded infrared face pattern recognition device structured flowchart.

[embodiment]:

Embodiment 1: the infrared face image-recognizing method

As shown in Figure 1, 2, infrared face image-recognizing method provided by the invention comprises infrared face image acquisition, people's face location, three parts of recognition of face comparison, is described as follows one by one:

1st, infrared face acquisition mode and images acquired specification

Gather facial image by the infrared image acquisition device, people's face is 15～100cm apart from the harvester distance, allows people's face that certain inclination and rotation are arranged.Wherein inclination being defined as with the neck is axle, and the head along continuous straight runs rotates to an angle left or to the right, allows the angle of inclination to be no more than 30 degree; It is axle that definition rotates to be with the neck, keeps the head horizontal level constant, and head rotates up or down, and the anglec of rotation requires to be no more than 10 degree.

Images acquired is a gray level image, and wherein the heat distribution of target is being represented in the distribution of gray scale.The high more place of people's face temperature, the gray-scale value of corresponding point is big more in the image.The image that device collects is divided into 256 gray levels, and promptly the target heat is by 256 levels of uniform discrete, and each rank is represented a gray-scale value.The image size that collects is 240 * 320, and promptly vertical direction comprises 240 pixels, and horizontal direction comprises 320 pixels.

In this example, the mode of images acquired is a video acquisition, promptly gathers the video of a period of time, as 20 seconds, according to Fixed Time Interval abstract image frame, sets up basic sample set then.

The 2.th people's face-positioning method (see figure 1)

The accurate location of people's face Detection of Existence and human face region is the primary link in the infrared face recognition methods, the problem of its processing is to confirm whether there is people's face in the image, if exist then determine the area coordinate at people's face place, if there is no people's face termination process then.

The inventive method adopts the method for adaptive classification to solve people's face orientation problem, the facial image detection and location sorter that have adaptive ability by training a large amount of facial image samples to set up, and use this sorter that image is handled automatically, obtain accurate human face region.

In this example, gather 18 people, everyone 120 pictures totally 2160 pictures altogether, human face region in the picture is carried out manual demarcation, and the zone of wherein choosing must comprise eyes and eyebrow, lip, cheek, simultaneously within the ears, the above zone of chin all will be in the calibration range.Randomly draw wherein 1000 pictures again, alternative is got 430 non-face picture training, the training degree of depth is set to 10 layers, finally choose 140 features, on 2160 complete sample set, correct locating effect can reach 2075/2160, tests on 2984 non-face images in addition, and the location of mistake ratio is 95/2984.

The general location effect:

Sample	Number of samples	Number is located in success	The no-fix number
				Aa	120	111	9
Aaa0	120	107	13
				Aaa1	120	119	1
Aab0	120	116	4
				Aab1	120	114	6
Csy	120	120	0
				Dhs	120	120	0
Lg	120	114	6

Lkm0	120	118	2
				Lkm1	120	120	0
Ryh	120	118	2
				Tzc	120	119	1
Yp	120	119	1
				Zkd	120	120	0
Zkp	120	93	27
				Zl	120	112	8
Zqw0	120	117	3
				Zqw1	120	118	2

Finish after the location, locating area need be unified size, choose bilinear interpolation method in this example the target area is normalized on 20 * 25 sizes to carry out subsequent treatment.Fig. 4-b is the human face region image through obtaining after the normalization.

3rd, face identification method

Based on the human face region image after the normalization, gray-scale value with its 500 pixels is a primitive character, adopt principal component analysis (Principle ComponentAnalysis, be called for short PCA) method, obtaining the space size is the eigentransformation matrix of 500 (line number) * 100 (columns), and the vector that calculates the space size on this basis and be 100 (line number) * 1 (columns) is as the human face region image feature information.It is as follows wherein to carry out the treatment scheme of feature dimensionality reduction based on PCA:

[face characteristic extracts and the dimensionality reduction flow process]

In following definition, establishing A is n rank matrixes, and x, y are n dimension non-vanishing vector, and k is a real number.

