CN101089874A - Identify recognising method for remote human face image - Google Patents

Identify recognising method for remote human face image Download PDF

Info

Publication number
CN101089874A
CN101089874A CN 200610087035 CN200610087035A CN101089874A CN 101089874 A CN101089874 A CN 101089874A CN 200610087035 CN200610087035 CN 200610087035 CN 200610087035 A CN200610087035 A CN 200610087035A CN 101089874 A CN101089874 A CN 101089874A
Authority
CN
China
Prior art keywords
image
gabor
facial image
histogram
standard faces
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN 200610087035
Other languages
Chinese (zh)
Other versions
CN101089874B (en
Inventor
邵刚
庄镇泉
庄连生
李斌
王睿斌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huawei Technologies Co Ltd
Original Assignee
Huawei Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huawei Technologies Co Ltd filed Critical Huawei Technologies Co Ltd
Priority to CN2006100870359A priority Critical patent/CN101089874B/en
Publication of CN101089874A publication Critical patent/CN101089874A/en
Application granted granted Critical
Publication of CN101089874B publication Critical patent/CN101089874B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Image Analysis (AREA)

Abstract

A method of using remote human face image (HFI) to identify status includes carrying out treatment on standard HFI by server end to obtain standard human face histogram character, forming weak sorter as per said histogram character, integrating optimum weak sorters to form strong sorter, using histogram corresponding to strong sorter as standard HFI optimum histogram character, making treatment on HFI to be identified to obtain optimum histogram character of HFI to be identified, comparing two said histogram characters to generate identification sample and confirming user status of HFI to be identified by classifying sample to be identified.

