CN107615298A - Face identification method and system - Google Patents
Face identification method and system Download PDFInfo
- Publication number
- CN107615298A CN107615298A CN201680030571.7A CN201680030571A CN107615298A CN 107615298 A CN107615298 A CN 107615298A CN 201680030571 A CN201680030571 A CN 201680030571A CN 107615298 A CN107615298 A CN 107615298A
- Authority
- CN
- China
- Prior art keywords
- face
- attribute
- image
- template
- vector
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
- G06V40/173—Classification, e.g. identification face re-identification, e.g. recognising unknown faces across different face tracks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Collating Specific Patterns (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of face identification method and system.The method comprising the steps of:A) one or more character image of display is read, b) judge whether described image shows the face of at least one personage, wherein, methods described only continues in the case where showing at least one face, c) image of the non-face attributive character of the face is analyzed, d) the facial attribute of the face is extracted from described image, e) face template being stored in database is ranked up and/or filtered by the non-face attributive character, f) database of the sequence and/or filtering is searched for, to obtain the face template to match with the face of described image.
Description
Technical field
The present invention relates to a kind of face identification method and system.
Background technology
The A1 of US 2014/0241574 disclose a kind of Face tracking and recognition method and device.By identifying the selected of face
The facial attribute simultaneously is compared to know by the facial attribute in region with the face data being stored in the database of known face
Others' thing.
The B2 of US 8,380,711 disclose a kind of method and system for being used to determine the order of classification of facial attribute.From face
Face area is estimated in portion's view data, and facial zone is determined in the attribute of these face areas and/or feature.By right
These attributes and feature carry out vector quantization, establish the grade figure of face recognition.The level plot shows point of the facial attribute
Level sequence.Therefore, one people of identification of facial attribute effectively can be passed through.
The A1 of US 2013/0129210 disclose a kind of introducing system identified based on face and style and method.Pass through face
Portion identifies, determines sex and age.Style identification includes the identification to the design and colors of clothes, and combine season, weather and
The information of time.The information identified using face and style is this life into one on hair, dressing, clothes and general
The recommendation of style.
The B2 of US 7,236,615 disclose a kind of Face datection based on energy model and expression appraisal procedure.This method
Multi views detector is set to be able to detect that the face of various expressions.So as to efficiently control the colour of skin, glasses, beard, light, ratio
The change of example and facial expression and other facial attributes or facial characteristics.
The A1 of US 2009/0087100 disclose a kind of device for calculating the crown position of people in image.This is by figure
Realize in the region for the people for finding out hair as carrying out oscillometry.By using this method, find and using in image
Face is as a reference point, to solve the problems, such as the Compositional balance in picture editting.
The A of CN 103679151 disclose a kind of method that face cluster is carried out in one or more image.This method
By the way that RGB image is converted into the gray level image for efficiency purpose to improve efficiency, and Gabor is extracted from gray level image
And/or local binary patterns (Local binary patterns, LBP) feature.The image for belonging to a people is clustered.
Other attributes, such as background, brightness, different facial expressions, body gesture, hair and hair style, and headwear etc. are obtained for
Effective control.
C.P. Georgia, the human relations of M. Austria and T. ripples are burnt;General framework for object detection;6th international computer vision
Meeting, 555-562 pages, 1998 be that first description Ha Er (Haar) small echo is used for one of publication of real-time objects detection.
From Borrow's viola and Michael Jones;Use the quick object detection of simple feature enhancing cascade, Mitsubishi's electronics
Co., Ltd of research laboratory, 2004 (TR-2004-043), positioned at the U.S. (computer vision and pattern-recognition meeting, 2001)
Cambridge of Massachusetts, and Borrow's viola and Michael Jones;Sane real-time objects detection;International computer vision
Journal, 57 (2):137-154,2002, a kind of method of face in automatic identification image, wherein Haar wavelet transform are used to detect Haar
Feature.This method uses so-called " general image ", and it is the middle representative of some image, wherein, all of above pixel and a left side
The summation of side pixel, plus the value of its own, each pixel as general image is signed.By using this distributed
A little mesh points, it can quickly calculate very much pixel in any rectangle of the image between four such complete pixels
Sum.Therefore, Haar wavelet transform very quickly can be applied in image.A kind of study based on AdaBoost faces is described to calculate
Method, the algorithm select a small amount of class Lis Hartel to levy from larger set, to obtain very efficient grader.These graders
A grader cascade is can be incorporated into, grader cascade allows the background area of image rapidly to be abandoned, and is having
More calculate is spent in desired subject area.
P.I. Wilson's, J. Fernandezs;Based on the facial features localization of Ha Er graders, JCSC 21,4 (2006
April), CCSC:In the meeting of the middle and south, another method that face in identification image is levied using class Lis Hartel is described.For
The image-region of face feature analysis is by region division to the position with maximum probability.By the detection zone of compartmentalization, disappear
Except reporting by mistake, reduce the area of detection, improve the speed of detection.
In Sebastian Schmidt, there is the real-time target detection of class Lis Hartel sign, on June 22nd, 2010, s-
Schmitt.de/ressourcen/haar_like_features.pdf, describe using detection object during class Lis Hartel levies in kind
Several projects.In these projects, the profiling feature of rotation has been used.In order to calculate the spy of the feature of rotation and axle alignment
Sign, the rotary area summation table (RSAT) of rotation integral image is used.
In F. A Badete, in C. horses AVM hereinafter and A. Bruces base (2010);The tracking of real-time face characteristic point, there is gold
Luas-Kanade algorithms, man-machine interaction, the great Zuo-Chu (chief editor) of word turriform, International Standard Book Number (ISBN):978-953-
307-051-3, InTech, it can be obtained from following network address:http://www.intechopen.com/books/human- robot-interaction/real-time-facial-feature-pointstracking-with-pyramidal- lucas-kanade-algorithm, a kind of method of the countenance tracking based on class Lis Hartel sign., can be with using this method
Selected facial feature points are tracked in the video sequence.