Definition covariance matrix (covariance) is

$\mathrm{cov}(x,y)=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})$

X wherein, y represents average, in covariance matrix, the variance of corresponding this this component of diagonal line, the covariance between each component of other lists of elements.

The definition variance is: $L_{\mathrm{xx}}=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(x_i-\stackrel{\‾}{x})}^2$

If Ax=kx is arranged, claim that then k is the eigenwert of upright A, x is a proper vector.

The proper vector that the 500 dimension * 1 that obtain based on the scanning facial image according to facial image are listed as is constructed its diagonal matrix.The flow process of its feature dimensionality reduction is as follows:

The compute matrix average is obtained covariance;

Ask covariance characteristic of correspondence value and characteristic of correspondence vector;

Eigenwert is sorted from big to small, choose the eigenwert of final employing according to the accumulation contribution rate;

Select selected eigenwert characteristic of correspondence vector, constitute transformation matrix;

For the data of new typing, directly multiply each other with this transformation matrix, can obtain the data behind the dimensionality reduction of correspondence;

Accumulation contribution rate computing formula is as follows: η _m=λ ₁+ λ ₂+ ...+λ _m/ (λ ₁+ λ ₂+ ...+λ _p)

Work as η _mDuring greater than threshold value S, can think that the major component number is m.Native system selects 100 as final feature master's composition number, and S is 99%.

The face characteristic training: all can be all necessary through features training by the facial image that native system is differentiated automatically.Native system provides the semi-automatic mode of man-machine interaction to gather people's face training image, and the image with 25 * 20 is according to line scanning, with the column vector of one 500 dimension of its conversion dimension.The image of at first selecting some is as training vector; Computing formula according to covariance matrix is calculated, and generates a symmetric matrix of one 500 * 500 dimension; This symmetric matrix is carried out finding the solution of eigenwert and proper vector, obtain transformation matrix, generate one 500 * 100 transformation matrix.Here it is eigentransformation matrix remembers that this matrix is M.When carrying out features training or when identification, at first image is converted to 500 dimensional vector X according to the mode of line scanning, the vectorial Y behind the dimensionality reduction can obtain according to following formula easily:

Y＝M ^TX，

Obtain one 100 * 1 vector, in the features training process, this vector is kept in the characteristic information storehouse as the features training result, aspect ratio to identifying in, use this vector to carry out object matching.

[face characteristic training flow process]

Make embedded infrared face recognition methods of the present invention can differentiate 17 people's facial image,, gather lineup's face image pattern, make it add up to 17 * 10 as undergoing training sample to wherein everyone; Gather one group of other people face image pattern that does not belong to these 17 people simultaneously as the refusal training sample, make it add up to 1 * 10;

Adopt the K Mean Method to set up the cluster system, be K class, K=17 among the present invention with 17 people's facial image is poly-;

Adopt the iteration training method to calculate the classification thresholds S of each class, in this example, choose S=515000, this threshold value can realize undergoing training correct receptance of sample is 99%, and the correct rejection ratio of refusing training sample is 99.9%.This threshold value will be used for face characteristic comparison identification.

Face characteristic comparison identification: at human face region image to be compared, adopt the face characteristic extraction algorithm of native system to obtain the human face region characteristics of image, and each proper vector of calculating in the people's face training image characteristic information storehouse that is obtained in this proper vector and the face characteristic training method is carried out distance relatively, calculate its Euclidean space distance, the threshold value S that obtains according to features training with judge face characteristic whether with the characteristic information storehouse in certain proper vector than in, as than in then return by than the pairing personally identifiable information of middle facial image, as not than in then return the refusal identifying information.

With Fig. 4-a is example, and images acquired is 320 * 240 sizes, through locating and carrying out after the normalization, obtains the target facial image of 20 * 25 sizes, shown in Fig. 4-b.Be the column vector of one 500 dimension again with this image transitions, project to transformation matrix, process and the sample calculation Euclidean distance of having registered, its distance is less than setting threshold value 515000, and this sample is by identification the most at last.