Description

A kind of personal identification method of remote human face image
Technical field
The present invention relates to technology, particularly a kind of personal identification method of remote human face image in the identification of communication system medium-long range.
Background technology
Along with the fast development of communication network, various services are also increasing, how to carry out remote identity identification and seem and become more and more important that remote identity identification has become prerequisite and the important component part that the Internet Service Provider provides various services in communication network.The mode of text password is adopted in traditional remote identity identification, be that terminal authenticates the certificate server that the text password sends to communication network, this mode not only requires the user to remember various loaded down with trivial details text passwords, and the text password is also lost easily with stolen.
In order to make the remote identity identifying efficient, novel and convenient, can utilize intrinsic physiological characteristic of human body or behavioural characteristic to carry out long-range identification at present, i.e. remote identity identification can be adopted biometrics identification technology.In numerous biometrics identification technologies, face recognition technology occupies an important position and obtained application in some special-purpose field.Recognition of face is to carry out remote identity identification according to facial image, comprises the detection of people's face and the identification of facial image.
At application number is CN200310120624.9, denomination of invention is in the patented claim of " people's face of multiple step format detects and recognition methods under the mobile computing environment ", disclosing people's face detects and the method for discerning, this method comprises: at first, detect and demarcate, camera from mobile device obtains image, and image is carried out light rectification simply and effectively, and adopts the fast face detection algorithm to carry out people's face and detect and demarcate; Secondly, encrypted transmission carries out the people's face scope that calibrates to send to server end by cordless communication network after digital watermarking is encrypted, and server end is verified the digital watermarking that embeds in people's face scope, judges the integrality and the correctness of facial image; At last, adopt recognition of face training algorithm to carry out facial image identification and authentication result is returned to mobile device based on built-in type hidden Markov model (HMM).
Present result of study shows, the facial image otherness of same individual under different illumination conditions has greatly may be also bigger than the otherness of the facial image of different people.Change under the little situation at indoor illumination condition, best's face recognition system discrimination can reach 95%, but under outdoor illumination condition, best's face identification discrimination can suddenly drop to about 50%.Therefore, illumination condition has become a key factor that influences remote human face image authentication discrimination.Because the complicacy of mobile device use occasion, the illumination condition of photographic images is inevitable complicated, above-mentioned this facial image authentication mode only carries out simple light to the image of having taken and corrects, do not use special illumination pretreatment algorithm that image is carried out pre-service, can cause the success ratio of remote identification facial image to reduce greatly like this.
In order not to be subjected to the influence of illumination condition when the remote validation facial image, improve the success ratio of identification facial image, can adopt from quotient images (SQI, Self Quotient Image) facial image is carried out being transferred to server end again behind the unitary of illumination, its concrete steps are:
(1) at first, a given facial image I, Gauss's smoothing operator G of the anisotropic of n different scale of selection 1, G 2..., G n, and give different weights W 1, W2 ..., Wn smoothly obtains a series of images after level and smooth with these operators to image I I ‾ i = I ⊕ 1 N W i G i , i = 1,2 , . . . , n ;
(2) calculate from quotient images Q i, its computing formula is: Q i = I I ‾ i , i = 1,2 , . . . , n ;
(3) utilize nonlinear function T to adjust value, make that the value of Qi drops between [0,255], remember that adjusted image is D from quotient images i: D i=T (Q i), i=1,2 ..., n;
(4) adjusted image is sued for peace, obtains final from quotient images Q: Q = Σ i = 1 n m i D i , I=1,2 ..., n, wherein, m iBe the weight from quotient images that the wave filter corresponding to different scale calculates, value can be 1;
(5) utilize the result after quotient images Q is as original image I illumination pretreatment to carry out subsequent treatment,, will send to server end again through the facial image of subsequent treatment and adopt the recognition of face training algorithm of HMM to discern such as feature extraction and identification etc.
But, adopt SQI that facial image is carried out being transferred to server end again behind the unitary of illumination and also exist shortcoming: the first, the parameter that SQI relates to is numerous, and particularly Gauss's smoothing operator parameter is selected difficulty in this method; The second, SQI is not fine to the place to go result of shade in the facial image; The 3rd, SQI needs further to improve to the precision of feature extraction in the facial image.
In order not to be subjected to the influence of illumination condition when the remote validation facial image, improve the success ratio of identification facial image, can also adopt following method: at first, set up standard faces image and strong classifier at server end, promptly utilize the Gabor wave filter standard faces image to be carried out the Gabor filtering of 8 directions of 5 yardsticks, utilize the filtered magnitude image of Gabor to describe the feature and the storage of image, with the amplitude after the Gabor conversion of single pixel in the magnitude image is that fundamental construction goes out Weak Classifier, it is comprehensive to carry out Weak Classifier by AdaBoost, and then constructs strong classifier; Then, the image that mobile device is photographed carries out obtaining facial image to be identified behind the geometrical normalization, sends to server end, facial image to be identified is authenticated according to the standard faces characteristics of image of storage and the strong classifier of facial image by server.The concrete steps of this method are:
(1) at server end, the standard faces image is carried out the Gabor filtering of 8 directions of 5 yardsticks, the Gabor wave filter is expressed as follows: ψ u , v ( z ) = | | k u , v | | 2 σ 2 e ( - | | k u , v | | 2 | | z | | 2 / 2 σ 2 ) [ e i k u , v z - e - σ 2 / 2 ] , Wherein, k u , v = k v e i φ u ; k v = k max f v Specified frequency, φ u = uπ 8 , φ u∈ [0, π) specified direction, and z=(x, y).
In above-mentioned formula, ν has controlled the yardstick of Gabor wave filter, is determining the center in frequency field of Gabor wave filter; U is controlling the filtering direction of Gabor wave filter.In experiment, the value of parameter is: ν ∈ 0,1,2,3,4}, u ∈ 0,1,2,3,4,5,6,7} and σ=2 π, k Max=pi/2, f=.After Gabor filtering, a standard faces image becomes 40 onesize images to be represented, is called Gaborface, and Gaborface is stored, as shown in Figure 1: the left side is not for passing through the standard faces image of Gabor filtering, and the right is the standard faces image through Gabor filtering;
(2) at server end, all standard faces images are carried out Gabor filtering after, produce the positive negative sample of standard faces, i.e. training sample.Wherein, positive sample is represented the difference between same people's the filtered Gaborface of various criterion facial image, and negative sample is represented the difference between the Gaborface behind the standard faces image filtering of different people.As shown in Figure 2, the positive sample shown in the top in the drawings for obtaining behind same individual's the different people face image filtering; Negative sample shown in the below in the drawings for obtaining after the facial image filtering of different people;
(3),, be a Weak Classifier with each pixel of training sample according to the training sample that obtains at server end, pick out optimum Weak Classifier combination with the AdaBoost algorithm, constitute strong classifier, the framework of training process is finally trained and is obtained a strong classifier as shown in Figure 3;
(4) at server end, utilize the Gaborface and the strong classifier of standard faces image to discern, the steps include: for one on mobile device through the facial image to be certified of geometrical normalization, at first this facial image is carried out Gabor filtering, Gaborface according to the standard faces image of the Gaborface of this facial image and server end storage compares one by one then, the difference that obtains both generates a sample to be certified, at the strong classifier that obtains according to the training stage sample to be certified is classified:, so just illustrate that people's face to be certified and this standard faces image belong to same individual's if this sample belongs to positive sample; Otherwise be different people; Thereby finally determine the identity of facial image to be certified.