N. Dalaro, the auspicious lattice of B., for the histogram of the orientation gradient of mankind's detection, lear.inrialpes.fr/
People/triggs/pubs/Dalal-cvpr05.pdf, is published in computer vision and pattern-recognition, and 2005, CVPR
2005, IEEE computer societies 2005 meeting on June 25, (volume 1), the 886-893 pages, volume 1, ISSN 1063-6919,
ISBN 0-7695-2372-2 are printed in, AAP describes a kind of method for detecting the mankind in the picture, and this method is led to
Cross histograms of oriented gradients (Histogram of Oriented Gradient, HOG) the identification mankind.This method is to be based on
The normalization local histogram in image gradient direction is assessed in dense grid.Basic idea is can to pass through local strength's ladder
The distribution of degree or edge direction represents the outward appearance of native object and shape, in this embodiment it is not even necessary to it is accurate understand associated gradients or
Marginal position.This is that each unit is in list by the way that image window is divided into small area of space (" cell ") come what is realized
A Local gradient direction or the one dimensional histograms of edge direction are accumulated in first pixel.The table in the form of combining histogram entries
Show.With the grid of intensive (really overlapping) HOG descriptors with conventional based on SVMs (Support Vector
Machine, SVM) assemblage characteristic vector, the human testing chain based on window grader.
The content of the invention
It is an object of the invention to provide a kind of face identification method and system, can quickly identify with high reliability
People.
The purpose of the present invention is that solved by a kind of method and system according to independent claims.Accordingly from
Belong to advantageous embodiment of the invention disclosed in claim.
A kind of face identification method, methods described include step:
A) one or more character image of display is read,
B) judge whether described image shows the face of at least one personage, wherein, methods described is only in display at least one
Continue in the case of individual face,
C) image of the non-face attributive character of the face is analyzed,
D) the facial attribute of the face is extracted from described image,
E) face template being stored in database is ranked up and/or filtered by the non-face attributive character,
F) database of the sequence and/or filtering is searched for, to obtain the face mould to match with the face of described image
Plate.
Before the face that is matched with image of search, using non-face attribute to the face template that is stored in database
It is ranked up and/or filters, the quantity of the face template of matching can be greatly reduced, so as to soon searches for face template
Database, face template can also be accurately matched.Inventor have appreciated that the non-face attribute of people is very
Specifically.By that using only a small amount of non-face attribute, can be carried out to the face template being stored in database very much effective
Ground sorts and/or filtering.
The facial attribute of major part of different faces is closely similar.All faces include two eyes, a nose, a mouth
Bar, these arrangement of elements obtain quite similar.Therefore, corresponding attribute is substantially closely similar.It is only multiple such in combination
Facial attribute, different faces could be distinguished.On the contrary, non-face attribute is often very specific people.For example, clothes can show
Show very specific pattern and/or color, and hair quality can be very specific.Therefore, it is possible to use a small amount of non-face
Subordinate's property, for abandoning the major part for the face template being stored in database, and these templates do not have corresponding non-face
Subordinate's property.
That is, non-face attribute can be used for efficiently selecting the face template in database.By using a small amount of
Non-face attribute, such as:The colour of skin, clothes, hair style and glasses, the correlated measure of face template to be matched is reduced to and is stored in
The 0.5%-5% of all face templates in database.Therefore, can significantly acceleration search with extraction facial thumbnail
The face template of matching, can also highly precisely it carry out.
The present invention realizes carries out extensive real-time face identification to multiple video cameras, particularly on the same day.
Step c) and order d) can change, so as to first determine facial attribute, followed by non-face attribute, or by this
A little steps are combined as a step to extract facial attribute and non-face attribute.
Non-face attribute can include:
The color of skin, particularly neck color,
The shape of hair style including hair, the length of hair, the color of hair, the hair quality etc. of hair,
Color of the style of clothes including clothes, the quality of clothes, the style of clothes, collar,
Build includes neck shaped, shoulder shape,
Wearing spectacles,
Colorized glasses.
For the purpose of face recognition, in some non-face attributes, as the style and hair style of clothes only have in a short time
Effect, such as one day.Other non-face features, such as neck shaped, neck color, shoulder shape, generally keep in a long time
It is stable.Therefore, timestamp is distributed into non-face attribute helps to mark the time of shooting image or extracts from image non-face
The time of subordinate's property.When the face template being stored in database is sorted and/or filtered by non-face attribute, non-face attribute
Timestamp can be combined according to the average effective phase of each non-face attribute with weight.
Before step d) is performed, face's thumbnail can be selected from image.Preferably, the face is determined in step b)
Portion's thumbnail.Face's thumbnail has the size of face.Facial attribute is extracted from this face's thumbnail.The face is contracted
Sketch map performs the face template that matching is searched for according to step f).
The attribute thumbnail bigger than face thumbnail can be extracted from the image including face's thumbnail.Therefore, attribute
Thumbnail shows other parts of the people in addition to his/her face.These parts particularly including the hair of people, chest, neck
Portion and/or shoulder.Attribute thumbnail is preferably dimensioned to be 2 to 4 times more than face's thumbnail, so as to special comprising non-face attribute
Sign.Because such region is sufficiently large, the attribute of surrounding can be captured, but the people of interference nearby is captured without substantial amounts of chance
And background, the size of attribute thumbnail are desirably no more than the 2 of face's thumbnail, 3 or 4 times.