Embodiment 2: the recognition device design

Embedded infrared face pattern recognition device provided by the invention comprises:

Core processor assembly: comprise based on the S3c2440 primary processor of ARM920T kernel with based on the TMS320DM642 coprocessor of TMS320C6000DSP platform;

Wherein coprocessor TMS320DM642 is responsible for receiving and handling the video image of input, obtains the infrared face view data, and sends data to primary processor S3c2440, is finished the location and the discriminator of infrared face image by s3c2440;

Infrared image acquisition assembly: comprise infrared thermal imaging device and coding and decoding video chip; The infrared thermal imaging device is used for captured video image, and output simulating signal, the coding and decoding video chip is a digital signal with the analog signal conversion of infrared thermal imaging device output, and output format is that the digital video signal of YUV4:2:2 of ITU-R BT.656 is to the TMS320DM642 coprocessor;

Memory module:: use the SDRAM chip of 16 of two 32MB to be connected, be used for the primary processor data storage, use the NandFlash chip of a slice 64MB to be connected, be used for the procedure stores of primary processor with primary processor S3c2440 with primary processor S3c2440.Use the SDRAM chip of 16 of a slice 32MB to be connected, be used for the coprocessor data storage, use the NandFlash chip of a slice 32MB to be connected, be used for the procedure stores of coprocessor with coprocessor TMS320DM642 with coprocessor TMS320DM642.

USB output interface assembly: the USB HOST controller that uses primary processor S3c2440 to carry is realized the USB output interface.

Power supply module: provide by 220V and exchange the direct power supply of the transformer that changes the 5V direct current and two kinds of patterns of powered battery, be used to drive the operation of hardware unit;

Miniature keyboard input module: be connected with primary processor, input password and steering order for the user;

LCD display unit: be connected with primary processor, be used to show facial image, identification comparison result;

Audio frequency output prompting assembly: be connected with primary processor, be used to provide voice suggestion, consumer-oriented operating process.

Embodiment 3: true application flow

Solidified people's face finder and recognition of face comparison program in the recognition device of the present invention, backer's face information registering and face recognition application in the real world applications environment, applying step comprises:

1st, people's face information registering

The user keys in instruction by the miniature keyboard input module, start people's face information registering flow process, with the infrared thermal imaging device in people's face aligning infrared image acquisition assembly, send the image that collects to coprocessor TMS320C6000 by video decoding chip, after coprocessor is seen the frequency fractionation, obtain the infrared face image, send primary processor S3c2440 to.The real human face image of Fig. 4-a for collecting through people's face information registering, size is 240 (height) * 320 (wide).

Primary processor S3c2440 operation people face finder is realized the accurate extraction to human face region in the image, and extraction of operation face characteristic and dimensionality reduction program, obtains people's face standard feature, and is kept among the NandFlash of memory module.The human face region image of Fig. 4-b for after human face region location, normalization, exporting, size is 25 (height) * 20 (wide), this face characteristic is converted into 100 * 1 column vector and is kept among the NandFlash after dimension-reduction treatment.

In this flow process, the LCD display assembly will show the collection situation of facial image, and the display process state.Audio frequency output prompting assembly is finished with voice mode prompting user's operation steps and registration.

2nd, recognition of face comparison

The user keys in instruction by the miniature keyboard input module, start people's face comparison flow process, with the infrared thermal imaging device in people's face aligning infrared image acquisition assembly, send the image that collects to coprocessor TMS320C6000 by video decoding chip, carry out obtaining the infrared face image after video splits through coprocessor, send primary processor S3c2440 to.Primary processor S3c2440 operation people face finder is realized the accurate extraction to human face region in the image, and extraction of operation face characteristic and dimensionality reduction program, obtains the face characteristic vector.Continue operation recognition of face comparison program, face characteristic vector and the standard feature that is kept among the NandFlash compared, as than in then compare successfully by LCD display assembly and audio prompt output precision prompting identification, otherwise prompting identification is compared and is failed.