Adopt said method to carry out the remote human face image authentication and also exist shortcoming: first, only utilize the filtered magnitude image of Gabor to describe the feature of image, and do not utilize the filtered phase image of Gabor, and in practice, phase image contains the more images textural characteristics than magnitude image, to illumination variation robust more; The second, directly with the amplitude of the pixel of magnitude image as feature, this feature for classifying quality not as the statistical nature based on the zone come robust; The 3rd, utilize the AdaBoost algorithm to carry out the comprehensive of Weak Classifier, what might obtain is not optimum strong classifier, this is because on essence, the AdaBoost algorithm is equivalent to the gradient descent method of function, be absorbed in local optimum probably, therefore the strong classifier that obtains might not be a global optimum, thereby has influenced the discrimination of remote human face image.
Summary of the invention
In view of this, fundamental purpose of the present invention is to provide a kind of personal identification method of remote human face image, and this method can not be subjected to the influence of illumination condition when the remote identification facial image, improves the discrimination of facial image.
According to above-mentioned purpose, technical scheme of the present invention is achieved in that
A kind of personal identification method of remote human face image, this method comprises:
In the training stage of standard faces image, server end is to standard faces image process illumination pretreatment, Gabor filtering and the normalization of Gabor coefficient, behind subwindow analysis and the Gabor statistics with histogram, obtain the Gabor histogram feature of standard faces image, the standard faces image is made up in twos, Gabor histogram feature according to the standard faces image makes up Weak Classifier, utilize optimization algorithm screening Weak Classifier to constitute strong classifier, the Gabor histogram feature of the pairing standard faces image of strong classifier is the optimum Gabor histogram feature of standard faces image;
In the identification stage of remote human face image, server end carries out illumination pretreatment to the facial image to be identified that receives from client, Gabor filtering and the normalization of Gabor coefficient, behind subwindow analysis and the Gabor statistics with histogram, extract the optimum Gabor histogram feature of facial image to be identified, one by one comparing with the optimum Gabor histogram feature standard faces image produces sample to be identified, the strong classifier that obtains according to the training stage is classified to the sample identified again, determine the user identity of facial image to be identified, recognition result is sent to client.
The process of described illumination pretreatment is:
A, image I is carried out log-transformation, obtain the facial image i=LogI after the conversion;
B, the image i after the conversion is estimated l by the full variation model of finding the solution expansion: min l ≥ s ∫ Ω [ | ▿ l | 2 + α ( l - i ) 2 + β | ▿ ( l - i ) | 2 ] dxdy , This formula adopts optimization algorithm to find the solution;
C, obtain after the l, l is carried out transformation transformation, obtain image L=exp (l);
D, image L is carried out nonlinear transformation, make it value and fall into [0,255];
E, with the merchant of image I and image L as illumination invariant R, promptly R = I L ;
F, illumination invariant R is carried out nonlinear transformation, the value that makes pixel has obtained the image after the illumination pretreatment between [0,255].
The described optimization algorithm of step b is multiple step format algorithm for estimating EDA.
The normalized process of described Gabor coefficient is: the Gabor coefficient that facial image is carried out each pixel of obtaining after the Gabor filtering carries out normalized, makes through the amplitude of the facial image of Gabor filtering and the value discretize of phase place.
The process of described subwindow analysis and Gabor statistics with histogram is: extract facial image at the statistic histogram of the corresponding Gabor coefficient in the subwindow zone Gabor histogram feature as facial image.
The facial image Gabor histogram feature of described optimum utilizes the EDA screening to obtain.
From such scheme as can be seen, method provided by the invention is in carrying out the training stage process of standard faces image, extract the feature of standard faces image and will pass through illumination pretreatment, Gabor filtering and the normalization of Gabor coefficient (amplitude and phase place value discretize), after filtered image being carried out the Gabor coefficient histogram of subwindow analysis and statistics subwindow zone correspondence, obtain the Gabor histogram feature of standard faces image, the standard faces image is matched in twos, constituted training sample according to the difference between the Gabor histogram feature basis of calculation facial image of standard faces image, difference between each Gabor histogram of image is then as the feature of training sample, each feature of training sample as a Weak Classifier, filter out optimum Weak Classifier combination by EDA and obtain strong classifier, and the Gabor histogram feature of the standard faces image of these strong classifier correspondences is exactly the feature that ultimate demand is extracted and participation is classified of every standard faces image, is referred to as the optimum Gabor histogram feature of standard faces image.In carrying out remote human face image identification phase process, to carry out after illumination pretreatment, the Gabor filtering and Gabor coefficient normalization (amplitude and phase place value discretize) to facial image to be certified accordingly, extract the optimum Gabor histogram feature of facial image to be identified, one by one comparing with the optimum Gabor histogram feature standard faces image obtains sample to be identified, again the strong classifier that the sample evidence to be identified training stage the is obtained identification result that classifies and obtain facial image to be identified.Because the present invention has carried out special illumination pretreatment to facial image, so be not subjected to the influence of illumination condition; Because the present invention adopts the feature of the Gabor histogram of subwindow as facial image, so can improve the discrimination and the robustness of face recognition algorithms.
Description of drawings
Fig. 1 is the synoptic diagram of standard faces image before and after Gabor filtering in the prior art;
The synoptic diagram of the negative sample that obtains after the positive sample that obtains behind the different people face image filtering of Fig. 2 in the prior art same individual and the facial image filtering of different people;
Fig. 3 is for training the framework synoptic diagram of the strong classifier that obtains facial image in the prior art;
Fig. 4 carries out the method flow diagram of remote identity identification to facial image for the present invention;
Fig. 5 is the method flow diagram of the training stage of standard faces image of the present invention;
Fig. 6 extracts each feature synoptic diagram through the facial image of illumination pretreatment for the present invention;
Fig. 7 carries out the synoptic diagram of normalized for the present invention to each that extract through the feature of the facial image of illumination pretreatment;
Fig. 8 is the method flow diagram of server end identification facial image of the present invention.
Embodiment
In order to make the purpose, technical solutions and advantages of the present invention clearer, below lift specific embodiment and, the present invention is described in more detail with reference to accompanying drawing.
The invention provides the method for carrying out identification according to facial image under a kind of variable illumination condition, emphasis solves the identity information of discerning captured facial image under the complex illumination condition.Method provided by the invention can be applied in the communication network, in cordless communication network.When carrying out the identification of facial image, at first, can behind mobile device collection facial image, send to server end by client by communication network; Secondly, server end is handled the facial image that receives and is extracted corresponding feature, and classifies according to the strong classifier that training obtains, and recognition result is sent to client by communication network.
Below send to server end with client facial image to be identified serve as that captured facial image after handling through geometrical normalization is illustrated.
Fig. 4 comprises the training stage of standard faces image and the remote identity cognitive phase of facial image for the present invention carries out the method flow diagram that remote identity is discerned to facial image, and its concrete steps are:
The training stage of standard faces image
After step 400, server end carry out illumination pretreatment respectively to the standard faces image, carry out the Gabor filtering of 8 directions of 5 yardsticks respectively, each standard faces image of opening through illumination pretreatment obtains 40 Gaborface, the pixel point value of this Gaborface is a plural number, and the value of its amplitude and phase place is a real number.
Step 401, server end carry out the discretize processing to the amplitude of each pixel among the Gaborface that obtains and the value of phase place, make it to normalize to the integer between [0,255].
Step 402, server end scan Gaborface with the subwindow of variable size, add up each corresponding Gabor histogram (comprising amplitude histogram and phase place histogram) in subwindow zone, a feature as facial image, be called the Gabor histogram feature, all Gabor histogram features lump together as the feature of standard faces image and storage.