The method being ranked up by the image of wavelet transformation pair realizes that step b) detects to face.Wavelet transformation is excellent
Choosing is levied using two-dimentional Quasi-Haar wavelet to detect class Lis Hartel.The sort method (can be protected based on above-mentioned levied by Lis Hartel
Sieve viola and Michael Jones;Use the quick object detection-P.I. Wilson's of simple feature enhancing cascade, J. Fernands
Hereby;Facial features localization based on Ha Er graders-Sebastian Schmidt;Real-time target inspection with class Lis Hartel sign
Survey) method that is detected to the object in image.Therefore, these documents are very comprehensive.
The non-face attribute related to one of shape is determined by method for checking object or edge detection method.It is preferable right
As detection method is histogram of gradients.But also have other suitable edge detection methods, as Canny edge detection operators,
Canny-Deriche edge detectors, difference edge detection, Sobel operators, Prewitt operators and Roberts crossover operators.
The non-face attribute related to color is determined by method for detecting color.Preferable method for detecting color is that color is straight
Fang Tu.
The non-face attribute related to texture or pattern, such as local binary pattern are determined by texture sort method
Or Gabor filter (LBP).
Local binary patterns are respectively adopted or Gabor filter carries out texture sequence, from image or face's thumbnail extraction
Facial attribute.
The non-face attribute of the image of shooting can form non-face vector.Each face template of database includes non-face
Non-face vector corresponding to subordinate's property.Step e) filtering, non-face are performed by selecting in database all people's face template
The non-face vector of subordinate's property is nearer with the non-face vector distance of image, and exceedes predetermined threshold distance.
Such non-face vector can be also used for arranging the face template being stored in database according to step e)
Sequence, the face template of database is ranked up according to the non-face vector distance of the non-face vector of captured image.
, can be to individual by determining the distance of non-face vector and the non-face vector of captured image from face template
Other non-face attribute is weighted.The weight of the other non-face attribute of this can correspond to a tolerance, to determine
The value of corresponding non-face attribute.For example, the determination that can be perfectly clear only includes a kind of clothes of color, have than " clothes
Color " higher weight is the segment pattern with many different colours.The weight can also be combined with above-mentioned timestamp
Use.The weight of non-attribute corresponds to the stability of attribute.The non-face attribute relevant with clothes does not have the duration generally
Stability more than one day.Therefore, the weight more than the time will be substantially reduced.Have with hair color, hair quality or hair shape
The attribute of pass is generally more stable so that these non-face attributes have will not be as the non-face attribute of clothes over time
The weighting function for elapsing and declining.The non-face attribute related to neck shaped or shoulder shape is generally highly stable, therefore, this
A little non-face attributes have constant time weighting.
, can be by being ranked up to selected face template, or to from less than certain threshold according to step f) search
The ordering face template of the limited quantity with non-face vector distance extracted in the non-face vector of the image of value enters
Row sequence, is further ranked up on the basis of the facial attribute.Facial attribute preferably forms facial vector, so as to
Carried out with the distance between facial vector based on the image captured by the facial vector relative to the face template stored
Sequence.It can be ranked up by multi-dimensional indexing.
Multiple video cameras can be used for shooting multiple images, and carry out face recognition to each image.This method can be used for
The someone of some time frame of track.In time frame, it is necessary to select non-face attribute.It is all above-mentioned non-for the time frame of one day
Facial attribute is all suitable.In the case where frame length is more than one day, non-face subordinate of the selection with stronger time stability
Property.This method is also applied for monitoring or tracks everyone in a large amount of crowds.This finds out stream for the masses of monitoring outdoor activity
The criminals such as the common people are very favorable.
The face recognition method can also be used to determine customer behavior, such as assess advertisement measure or product introduction.This method
It can also be used to identify acceptance level of the customer to service and support center.
People of this method especially suitable for tracking and calculating sales department and public place, particularly and multi-camera system
It is combined.
The image for including face handled by the inventive method, can be caught by one or more video cameras.
These images can also extract from the database including multiple display facial images.
The invention further relates to a kind of system for recognition of face, the system includes taking the photograph at least one of shooting image
Camera and the control unit for being connected at least one video camera.The control unit is used to carry out face knowledge according to the above method
Not.
The system preferably includes multiple video cameras, for example, at least including five video cameras, preferably at least including ten
Video camera, and more preferably at least include 100 video cameras.Video camera can be placed on certain closed area.Video camera
Incoherent region, such as railway station, airport can be distributed in, with the personal movement of tracking.
Brief description of the drawings
The present invention is illustrated in greater detail by brief description of the drawings, wherein:
Fig. 1 is the block diagram of face identification system,
Fig. 2 is the schematic flow sheet of face identification method,
Fig. 3 is the module diagram of statistical data sampling program,
Fig. 4 a are a kind of simple class Lis Hartel collection,
Fig. 4 b are that a kind of class Lis Hartel of extension is collected,
Fig. 5 is levied to be a kind of by the first kind of AdaBoost algorithms selections and the second class Lis Hartel.
Embodiment
List of reference numbers
1 system
2 shopping paths
3 shopping centers
4 imports
5 outlets
6 bifurcated sections
7 central control units
8 processor units
9 storage mediums
10 video cameras
11 data wires
12 statistical data collection softwares
13 change detection modules
14 human detection modules
15 face detection modules
16 Haar features
17 subgraphs
18 images
19 subwindows
20 non-face property extracting modules
21 facial feature extraction modules
The preselected module of 22 templates
23 matching modules
24 statistical analysis modules
A kind of 25 internets
Fig. 1 show the embodiment for face identification system 1 according to the present invention, and the system designed to be used monitoring
The service condition of shopping path 2 in shopping center 3.