With Fig. 4 is example, and the people's face that shows among Fig. 4 has been registered in the system, has people's face information of other 16 people to be registered simultaneously.When this user aims at people's face infreared imaging device and starts the comparison flow process, primary processor S3c2440 will carry out the location and extract human face region, further extract face characteristic, with the standard feature storehouse in all 17 face characteristics whens comparison that are registered, by calculating the Euclidean distance of proper vector to be identified and standard feature vector, find that this face characteristic people face standard feature distance corresponding with Fig. 4 is the most approaching, function signal is compared in output identification, judges this people's identity simultaneously.

Claims (6)

1, a kind of embedded infrared face image-recognizing method is characterized in that this method comprises:

1st, the infrared face image detects and the location automatically

1.1st, infrared face image acquisition: utilize the infrared thermal imaging device, adopt the passive infrared formation method, gather 256 gray scale infrared face images, should comprise imaging background and people's head complete image information in the image;

1.2nd, human face region is located automatically: at the infrared face image of being gathered, set up the Detection of Existence method of people's face in the image and the automatically accurate location mechanism of human face region according to the Adaboost method, this mechanism can be rejected the image that does not comprise people's face automatically, and extracts the human face region image that only comprises face's information automatically from the infrared face image that comprises background;

1.3rd, positioning result normalization: the human face region image transitions is become to be of a size of 256 gray level images of wide * height=25 * 20 pixels, be used for follow-up recognition of face;

2nd, the infrared face image is discerned comparison automatically

2.1st, face characteristic extracts: based on the human face region image after the normalization, the image with 25 * 20 is converted into the column vector of one 500 dimension according to line scanning; Adopt principal component analytical method, obtaining the space size is the eigentransformation matrixes of 500 row * 100 row, and the vector that calculates the space size on this basis and be 100 row * 1 row is as the human face region image feature information;

2.2nd, face characteristic training: for realizing differentiating based on the identity of facial image, all can the facial image of differentiation all must be through features training automatically by native system; Native system provides the semi-automatic mode of man-machine interaction to gather people's face training image, and uses the method in the 2.1st step to extract face characteristic, and this feature will be kept in the characteristic information storehouse as standard feature; In the features training process, determine threshold value S simultaneously, for the use of the 2.3rd step;

2.3rd, face characteristic comparison identification: human face region image feature information and each standard feature in people's face training image characteristic information storehouse that the 2.1st step obtained are carried out distance relatively, calculate its Euclidean space distance, according to the threshold value S that obtains in the 2.1st step with judge face characteristic whether with the characteristic information storehouse in certain proper vector than in, as than in then return by than the pairing personally identifiable information of middle facial image, as not than in then return the refusal identifying information.

2, embedded infrared face image-recognizing method according to claim 1, it is characterized in that the Detection of Existence method of people's face in the image described in the 1.2nd step and the automatic accurate positioning method of human face region, be the facial image detection and location sorter that has adaptive ability by training a large amount of facial image samples to set up, and use this sorter that image is differentiated automatically, reject the image that does not comprise people's face, and from facial image, obtain accurate human face region, concrete steps are as follows:

1.2.1, facial image detect the training process with the location sorter

1.2.1.1, images acquired sample set (x _i, y _i), x wherein _iRepresent an image pattern, y _iRepresent whether this image pattern is facial image, being then is 1, is not to be 0 then, makes that n is the sum of image pattern.It is 5000 that the present invention selects n, and wherein the facial image sample is 3000,2000 of non-face image patterns; Human face region position in the manual sign facial image sample, definition human face region be R (W, H), wherein W is the width of human face region, H is the height of human face region, selection W is 80 among the present invention, H is 100;