Step 403, server end makes up in twos to all standard faces images, calculate difference between the standard faces image according to the Gabor histogram feature of every standard faces image, this difference is as training sample, difference between the Gabor histogram feature of standard faces image correspondence has constituted the feature of training sample, each feature of training sample as a Weak Classifier, adopt distributed algorithm for estimating (EDA) to filter out optimum Weak Classifier and constitute strong classifier, and the Gabor histogram feature of these strong classifier correspondences is to the most effective feature of standard faces image identification, be called optimum Gabor histogram feature, only need extract the optimum Gabor histogram feature of standard faces image at the cognitive phase facial image.
The remote identity cognitive phase of facial image
Step 404, client send to server end with facial image to be identified through after the processing of geometrical normalization.
Step 405, server end carry out illumination pretreatment, carry out the Gabor filtering of 8 directions of 5 yardsticks the facial image to be identified that receives, and obtain 40 Gaborface, and the picture element of this Gaborface is a plural number, and the value of its phase place and amplitude is a real number.
Step 406, server end carry out discretize to the amplitude of each pixel among the Gaborface that obtains and the value of phase place, make it value and normalize to integer between [0,255], extract the optimum Gabor histogram feature of facial image to be identified.
Step 407, server end compare facial image to be identified seriatim with the standard faces image of all storages, according to the optimum Gabor histogram feature of facial image to be identified calculate and the optimum Gabor histogram feature of standard faces image between difference, obtain sample to be identified, the strong classifier that obtains according to the training stage is classified to the sample identified again: if sample to be identified is positive sample, facial image then to be identified and standard faces image belong to same individual; If sample to be identified is a negative sample, facial image then to be identified belongs to different people with the standard faces image.
Step 408, server end send to client to recognition result.
Server end of the present invention need be trained the standard faces image that has got access in order to carry out identification to the facial image that receives from client, below concrete narration how the standard faces image is trained.
The training stage of standard faces image
Because the facial image identification problem is the multiclass problem, the present invention for convenience is converted into two class problems, promptly positive sample and negative sample.Wherein, positive sample is the difference between same individual's the different facial images; Negative sample is the difference between the facial image of different people.Positive sample and negative sample are commonly referred to as training sample here.
The purpose of the training stage of standard faces image is to pick out from the standard faces image for the most effective characteristics combination of facial image identification, utilize these the most effective characteristics combination to constitute strong classifier, utilize the feature of standard faces image and strong classifier that facial image to be identified is carried out identification at the cognitive phase of server end.
In order can normally to train the standard faces image, same individual comprises two different facial images at least, and Fig. 5 is the method flow diagram of the training stage of standard faces image of the present invention, and its concrete steps are:
Step 500, server end carry out illumination pretreatment respectively to all standard faces images, extract the illumination invariant of facial image respectively.
Step 501, server end carry out the Gabor filtering of 8 directions of 5 yardsticks respectively to the facial image through illumination pretreatment, and each facial image through illumination pretreatment all obtains 40 filtered facial images, i.e. Gaborface respectively.Every each pixel point value of Gaborface is a plural number, and the value of its amplitude and phase place is a real number.As shown in Figure 6, Fig. 6 is for carrying out obtaining after the Gabor filtering synoptic diagram of 40 Gaborface to each through facial image of illumination pretreatment.
Step 502, server end carry out discretize respectively to each pixel value of every Gaborface (comprising amplitude and phase place) to be handled, and makes it to normalize to the integer between [0,255].
Because each value through the pixel of the filtered Gaborface as a result of Gabor of the facial image of illumination pretreatment is a plural number, the value of its amplitude and phase place is a real number, for amplitude histogram and the phase place histogram (being referred to as the Gabor histogram) of adding up the subwindow zone, the present invention need carry out discretize respectively to the value of the amplitude of each pixel of the Gaborface that obtains after the filtering and phase place and handle, make it to normalize to the integer between [0,255].
Be illustrated: amplitude normalization concrete grammar is that more current unique point P is adjacent the amplitude size of 8 unique points, constructs one 8 bit (b 1b 2b 3b 4b 5b 6b 7b 8) 2If the amplitude of i unique point is bigger than the amplitude of current unique point P, b so i=1, otherwise b i=0, at last the amplitude of current unique point P (b 1b 2b 3b 4b 5b 6b 7b 8) 2Replace, thereby realize the amplitude of type real is normalized to integer between [0,255], as shown in Figure 7.
Step 503, server end scan the Gaborface that obtains respectively by the subwindow of variable size, add up the Gabor histogram (comprising amplitude histogram and phase place histogram) in each subwindow, all subwindow Gabor histograms lump together the Gabor histogram feature that has just constituted facial image.
In the present invention, extracting each process through the Gabor histogram feature of the facial image of illumination pretreatment is:
Every Gaborface of subwindow scanning by a series of variable sizes extracts the Gabor histogram (comprising amplitude histogram and phase place histogram) in corresponding subwindow zone, as the feature in this subwindow zone in every Gaborface.Like this, each is through each subwindow zone of the facial image of illumination pretreatment, and its feature is described by 80 Gabor histograms, comprises 40 amplitude histograms and 40 phase place histograms.In the present invention, suppose that every facial image size is 100*100, if the size of subwindow changes to 100*100 by 10*10, the step-length that the length of subwindow and width always equate and length and width change is 2, about each subwindow, moving step length up and down is 3, so the subwindow number that obtains of the final scanning of every facial image is 497025, each subwindow is described with 80 Gabor histograms, and so final every facial image through illumination pretreatment is described its feature by 497025*80=39762000 Gabor histogram exactly.In the present invention, the information of corresponding three aspects of each Gabor histogram: the direction of subwindow position, filtering and yardstick and Gabor histogram type, wherein, Gabor histogram type can be amplitude histogram and phase place histogram.Each Gabor histogram is with a column vector H i, i=1,2 ..., N represents, each all Gabor histogram through the facial image of illumination pretreatment is arranged in sequence constituted a matrix H=[H together 1H 2... H N], matrix H is exactly each feature through the facial image of illumination pretreatment.
Step 504, server end make up all standard faces images in twos, calculate difference between the standard faces image according to the Gabor histogram feature of standard faces image, and as training sample, the difference between the Gabor histogram of standard faces image correspondence has then constituted the feature of training sample.
In the present invention, the difference between same individual's the various criterion facial image has formed positive sample, and the difference between the standard faces image of different people has formed negative sample.By means of the notion of positive negative sample, can be converted to two class problems to the multiclass problem in the facial image identification, judge that promptly the difference between any two facial images is to belong to positive sample class, i.e. same individual; Still belong to the negative sample class, i.e. different people.
The production method of training sample, promptly the otherness computing method between two facial images are: for any a pair of facial image I mAnd I n, its eigenmatrix is the H through normalized mAnd H n, the difference between the facial image is with a N dimensional vector D Mn=[d 1, d 2..., d N] represent that N equals the texture histogram number of facial image, wherein d iThe representation feature matrix H mWith H nI column vector between similarity, i.e. similarity between Dui Ying two Gabor histograms.Therefore, training sample D MnBe defined in the space of a N dimension, the value of all similarities has constituted whole feature space.
Step 505, server end as a Weak Classifier, adopt each feature of training sample EDA filter out optimum Weak Classifier and constitute strong classifier, obtain the Gabor histogram feature of optimum facial image.
In the present invention, because training sample is a N dimensional vector, N is often bigger, therefore need carry out dimensionality reduction to the feature space of training sample, picks out optimal characteristics, and the intrinsic dimensionality that reduces training sample improves classifying quality.
Adopt EDA to filter out the most effective characteristics combination and constitute strong classifier, its concrete steps are:
With length is the sign indicating number string c of N 1c 2... c N(c wherein i=0 or 1) characteristics combination that expresses possibility, all possible characteristics combination has constituted the search volume of EDA, and each yard string is called the body one by one in the EDA search volume.Each c in the sign indicating number string iCorresponding to a specific eigenwert component d ic iThe selected participation classification of=1 expression characteristic of correspondence component, c i=0 expression characteristic of correspondence component does not participate in classification, represents to have only first characteristic component and the 3rd characteristic component to participate in classification as 10100...000, and final strong classifier is made of two characteristic components, i.e. first and the 3rd characteristic component.
The present invention utilizes this strong classifier to align negative sample to classify, calculate the discrimination of this strong classifier, and this discrimination is defined as individual fitness, and promptly fitness function is defined as individual discrimination.In the space, search out optimum individual with EDA, i.e. the characteristics combination that discrimination is the highest and number of features is minimum, the characteristic component of optimum individual representative has constituted final strong classifier.
Suppose that having obtained one at last through EDA screening forms strong classifier by the K characteristic component, K＜＜N, so finally each training sample can be described by this K characteristic component, is designated as D mn / = [ d 1 ′ , d 2 ′ , . . . , d K ′ ] . D wherein mThe new proper vector of n ' expression training sample, d j' (j=1,2 ..., K) corresponding to the selected optimal characteristics component that comes out.Each characteristic component is all corresponding to a texture histogram, and the final feature of extracting of standard faces image is exactly the texture histogram of these characteristic component correspondences, remembers that new image characteristic matrix is H '=[H 1' H 2' ... H k'], H wherein i' corresponding to the selected characteristic component d that comes out i' pairing Gabor histogram vectors, these select Gabor histograms are exactly optimum facial image Gabor histogram feature, utilize these optimum Gabor histogram features to represent facial image and carry out identification.
In server stores after the optimum Gabor histogram feature of standard faces image and training stage obtain strong classifier, just can discern, below the remote identity cognitive phase of facial image has been elaborated the facial image to be identified that client sends.
The remote identity cognitive phase of facial image
The remote identity cognitive phase of facial image is the feature according to facial image to be identified, determines user identity.At server end, the same with the training stage of standard faces image, the present invention is converted to two class problems to the multiclass problem of facial image identification, belongs to same individual by which standard faces image of judging facial image to be identified and storage and determines user identity.
Fig. 8 is the method flow diagram of server end identification facial image of the present invention, and its concrete steps are:
Step 800, server end carry out illumination pretreatment to the facial image to be identified that client sends over, and extract the illumination invariant of facial image to be identified.
Step 801, server end carry out the Gabor filtering of 8 directions of 5 yardsticks to the facial image to be identified through illumination pretreatment, obtain 40 filtered facial images to be identified, Gaborface promptly to be identified.
In the present invention, the pixel value of 40 Gaborface of facial image to be identified also is a plural number, and its amplitude and phase place value are real number.
The amplitude of each pixel among each Gaborface that step 802, server end obtain after to filtering and the value of phase place are carried out discretize and are handled, and make it value and normalize to integer between [0,255].
Step 803, server end scan 40 Gaborface of the facial image to be identified that obtains respectively by the subwindow of variable size, extract the optimum Gabor histogram feature of facial image to be identified.
In the present invention, extract the optimum Gabor histogram feature of facial image to be identified, these feature constitutive characteristic matrix H=[H 1H 2... H K], each Gabor histogram feature represents that with a column vector this column vector is exactly H i
Step 804, server end mate facial image to be identified the subscriber identity information of final decision facial image to be identified one by one with the standard faces image of being preserved.
The step of mating is: calculate the difference between the optimum Gabor histogram feature of the optimum Gabor histogram feature of facial image to be identified and standard faces image, obtain sample to be identified, the strong classifier that utilizes the training stage the to obtain sample to be identified of classifying: if sample to be identified is positive sample, then the match is successful, and facial image to be identified and standard faces image belong to same user; If sample to be identified is a negative sample, then it fails to match, and facial image to be identified and standard faces image belong to different user.
Step 805, server end send to client with recognition result.
In Fig. 5 or Fig. 8, the detailed process of illumination pretreatment is as described below.
Illumination is a key factor that influences facial image identification, and the variation of illumination condition can cause the fierceness of facial image gray-scale value to change, and the influence of removing illumination condition can realize by the illumination invariant that extracts facial image.In the present invention, the illumination condition problem is converted to an optimization problem, utilizes EDA to realize illumination pretreatment, extract the illumination invariant of facial image.
If a facial image is I, the algorithm of its illumination pretreatment is:
1) facial image I is carried out log-transformation, obtain the facial image i after the conversion, i.e. i=LogI;
2) the facial image i after the conversion is estimated l by the full variation model of finding the solution expansion: min l ≥ s ∫ Ω [ | ▿ l | 2 + α ( l - i ) 2 + β | ▿ ( l - i ) | 2 ] dxdy , Finding the solution of this formula is actually an optimization problem, and method for solving can be varied, selects for use in the present invention to optimize effective and find the solution fireballing EDA and find the solution;
3) obtain after the l, l is carried out transformation transformation, obtain facial image L, be i.e. L=exp (l);
4) facial image L is carried out nonlinear transformation, make it value and fall into [0,255];
5) utilize the merchant of facial image I and facial image L as illumination invariant R, promptly R = I L ;
6) illumination invariant R is carried out nonlinear transformation, the value that makes pixel has obtained the facial image after the illumination pretreatment between [0,255].
Server end carries out according to the method described above to the illumination pretreatment process of all standard faces images.
Method provided by the invention has been carried out the experimental verification of principle, the face database that is adopted is the FERET storehouse, and training set is 270 people totally 540 fronts (accurate positive) facial images, and test set comprises two subclass: fa subclass 1196 people, totally 1196, as the standard faces image; Fb subclass 1195 people, totally 1195, but identical with fa subclass personnel express one's feelings different.The algorithm training time was approximately for 1 week, and average each facial image recognition time is all less than 1.5 seconds, discrimination＞98%.
The method of remote human face image identification provided by the invention, imagery exploitation is expanded full variation model carry out illumination pretreatment, realize the unitary of illumination of image, help to improve the robustness and the discrimination of image recognition algorithm, make the recognition system that adopts this method can be applied in the occasion of illumination condition complexity; The method of remote human face image identification provided by the invention, when the training strong classifier, adopt the EDA algorithm to carry out the feature screening, help to reduce the characteristic number of standard faces image, when recognition image, can reduce processing time and characteristic storage space, improve recognition effect, in addition, because global optimum's search capability that EDA self has makes that the strong classifier that finally obtains is a global optimum; The Gabor histogram feature that the method for remote human face image identification provided by the invention proposes is more stable based on the feature of single pixel than prior art, can improve the robustness of the recognition system that adopts this method, this be because based on the Feature Recognition algorithm of single pixel when the identification, facial image is alignment accurately, otherwise recognition effect can be subjected to very big influence, and Gabor histogram feature of the present invention has utilized regional statistical property, when identification, positioning requirements to unique point is looser, the feature of extracting is relatively stable, and the recognition result of recognizer is also relatively stable.
The above only is preferred embodiment of the present invention, not in order to restriction the present invention, all any modifications of being made within the spirit and principles in the present invention, is equal to and replaces and improvement etc., all should be included within protection scope of the present invention.