Shopping path 2 is extended between the entrance 4 in shopping center 3 and outlet 5.Shopping path 2 includes having multiple bifurcateds
The bifurcated of section 6.Customer passes through one or more bifurcated sections 6 in from No. 4 porch to the way of No. 5 outlets.Customer is according to its need
Wherein one or more bifurcated sections 6 are selected, product and advertising campaign are shown in bifurcated section 6.The behavior of customer is mainly produced
The influence of distribution and the advertising campaign of product.Therefore, statistics is shown, shown in the position of shopping route 2 some products or
Advertising campaign is very attractive to customer, and this handles very helpful to shopping center.
System 1 for face recognition allows to collect this statistics.
System 1 includes the central control unit 7 with processor unit 8 and the storage medium 9 for data storage storehouse.Place
Reason device unit 8 includes CPU, RAM (random access memory) and ROM (read-only storage).
Multiple video cameras 10 and central control unit 7 are connected by data wire 11.In the present embodiment, video camera 10 is still
It is image camera.Substantially, the combination of video camera or still image video camera and video camera can also be used.
Video camera 10 can also arrange the remote sites such as parking lot at the mall, and is connected to by internet 25
Entreat control unit 7.
Video camera 10 is the DV for producing electronically readable image file.During these image files are sent to
Entreat control unit 7.Software 12 is stored on central control unit 7 and is used for collection of statistical data, video camera 10 is delivered automatically
Image carry out recognition of face.
Statistical data collection software 12 includes multiple software modules (Fig. 3).Change detection module 13, for detecting input figure
It seem the change of the no prior images including same video camera.If an image is identical with previous image, need not enter
Row analysis, can abandon it.
Human detection module 14, at least one people whether is shown for detection image.
Face detection module 15, for one or more of detection image face.If face detection module 15 detects
To face, then it extracts face's thumbnail and attribute thumbnail.Face's thumbnail is the rectangle part of image, it is shown that from forehead
To the face of chin.Attribute thumbnail is a part for described image, that surrounds corresponding face's thumbnail and around described
The edge of face's thumbnail, the edge at least show the hair, neck and shoulder of the people related to the face.
Face detection module 15 make use of the object detection technique to image based on so-called class Lis Hartel sign.Class Lis Hartel
Sign represents the first feature not being rendered obvious by the image pixel intensities of image.In general, a class Lis Hartel sign is to figure
The difference of the mean intensity of subregion is encoded as in.Simplest feature set is by including the squares of two or four formed objects
The second zone composition (Fig. 4 a) of shape subregion.These class Lis Hartels sign is applied in image, calculates the pixel in subregion
It is worth sum, intensity difference is determined according in white subregions of Fig. 4 a in side and the shade subregion in opposite side.The difference table
Show characteristic value.
Feature can zoom in and out in its size, to obtain the characteristic information of different amplitudes.
The feature set extended as shown in Figure 4 b, including edge feature, line feature and center ring characteristics.Some 45 ° of rotations
Class Lis Hartel sign.
In order to calculate characteristic value in real time, so-called integral image or region summation table (SAT) are converted images into.
This region summation table has is assigned to the left side of original image and all pictures with original image identical size, each pixel
The summation of element.Once region summation table is calculated, just can be effectively to any in original image as long as calculating the summation of four values
Image pixel intensities in the sub-rectangular areas of axle alignment are summed.
In order to calculate the feature of the feature of rotation and axle alignment, rotary area summation table (RSAT) is used.In Rotary District
In domain summation table, the sum for the pixel distributed in original image each pixel, pixel arrangement in the original image is into edge
45 ° of rectangular area is tilted, forms the most right corner of rectangular area.
In order to further improve calculating speed, it is preferable that class Lis Hartel sign is applied in cascade, for image 17
Subwindow 16 is ranked up, and the subwindow will be analyzed existing face.Class Lis Hartel is taken over for use to be carried out in child windows 19
Sequence, so as to be referred to as Ha Er graders when applied to described image.
By the characteristic value of each Ha Er graders compared with feature weight, wherein, if characteristic value is more than or less than
Feature weight, then Ha Er graders are true or false, and vice versa.In the cascade of Ha Er graders, if a Ha Er classification
Device is false, then refuses subwindow 19, and terminates cascaded computation, and is further analyzed using the cascade of Ha Er graders
Another subwindow 19.
In order to detect the face feature of people, such as face, eyes and nose, it is necessary to train Ha Er graders to cascade.Many machines
Device learning method can be used for learning Ha Er graders.Preferable algorithm is AdaBoost learning process.Alternative
Habit process is the Variance feature selection of feature based, based on Wimnow indexes subscriber loops rule or uses neutral net or support
The feature selection process of the learning process of vector machine.
Fig. 5 is shown by the first kind of AdaBoost method choices and the second class Lis Hartel sign.The two Lis Hartels sign is aobvious
Show and gone on top, is then covered on the typical training face of bottom row.First feature is between measurement eye areas and cheek region
Strength difference.This feature is often more darker than cheek using eye areas.Second feature is by the intensity and mouth of eye areas
The intensity of bridge is compared.The embodiment is what Borrow's viola from the discussion above et al. obtained.
Can rapidly analyze multiple subwindows 19 using the face detection module 15, analyze different sizes in image,
The subwindow of diverse location.By first or at least by the subwindow of second Ha Er graders discarding only display background.
If detecting face, corresponding subwindow forms face's thumbnail.Based on face's thumbnail generation attribute contracting
Sketch map.Attribute thumbnail includes face's thumbnail and certain back gauge around face's thumbnail.Preferably, attribute thumbnail is face
Twice to four times of portion's thumbnail.
Non-face property extracting module 20, for extracting the non-face attribute of the people shown in image, these non-face subordinate
Property does not include the feature of this face.These non-face attributes include one or more with properties:Skin color, hair style, color development,
Hair quality, clothes color, clothes quality, clothes pattern, neck shape, neck sub-color, shoulder shape, wearing spectacles, colorized glasses,
Hair style at collar and/or whether there is collar.