1.2.1.2, initial weight W is set for each width of cloth image pattern _{1, j}, wherein subscript 1 is represented the number of times that weights are provided with, the numbering of i presentation video sample in sample set.Suppose that the concentrated facial image of image pattern adds up to L, non-face total number of images is M.Then the initial weight of face images is 1/2L, and the initial weight of non-face image is 1/2M.Set L=3000 among the present invention, M=2000;

1.2.1.3, use Haar feature are carried out the sorter training, Haar is characterized as rectangular characteristic, have 14 kinds of three classes, its definition as shown in Figure 3, to each Haar feature, calculate the difference of deceiving pixel value sum and white pixel value sum in its corresponding region, this difference is the value of this Haar feature.To each pixel in the human face region, all add up 14 kinds of Haar features of its three class, thereby obtain all Haar eigenwerts of whole human face region;

1.2.1.4 is based on the sorter training of Adaboost method, adopt the method for iteration training to carry out detection of people's face and location sorter training, set up two layers of classified body system, wherein ground floor is a strong classifier, form the cascade pattern between a plurality of strong classifiers, each strong classifier is divided into 14 Weak Classifiers, and each Weak Classifier comprises 1 Haar feature, it is 10 that the present invention uses the strong classifier number, carries out following iteration training flow process:

To t=1,2,3 ... 10,

1.2.1.4.1, find the solution the normalization weights q of every width of cloth image pattern _{T, i}, wherein t is an iterations, the numbering of i presentation video sample in sample set;

$q_{t,i}=\frac{w_{t,i}}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{tj}}}$

1.2.1.4.2, (p Θ), makes it can distinguish facial image feature and non-face characteristics of image for x, f, calculates the weighting (q of corresponding all Weak Classifiers to train each Weak Classifier h _i) error rate ε _f

${\mathrm{\ϵ}}_f=\underset{i}{\mathrm{\Σ}}{q}_i|h(x_i,f,p,\mathrm{\θ})-y_i|$

1.2.1.4.3, the Weak Classifier of weighting error rate minimum is chosen for best Weak Classifier, the minimum weighting error rate of order is ε _t, pairing Weak Classifier is h _t

1.2.1.4.4, upgrade the weights of each image pattern

$w_{t+1,i}=w_{t,i}{\mathrm{\β}}_t^{1-\mathrm{ei}}$

Wherein, if current Weak Classifier can correctly be discerned x _i, ei=0 then; Otherwise, ei=1 and ${\mathrm{\β}}_t=\frac{{\mathrm{\ϵ}}_t}{1-{\mathrm{\ϵ}}_t};$

1.2.1.4.5, the strong classifier that finally obtains are as follows

Wherein ${\mathrm{\α}}_t=\log \frac{1}{{\mathrm{\β}}_t};$

1.2.1.5, through the iteration training in 1.2.1.4 step, can obtain to comprise the two-stage classification body system of 10 strong classifiers and 140 Weak Classifiers, based on this sorter system, can realize the people's face Detection of Existence and the location of the image gathered at the present invention.

1.2.2, facial image detect and location algorithm

1.2.2.1, to pending facial image, according to each pixel in the sequential scanning image of Row Column,, calculate the pairing Haar eigenwert of current pixel according to the Haar feature of using in 1.2.1.4 step;

1.2.2.2, the eigenwert to obtaining in 1.2.2.1 step use 10 strong classifiers to differentiate successively, if the differentiation result of certain strong classifier is 0, illustrate that the current Haar characteristic area of selecting for use of this pixel is not a human face region; If the differentiation result of 10 strong classifiers is 1 all, illustrate that the current Haar characteristic area of selecting for use of this pixel is a human face region;

1.2.2.3, all pixels in the image have all been carried out going on foot described differentiation process as 1.2.2.2 after, can obtain that all are differentiated the pixel region for people's face zone in the image, these pixel regions are communicated with, can obtain final human face region.