Claims (6)

1, a kind of personal identification method of remote human face image is characterized in that, this method comprises:
In the training stage of standard faces image, server end is to standard faces image process illumination pretreatment, Gabor filtering and the normalization of Gabor coefficient, behind subwindow analysis and the Gabor statistics with histogram, obtain the Gabor histogram feature of standard faces image, the standard faces image is made up in twos, Gabor histogram feature according to the standard faces image makes up Weak Classifier, utilize optimization algorithm screening Weak Classifier to constitute strong classifier, the Gabor histogram feature of the pairing standard faces image of strong classifier is the optimum Gabor histogram feature of standard faces image;
In the identification stage of remote human face image, server end carries out illumination pretreatment to the facial image to be identified that receives from client, Gabor filtering and the normalization of Gabor coefficient, behind subwindow analysis and the Gabor statistics with histogram, extract the optimum Gabor histogram feature of facial image to be identified, one by one comparing with the optimum Gabor histogram feature standard faces image produces sample to be identified, the strong classifier that obtains according to the training stage is classified to the sample identified again, determine the user identity of facial image to be identified, recognition result is sent to client.
2, the method for claim 1 is characterized in that, the process of described illumination pretreatment is:
A, image I is carried out log-transformation, obtain the facial image i=LogI after the conversion;
B, the image i after the conversion is estimated l by the full variation model of finding the solution expansion: min l ≥ s ∫ Ω [ | ▿ l | 2 + α ( l - i ) 2 + β | ▿ ( l - i ) | 2 ] dxdy , This formula adopts optimization algorithm to find the solution;
C, obtain after the l, l is carried out transformation transformation, obtain image L=exp (l);
D, image L is carried out nonlinear transformation, make it value and fall into [0,255];
E, with the merchant of image I and image L as illumination invariant R, promptly R = I L ;
F, illumination invariant R is carried out nonlinear transformation, the value that makes pixel has obtained the image after the illumination pretreatment between [0,255].
3, method as claimed in claim 1 or 2 is characterized in that, the described optimization algorithm of step b is multiple step format algorithm for estimating EDA.
4, the method for claim 1, it is characterized in that, the normalized process of described Gabor coefficient is: the Gabor coefficient that facial image is carried out each pixel of obtaining after the Gabor filtering carries out normalized, makes through the amplitude of the facial image of Gabor filtering and the value discretize of phase place.
5, the method for claim 1 is characterized in that, the process of described subwindow analysis and Gabor statistics with histogram is: extract facial image at the statistic histogram of the corresponding Gabor coefficient in the subwindow zone Gabor histogram feature as facial image.
6, the method for claim 1 is characterized in that, the facial image Gabor histogram feature of described optimum utilizes the EDA screening to obtain.
CN2006100870359A 2006-06-12 2006-06-12 Identify recognizing method for remote human face image Expired - Fee Related CN101089874B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN2006100870359A CN101089874B (en) 2006-06-12 2006-06-12 Identify recognizing method for remote human face image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN2006100870359A CN101089874B (en) 2006-06-12 2006-06-12 Identify recognizing method for remote human face image