The non-face attribute related to shape is determined by method for checking object or edge detection method.In preferred embodiment
In, histogram of gradients is used as the method for checking object of extraction shape association attributes.N. Dalaro et al.;Towards people's physical examination of gradient
Histogram is surveyed, as described above, discloses a kind of histogram of gradients for being used for extraction and shape association attributes.Therefore, by this article
Offer and be incorporated herein in full.
The non-face attribute of particular color in particular segment in image is determined by method for detecting color.In the present embodiment,
Using color histogram as method for detecting color, the dot frequency of particular color is determined according to the particular segment.
The non-face attribute related to texture or pattern is determined by texture sort method.The texture sequence of preferred embodiment
Method is local binary patterns case (LBP).
Facial property extracting module 21, the related feature of face for extracting to being detected.Facial property extracting module
The Lis Hartel determined by face detection module 15 can be replicated to levy, and be stored as facial attribute.In addition, or alternatively, enter
The facial attribute of one step can extract for example, by the texture sort method of local binary pattern.
Template preselects module 22, for the face being stored according to non-face Attributions selection in storage medium 9 in database
Template.Database in storage medium 9 includes the data set of multiple face templates.Each data set includes at least one non-face
Portion's vector, the non-face vector include non-face attribute and at least one facial vector, and the facial vector includes corresponding face
The facial attribute in portion.Preferably, data set also includes corresponding face and/or face's thumbnail of data seal or timestamp
And/or attribute thumbnail.
The preselected module 22 of template includes filtering and/or sort algorithm, for the people based on non-face attribute to database
Face template is filtered and/or sorted.This detects the non-of face by face detection module 15 by calculating in real image
The distance between non-face vector is carried out in the face template of facial vector database.
Face template is by the distance-taxis calculated, or is filtered according to this distance.If face template is carried out
Sequence, then selection have a number of face template of minimum range.This numeral can change from 10 to 10000, best
Not less than 100, particularly not less than 200, desirably no more than 2000, especially not greater than 1000 or 500.Selected face template
Quantity generally in the range of 0.5% to the 5% of non-selected face template.
If selecting face template using filter, only select those that there is the face mould less than certain threshold distance
Plate.Both selection modes are, it is necessary to further consider to be significantly reduced the quantity of face template.Preferably, it is preselected to template
Module 22 is adjusted, to no more than 10%, especially not greater than 5%, and the face template of preferably more than 2% database is done
Further processing.
The preselected module of template can be also used for abandoning the face template for showing non-face attribute.At the mall, employee
Often to wear specific clothes.Due to there was only customer, rather than staff comes under observation, and is related to this clothing so as to abandon
Take the related face template of the shopping center employee of attribute.
Matching module 23, optimal is carried out with the face detected in real image for searching for face template in database
Match somebody with somebody.
Best match is searched for according to facial attribute, especially by the face portion vector people detected in real image
Face's vector of face template scans for.Best match is between corresponding facial vector of the facial vector with face thumbnail
Face template with minimum range.Preferably, scanned for by multi-dimensional indexing.If it is not less than predetermined threshold distance
Matching, then result is " mismatch ".
Statistical analysis module 24, for carrying out statistical analysis to the face that detects, and by the information and additional information phase
With reference to, such as the time, corresponding to photograph during picture, or the position of people or the position of video camera in picture.
Based on above-mentioned face identification system, the invention discloses a kind of method of 3 collection statistics at the mall
(flow chart as shown in Figure 2).
Methods described is since step S1.
In step s 2, with one of shooting image of video camera 10.Video camera 10 can each certain time interval
Carry out shooting image.These intervals can be for example between 0.1s to 10s.Video camera 10 can also be connected with proximity transducer, with
The people in front of video camera is detected by the proximity transducer.The proximity transducer triggers the seizure to image.
Preferably, when shooting image, generate date stamp and attach it in described image.The date stamp includes clapping
Take the photograph the time of image, and/or the description of position shown in described image.The description of the position can be coordinate or art
Language, such as " shopping center entrance ".
Video camera 10 sends image by data wire 11 to central control unit 7.
If having any change in the last image shot with same video camera 10, checked by change detection module 13
Input picture (step S3).Due to having analyzed identical image before described image, if image does not change, lose
Abandon the image.If for nobody before specific video camera 10, video camera is continuously shot several identicals in 3 at the mall
Image.It is worth noting that, it is without in all senses that analysis again is carried out to same image.
If determining that image does not change in step s3, flow returns to step S2.If detect in step s3
To the change of image, then people (step S4) whether is shown in check image.Can be easily by histograms of oriented gradients
Detect the representative profiles of human body.If return to step S2 without display people, flow in image.If examine in step s 4
People is measured, then preferably, it is determined that the quantity with people in storage image.
Face detection module 15 passes through the face (step S5) in the analysis of above-mentioned class Lis Hartel sign and detection image.In the step
Face's thumbnail and attribute thumbnail are also generated in rapid.
Non-face property extracting module 20 extracts non-face attribute.In the present embodiment, it is only able to detect at the mall
Stop the people of at most several hours.Therefore, appropriate use non-face attribute is very important, but not in a long time
Effectively.The non-face attribute refers to all properties relevant with clothes and/or hair style.Duration of stay at the mall, anyone
All it is less likely to change the clothes or hair style of oneself.In other application, the different non-face attribute of selection that can be suitably.From
Non-face attribute is extracted in attribute thumbnail.
Facial feature extraction module 21 extracts (step S7) facial characteristics from face's thumbnail.Copy step can be passed through
The facial feature extraction face attribute being had determined in S5, such as class Lis Hartel sign, or by applying spy on face's thumbnail
Fixed extraction procedure.