3, embedded infrared face image-recognizing method according to claim 1 and 2, it is as follows to it is characterized in that adopting in the 2.1st step the principal component analysis method to carry out the algorithm steps of feature dimensionality reduction:

2.1.1,500 dimensional vectors are transformed to the matrix of 500*500, computation of mean values is obtained covariance;

2.1.2, ask covariance characteristic of correspondence value and characteristic of correspondence vector;

2.1.3, eigenwert is sorted from big to small, choose the eigenwert of final employing according to the accumulation contribution rate;

2.1.4, the selected eigenwert characteristic of correspondence vector of selection constitute transformation matrix;

2.1.5, for the data of new typing, directly multiply each other with this transformation matrix, obtain the data behind the dimensionality reduction of correspondence;

Wherein, accumulation contribution rate η _mComputing formula as follows: η _m=λ ₁+ λ ₂+ ... + λ _m/ (λ ₁+ λ ₂+ ... + λ _p)

λ in the formula is a proper vector characteristic of correspondence value, and m is the number that is chosen for the proper vector of major component, and p is the sum of proper vector; It is 99% that the present invention selects the contribution rate of accumulative total threshold value, and major component proper vector number is 100.

4, embedded infrared face image-recognizing method according to claim 1 is characterized in that the features training algorithm steps that adopts in the 2.2nd step is as follows:

2.2.1, make embedded infrared face recognition methods of the present invention can differentiate A people's facial image,, gather lineup's face image pattern, make it add up to P as undergoing training sample to wherein everyone; Gather one group of other people face image pattern that does not belong to this A people simultaneously as the refusal training sample, make it add up to Q;

2.2.2, adopt the K Mean Method to set up the cluster system, with A people's facial image poly-be the K class, K=A among the present invention;

2.2.3, adopt the iteration training method to calculate the classification thresholds S of each class, this threshold value can realize undergoing training correct receptance of sample is 99%, and the correct rejection ratio of refusing training sample is 99.9%; This threshold value will be used for the recognition of face in the 2.3rd step.

5, a kind of embedded infrared face pattern recognition device is characterized in that this recognition device comprises:

Core processor assembly: comprise based on the S3c2440 primary processor of ARM920T kernel with based on the TMS320DM642 coprocessor of TMS320C6000DSP platform;

Wherein coprocessor TMS320DM642 is responsible for receiving and handling the video image of input, obtains the infrared face view data, and sends data to primary processor S3c2440, is finished the location and the discriminator of infrared face image by s3c2440;

Infrared image acquisition assembly: comprise infrared thermal imaging device and coding and decoding video chip; The infrared thermal imaging device is used for captured video image, and output simulating signal, the coding and decoding video chip is a digital signal with the analog signal conversion of infrared thermal imaging device output, and output format is that the digital video signal of YUV4:2:2 of ITU-R BT.656 is to the TMS320DM642 coprocessor;

Memory module: use the SDRAM chip of two 32MB16 positions to be connected, be used for the primary processor data storage, use the NandFlash chip of a slice 64MB to be connected, be used for the procedure stores of primary processor with primary processor S3c2440 with primary processor S3c2440; Use the SDRAM chip of a slice 32MB16 position to be connected, be used for the coprocessor data storage, use the NandFlash chip of a slice 32MB to be connected, be used for the procedure stores of coprocessor with coprocessor TMS320DM642 with coprocessor TMS320DM642;

USB output interface assembly: the USB HOST controller that uses primary processor S3c2440 to carry is realized the USB output interface.

6, embedded infrared face pattern recognition device according to claim 5 is characterized in that this recognition device also comprises:

Power supply module: provide by 220V and exchange the direct power supply of the transformer that changes the 5V direct current and two kinds of patterns of powered battery, be used to drive the operation of hardware unit;

Miniature keyboard input module: be connected with primary processor, input password and steering order for the user;

LCD display unit: be connected with primary processor, be used to show facial image, identification comparison result;

Audio frequency output prompting assembly: be connected with primary processor, be used to provide voice suggestion, consumer-oriented operating process.

Images