Publications (2)

Publication Number Publication Date
CN101089874A true CN101089874A (en) 2007-12-19
CN101089874B CN101089874B (en) 2010-08-18

Family

ID=38943228

Family Applications (1)

Application Number Title Priority Date Filing Date
CN2006100870359A Expired - Fee Related CN101089874B (en) 2006-06-12 2006-06-12 Identify recognizing method for remote human face image

Country Status (1)

Country Link
CN (1) CN101089874B (en)

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011091593A1 (en) * 2010-01-29 2011-08-04 Nokia Corporation Methods and apparatuses for facilitating object recognition
CN101771539B (en) * 2008-12-30 2012-07-04 北京大学 Face recognition based method for authenticating identity
CN102714558A (en) * 2009-10-30 2012-10-03 三星电子株式会社 Image providing system and method
CN103020589A (en) * 2012-11-19 2013-04-03 山东神思电子技术股份有限公司 Face recognition method for single training sample
CN101604386B (en) * 2008-06-09 2013-06-12 株式会社电装 Image recognizing apparatus and method
WO2015103970A1 (en) * 2014-01-09 2015-07-16 Tencent Technology (Shenzhen) Company Limited Method, apparatus and system for authenticating user
CN105528616A (en) * 2015-12-02 2016-04-27 深圳Tcl新技术有限公司 Face recognition method and device
CN106156568A (en) * 2015-03-24 2016-11-23 联想(北京)有限公司 A kind of biometric information identification module and electronic equipment
CN106507199A (en) * 2016-12-20 2017-03-15 深圳Tcl数字技术有限公司 TV programme suggesting method and device
CN106934335A (en) * 2015-12-31 2017-07-07 深圳光启合众科技有限公司 The method and apparatus of image recognition
CN107256407A (en) * 2017-04-21 2017-10-17 深圳大学 A kind of Classification of hyperspectral remote sensing image method and device
CN109492601A (en) * 2018-11-21 2019-03-19 泰康保险集团股份有限公司 Face comparison method and device, computer-readable medium and electronic equipment
CN112149564A (en) * 2020-09-23 2020-12-29 上海交通大学烟台信息技术研究院 Face classification and recognition system based on small sample learning
CN112699355A (en) * 2020-12-22 2021-04-23 湖南麒麟信安科技股份有限公司 Dynamic face authentication method and system with user and host decoupled
CN114241534A (en) * 2021-12-01 2022-03-25 佛山市红狐物联网科技有限公司 Rapid matching method and system for full-palmar venation data
US11651447B2 (en) 2019-10-31 2023-05-16 Kyndryl, Inc. Ledger-based image distribution permission and obfuscation

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6301370B1 (en) * 1998-04-13 2001-10-09 Eyematic Interfaces, Inc. Face recognition from video images
US6826300B2 (en) * 2001-05-31 2004-11-30 George Mason University Feature based classification
US7369686B2 (en) * 2001-08-23 2008-05-06 Sony Corporation Robot apparatus, face recognition method, and face recognition apparatus
CN1204531C (en) * 2003-07-14 2005-06-01 中国科学院计算技术研究所 Human eye location method based on GaborEge model
CN1258737C (en) * 2003-10-29 2006-06-07 中国科学院计算技术研究所 Method of sicriminating handwriting by computer based on analyzing local feature