By extracting non-face attribute (step S6), the preselected module 22 of template extracts the face mould in database in advance
Plate.By being pre-selected, a small amount of face template being stored in database only have selected.
The face template of these selections is used to search for the face's thumbnail generated in step s 5 and the face in database
Matching (step S9) between template.
If not finding matching in step s 9, into step S10.In step slo, new data set is added
In the database related to the face to being detected in actual photographed image.The data set comprises at least corresponding facial vector phase
The attribute vector answered.Preferably, the data set also includes the attribute of face's thumbnail and/or thumbnail.The data set can be with
Including the date stamp generated in step s 2, including time during shooting image and/or position.
The new face mould stored in the face template or step S10 that are matched in step S11 or step S9 in database
Plate carries out statistical analysis.In the present invention, which bifurcated section 6 of any personal use shopping path 2 analyzed.Further, it is also possible to
Analyze residence time of this person in the specific bifurcated section 6 of shopping path 2.These information can also be with the production of this people's actual purchase
Condition associates.Product that someone buys is determined by people corresponding to the detection in point of sale (POS), the information with cash register
The data registered on machine are associated.
In step s 12, further check whether to detect people in the image of reality.If it is the case, then flow
Journey returns to step S5 to detect next face.Otherwise, flow enters step S13, checks that center control is single in step s 13
Whether member 7 receives further image.Then, flow is back to step S3.Otherwise, this method terminates in step S14.
The above method is the example for collecting data at the mall.In the present example, the face shown in face recognition process
Portion's information is used for statistical analysis.This face recognition process can also be used for other application.By this face recognition process, such as
Group can be monitored, can be by everyone in non-face attribute easily tracking crowd.This can be used for outside monitoring room
The masses of activity, the outdoor activity may be disturbed by criminals such as such as rogues.The process can analyze multiple cameras simultaneously
Image or the multiple faces of display image.Once in database record a people, even if he change the position of oneself and
The image of different cameras shooting, can also find same person in real time.If some criminal is identified and found out of doors, very
Hardly possible isolates the criminal, and then, as long as this video camera is connected to face identification system, by camera supervised, this criminal holds very much
Easily it is caught on railway station or any other public place.
In the above-described embodiments, number is detected in step s 4, and detects face in step s 5.The two steps
Suddenly a step can also be merged into, Face datection is also used for detecting multiple faces or more personal respectively, and for calculating image
The number of middle display.
Further, thus it is possible to vary step S6 and S7 order.Step S5 and S7 can also be merged into a step, led to
Cross while detect face, extract face characteristic.This is particularly suitable as facial characteristics, such as class Lis Hartel sign.
This method and system can also be used for monitoring the safety-related field of such as bank.This method can be identified in one day
Several times close to the people of safety zone.
This method and system are additionally operable to the service processes in Analysis Service center, can reliably detect a certain customer
The time of service centre, and which place of customer Qu Guo service centres must stay in.
The general principle of the present invention is considered from the preselected a small amount of non-face attribute of template being stored in database.By
It is big in non-face attribute information amount, it is more likely that a small amount of potential correlate template is soon selected under high reliability.Therefore, may be used
With very quickly and efficiently find face template (" face ").The system and method are especially suitable within the limited time
People are monitored, such as during one to five hour, one to five day or one month.Must be non-to select according to the cycle of monitored people
Facial attribute.
The template in step S8 is carried out according to non-face attribute it is preselected, calculate corresponding between non-face vector
Distance.By calculating the distance, the time correlation weight of each attribute can also be used, due to attribute be present, is more likely sent out
Changing, and other attributes are stable.In addition, the tolerance for being determined or being estimated according to the value of respective attributes, can enter to attribute
Row weighting.Tolerance is smaller, and the respective weights of attribute are bigger.
Claims (15)
1. a kind of face identification method, it is characterised in that methods described includes step:
A) one or more character image of display is read,
B) judge whether described image shows the face of at least one personage, wherein, methods described is only showing at least one face
Continue in the case of portion,
C) image of the non-face attributive character of the face is analyzed,
D) the facial attribute of the face is extracted from described image,
E) face template being stored in database is ranked up and/or filtered by the non-face attributive character,
F) database of the sequence and/or filtering is searched for, to obtain the face template to match with the face of described image.
2. according to the method for claim 1, it is characterised in that before step d) is performed, picked out from described image
Face's thumbnail with face size size, face's attribute is extracted from face's thumbnail, carried by matching
Face's attribute of the thumbnail taken performs the search face template in the step f).
3. method according to claim 1 or 2, it is characterised in that the image ranking method based on wavelet transformation performs step
It is rapid b) in personage face detection.
4. according to the method for claim 3, it is characterised in that wavelet transformation is breathed out using two-dimentional Quasi-Haar wavelet to detect class
That feature.
5. according to any described methods of claim 1-4, it is characterised in that the non-face attribute include it is one or more with
It is properties:
Hair style, color development, hair quality, clothes color, clothes quality, clothes pattern, neck shape, neck sub-color, shoulder shape, eye
Mirror, collar.
6. according to the method for claim 5, it is characterised in that by edge detection method (histogram of gradients) determination and shape
The related non-face attribute, and/or
The related non-face attribute of color is determined at by method for detecting color (color histogram), or
The non-face attribute related to texture or pattern is determined by texture sort method (local binary patterns case).
7. according to any described methods of claim 1-6, it is characterised in that pass through texture sort method (local binary patterns case)
Face feature is extracted from described image.
8. according to any described methods of claim 1-7, it is characterised in that the non-face attribute of described image forms non-face
Vector, and each template of database includes non-face vector corresponding to non-face attribute, by selecting to own in database
Face template perform step e) filtering, the non-face vector distance of the non-face vector and image of the non-face attribute
It is relatively near, and exceed predetermined threshold distance.
9. according to any described methods of claim 1-8, it is characterised in that the non-face attribute of described image forms non-face
Vector, and each face template of database includes non-face vector corresponding to non-face attribute, by selecting in database
All people's face template performs step e) sequence, according to the distance pair of the non-face vector of the non-face vector of taken image
The face template of the database is ranked up.
10. method according to claim 8 or claim 9, it is characterised in that by determining that non-face vector is with being clapped from template
The distance of the non-face vector of image is taken the photograph, individual other non-face attribute is weighted.
11. according to any described methods of claim 1-10, it is characterised in that facial thumbnail is determined in step b), it is described
Facial thumbnail is used to extract facial characteristics, and the attribute thumbnail including facial thumbnail and more than facial thumbnail, institute
Attribute thumbnail is stated to be used to analyze non-face attribute.
12. according to any described methods of claim 1-11, it is characterised in that, can be by right according to step f) search
Selected face template is ranked up, or has non-face to what is extracted from the non-face vector of the image less than certain threshold value
The ordering face template of the limited quantity of portion's vector distance is ranked up, and is further entered on the basis of the facial attribute
Row sequence.
13. according to the method for claim 12, it is characterised in that be ranked up by multi-dimensional indexing.
14. according to any described methods of claim 1-12, it is characterised in that multiple video cameras can be used for shooting multiple figures
Picture, and recognition of face is carried out to each image.
15. a kind of face identification system, including:
At least one video camera for shooting image,
The control unit of at least one video camera is connected to, wherein, server is used to perform as described in claim 1 to 14
Method.
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
SG10201504080WA SG10201504080WA (en) | 2015-05-25 | 2015-05-25 | Method and System for Facial Recognition |
SG10201504080W | 2015-05-25 | ||
PCT/SG2016/050244 WO2016190814A1 (en) | 2015-05-25 | 2016-05-23 | Method and system for facial recognition |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107615298A true CN107615298A (en) | 2018-01-19 |
Family
ID=57392166
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201680030571.7A Pending CN107615298A (en) | 2015-05-25 | 2016-05-23 | Face identification method and system |
Country Status (6)
Country | Link |
---|---|
CN (1) | CN107615298A (en) |
AU (1) | AU2016266493A1 (en) |
HK (1) | HK1248018A1 (en) |
PH (1) | PH12017502144A1 (en) |
SG (1) | SG10201504080WA (en) |
WO (1) | WO2016190814A1 (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108724178A (en) * | 2018-04-13 | 2018-11-02 | 顺丰科技有限公司 | The autonomous follower method of particular person and device, robot, equipment and storage medium |
CN108805140A (en) * | 2018-05-23 | 2018-11-13 | 国政通科技股份有限公司 | A kind of feature rapid extracting method and face identification system based on LBP |
CN109448026A (en) * | 2018-11-16 | 2019-03-08 | 南京甄视智能科技有限公司 | Passenger flow statistical method and system based on head and shoulder detection |
CN109670451A (en) * | 2018-12-20 | 2019-04-23 | 天津天地伟业信息***集成有限公司 | Automatic face recognition tracking |
CN110213632A (en) * | 2019-04-23 | 2019-09-06 | 浙江六客堂文化发展有限公司 | A kind of audio/video player system and its application method comprising user data processing |
CN111161312A (en) * | 2019-12-16 | 2020-05-15 | 重庆邮电大学 | Object trajectory tracking and identifying device and system based on computer vision |
CN111554007A (en) * | 2020-04-20 | 2020-08-18 | 陈元勇 | Intelligent personnel identification control cabinet |
CN112749290A (en) * | 2019-10-30 | 2021-05-04 | 青岛千眼飞凤信息技术有限公司 | Photo display processing method and device and video display processing method and device |
CN113128356A (en) * | 2021-03-29 | 2021-07-16 | 成都理工大学工程技术学院 | Smart city monitoring system based on image recognition |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019083508A1 (en) | 2017-10-24 | 2019-05-02 | Hewlett-Packard Development Company, L.P. | Facial recognitions based on contextual information |
TR201721657A2 (en) * | 2017-12-25 | 2019-07-22 | Arcelik As | ONE FACE RECOGNITION SYSTEM AND METHOD |
US10936854B2 (en) * | 2018-04-27 | 2021-03-02 | Ncr Corporation | Individual biometric-based tracking |
CN112651268B (en) * | 2019-10-11 | 2024-05-28 | 北京眼神智能科技有限公司 | Method and device for eliminating black-and-white photo in living body detection and electronic equipment |
CN111597872A (en) * | 2020-03-27 | 2020-08-28 | 北京梦天门科技股份有限公司 | Health supervision law enforcement illegal medical practice face recognition method based on deep learning |
CN113822367B (en) * | 2021-09-29 | 2024-02-09 | 重庆紫光华山智安科技有限公司 | Regional behavior analysis method, system and medium based on human face |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1794264A (en) * | 2005-12-31 | 2006-06-28 | 北京中星微电子有限公司 | Method and system of real time detecting and continuous tracing human face in video frequency sequence |
US20060140455A1 (en) * | 2004-12-29 | 2006-06-29 | Gabriel Costache | Method and component for image recognition |
US20130121584A1 (en) * | 2009-09-18 | 2013-05-16 | Lubomir D. Bourdev | System and Method for Using Contextual Features to Improve Face Recognition in Digital Images |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7739221B2 (en) * | 2006-06-28 | 2010-06-15 | Microsoft Corporation | Visual and multi-dimensional search |
CN100568262C (en) * | 2007-12-29 | 2009-12-09 | 浙江工业大学 | Human face recognition detection device based on the multi-video camera information fusion |