Cited By (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101604386B (en) * 2008-06-09 2013-06-12 株式会社电装 Image recognizing apparatus and method
CN101771539B (en) * 2008-12-30 2012-07-04 北京大学 Face recognition based method for authenticating identity
CN102714558A (en) * 2009-10-30 2012-10-03 三星电子株式会社 Image providing system and method
CN102741862A (en) * 2010-01-29 2012-10-17 诺基亚公司 Methods and apparatuses for facilitating object recognition
KR101479384B1 (en) * 2010-01-29 2015-01-05 노키아 코포레이션 Methods and apparatuses for facilitating object recognition
US9286544B2 (en) 2010-01-29 2016-03-15 Nokia Technologies Oy Methods and apparatuses for facilitating object recognition
WO2011091593A1 (en) * 2010-01-29 2011-08-04 Nokia Corporation Methods and apparatuses for facilitating object recognition
EP2529334A4 (en) * 2010-01-29 2017-07-19 Nokia Technologies Oy Methods and apparatuses for facilitating object recognition
CN103020589B (en) * 2012-11-19 2017-01-04 山东神思电子技术股份有限公司 A kind of single training image per person method
CN103020589A (en) * 2012-11-19 2013-04-03 山东神思电子技术股份有限公司 Face recognition method for single training sample
WO2015103970A1 (en) * 2014-01-09 2015-07-16 Tencent Technology (Shenzhen) Company Limited Method, apparatus and system for authenticating user
CN106156568B (en) * 2015-03-24 2020-03-24 联想(北京)有限公司 Biological information identification module and electronic equipment
CN106156568A (en) * 2015-03-24 2016-11-23 联想(北京)有限公司 A kind of biometric information identification module and electronic equipment
CN105528616A (en) * 2015-12-02 2016-04-27 深圳Tcl新技术有限公司 Face recognition method and device
CN105528616B (en) * 2015-12-02 2019-03-12 深圳Tcl新技术有限公司 Face identification method and device
CN106934335A (en) * 2015-12-31 2017-07-07 深圳光启合众科技有限公司 The method and apparatus of image recognition
CN106507199A (en) * 2016-12-20 2017-03-15 深圳Tcl数字技术有限公司 TV programme suggesting method and device
CN107256407A (en) * 2017-04-21 2017-10-17 深圳大学 A kind of Classification of hyperspectral remote sensing image method and device
CN107256407B (en) * 2017-04-21 2020-11-10 深圳大学 Hyperspectral remote sensing image classification method and device
CN109492601A (en) * 2018-11-21 2019-03-19 泰康保险集团股份有限公司 Face comparison method and device, computer-readable medium and electronic equipment
US11651447B2 (en) 2019-10-31 2023-05-16 Kyndryl, Inc. Ledger-based image distribution permission and obfuscation
CN112149564A (en) * 2020-09-23 2020-12-29 上海交通大学烟台信息技术研究院 Face classification and recognition system based on small sample learning
CN112149564B (en) * 2020-09-23 2023-01-10 上海交通大学烟台信息技术研究院 Face classification and recognition system based on small sample learning
CN112699355A (en) * 2020-12-22 2021-04-23 湖南麒麟信安科技股份有限公司 Dynamic face authentication method and system with user and host decoupled
CN114241534A (en) * 2021-12-01 2022-03-25 佛山市红狐物联网科技有限公司 Rapid matching method and system for full-palmar venation data
CN114241534B (en) * 2021-12-01 2022-10-18 佛山市红狐物联网科技有限公司 Rapid matching method and system for full-palm venation data

Also Published As

Publication number Publication date
CN101089874B (en) 2010-08-18

Similar Documents

Publication Publication Date Title
CN101089874B (en) Identify recognizing method for remote human face image
JP4543423B2 (en) Method and apparatus for automatic object recognition and collation
KR100734849B1 (en) Method for recognizing face and apparatus thereof
CN102902959B (en) Face recognition method and system for storing identification photo based on second-generation identity card
Messer et al. Face verification competition on the XM2VTS database
Froba et al. Robust face detection at video frame rate based on edge orientation features
US20050207614A1 (en) Iris-based biometric identification
WO2008072622A1 (en) Face authentication device
CN102938065A (en) Facial feature extraction method and face recognition method based on large-scale image data
CN107798308B (en) Face recognition method based on short video training method
CN111104852B (en) Face recognition technology based on heuristic Gaussian cloud transformation
Deshpande et al. CNNAI: a convolution neural network-based latent fingerprint matching using the combination of nearest neighbor arrangement indexing
Borra et al. An efficient fingerprint identification using neural network and BAT algorithm
JPH07302327A (en) Method and device for detecting image of object
Fritz et al. Building detection from mobile imagery using informative SIFT descriptors
Sahbi et al. Coarse to fine face detection based on skin color adaption
Kauba et al. Identifying the origin of iris images based on fusion of local image descriptors and PRNU based techniques
US20030044067A1 (en) Apparatus and methods for pattern recognition based on transform aggregation
Han et al. Object-based regions of interest for image compression
CN112418078B (en) Score modulation method, face recognition device and medium
CN111191519B (en) Living body detection method for user access of mobile power supply device
Xu et al. Coin recognition method based on SIFT algorithm
Triantafyllou et al. Iris authentication utilizing co-occurrence matrices and textile features
Taha et al. Speeded up robust features descriptor for iris recognition systems
Perez-Hernandez et al. Simplified stroke-based approach for off-line signature recognition

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C14 Grant of patent or utility model
GR01 Patent grant
CI01 Correction of invention patent gazette

Correction item: Claims

Correct: 6 items

False: 3 items

Number: 33

Volume: 26

CI03 Correction of invention patent

Correction item: Claims

Correct: 6 items

False: 3 items

Number: 33

Page: Description

Volume: 26

ERR Gazette correction

Free format text: CORRECT: CLAIM OF RIGHT; FROM: ITEM 3 TO: ITEM 6

CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20100818

Termination date: 20170612