US8379917B2 (en) * | 2009-10-02 | 2013-02-19 | DigitalOptics Corporation Europe Limited | Face recognition performance using additional image features |
-
2015
- 2015-05-25 SG SG10201504080WA patent/SG10201504080WA/en unknown
-
2016
- 2016-05-23 WO PCT/SG2016/050244 patent/WO2016190814A1/en active Application Filing
- 2016-05-23 AU AU2016266493A patent/AU2016266493A1/en not_active Abandoned
- 2016-05-23 CN CN201680030571.7A patent/CN107615298A/en active Pending
-
2017
- 2017-11-24 PH PH12017502144A patent/PH12017502144A1/en unknown
-
2018
- 2018-06-07 HK HK18107418.9A patent/HK1248018A1/en unknown
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060140455A1 (en) * | 2004-12-29 | 2006-06-29 | Gabriel Costache | Method and component for image recognition |
CN1794264A (en) * | 2005-12-31 | 2006-06-28 | 北京中星微电子有限公司 | Method and system of real time detecting and continuous tracing human face in video frequency sequence |
US20130121584A1 (en) * | 2009-09-18 | 2013-05-16 | Lubomir D. Bourdev | System and Method for Using Contextual Features to Improve Face Recognition in Digital Images |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108724178A (en) * | 2018-04-13 | 2018-11-02 | 顺丰科技有限公司 | The autonomous follower method of particular person and device, robot, equipment and storage medium |
CN108724178B (en) * | 2018-04-13 | 2022-03-29 | 顺丰科技有限公司 | Method and device for autonomous following of specific person, robot, device and storage medium |
CN108805140A (en) * | 2018-05-23 | 2018-11-13 | 国政通科技股份有限公司 | A kind of feature rapid extracting method and face identification system based on LBP |
CN109448026A (en) * | 2018-11-16 | 2019-03-08 | 南京甄视智能科技有限公司 | Passenger flow statistical method and system based on head and shoulder detection |
CN109670451A (en) * | 2018-12-20 | 2019-04-23 | 天津天地伟业信息***集成有限公司 | Automatic face recognition tracking |
CN110213632B (en) * | 2019-04-23 | 2021-07-30 | 浙江六客堂文化发展有限公司 | Video playing system containing user data processing and use method thereof |
CN110213632A (en) * | 2019-04-23 | 2019-09-06 | 浙江六客堂文化发展有限公司 | A kind of audio/video player system and its application method comprising user data processing |
CN112749290A (en) * | 2019-10-30 | 2021-05-04 | 青岛千眼飞凤信息技术有限公司 | Photo display processing method and device and video display processing method and device |
CN111161312A (en) * | 2019-12-16 | 2020-05-15 | 重庆邮电大学 | Object trajectory tracking and identifying device and system based on computer vision |
CN111161312B (en) * | 2019-12-16 | 2022-03-22 | 重庆邮电大学 | Object trajectory tracking and identifying device and system based on computer vision |
CN111554007A (en) * | 2020-04-20 | 2020-08-18 | 陈元勇 | Intelligent personnel identification control cabinet |
CN111554007B (en) * | 2020-04-20 | 2022-02-01 | 陈元勇 | Intelligent personnel identification control cabinet |
CN113128356A (en) * | 2021-03-29 | 2021-07-16 | 成都理工大学工程技术学院 | Smart city monitoring system based on image recognition |
Also Published As
Publication number | Publication date |
---|---|
HK1248018A1 (en) | 2018-10-05 |
PH12017502144A1 (en) | 2018-05-28 |
SG10201504080WA (en) | 2016-12-29 |
WO2016190814A1 (en) | 2016-12-01 |
AU2016266493A1 (en) | 2017-12-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107615298A (en) | Face identification method and system | |
Bialkowski et al. | A database for person re-identification in multi-camera surveillance networks | |
Liciotti et al. | Person re-identification dataset with rgb-d camera in a top-view configuration | |
Vittayakorn et al. | Runway to realway: Visual analysis of fashion | |
CN108520226B (en) | Pedestrian re-identification method based on body decomposition and significance detection | |
US9396412B2 (en) | Machine-learnt person re-identification | |
CN104166841B (en) | The quick detection recognition methods of pedestrian or vehicle is specified in a kind of video surveillance network | |
US20110142335A1 (en) | Image Comparison System and Method | |
Conde et al. | HoGG: Gabor and HoG-based human detection for surveillance in non-controlled environments | |
CN108416336A (en) | A kind of method and system of intelligence community recognition of face | |
CN110414441B (en) | Pedestrian track analysis method and system | |
CN107341688A (en) | The acquisition method and system of a kind of customer experience | |
CN109934047A (en) | Face identification system and its face identification method based on deep learning | |
Ancheta et al. | FEDSecurity: implementation of computer vision thru face and eye detection | |
Denman et al. | Searching for people using semantic soft biometric descriptions | |
CN109271932A (en) | Pedestrian based on color-match recognition methods again | |
Moctezuma et al. | Person detection in surveillance environment with HoGG: Gabor filters and histogram of oriented gradient | |
AU2017231602A1 (en) | Method and system for visitor tracking at a POS area | |
CN110443179A (en) | It leaves the post detection method, device and storage medium | |
Galiyawala et al. | Visual appearance based person retrieval in unconstrained environment videos | |
CN115661903B (en) | Picture identification method and device based on space mapping collaborative target filtering | |
CN112131477A (en) | Library book recommendation system and method based on user portrait | |
CN109858308B (en) | Video retrieval device, video retrieval method, and storage medium | |
Kumar et al. | Real-Time Face Mask Detection using Computer Vision and Machine Learning | |
Pavithra et al. | Survey on Face Recognition with pre filtering Techniques and their comparative study |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 1248018 Country of ref document: HK |
|
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180119 |
|
WD01 | Invention patent application deemed withdrawn after publication |