CN102426644A - Human eye cross localization method based on GentleBoost machine learning - Google Patents
Human eye cross localization method based on GentleBoost machine learning Download PDFInfo
- Publication number
- CN102426644A CN102426644A CN2011102437781A CN201110243778A CN102426644A CN 102426644 A CN102426644 A CN 102426644A CN 2011102437781 A CN2011102437781 A CN 2011102437781A CN 201110243778 A CN201110243778 A CN 201110243778A CN 102426644 A CN102426644 A CN 102426644A
- Authority
- CN
- China
- Prior art keywords
- eye
- sorter
- face
- left eye
- human
- 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
Links
Abstract
The invention provides a human eye cross localization method based on GentleBoost machine learning. The method comprises the following steps of: obtaining six classifiers by using a classifier training process on the basis of a Gentle Boost machine learning algorithm, wherein the six classifiers comprises a human face classifier, a double eye classifier, a left eye transverse classifier, a right eye transverse classifier, a left eye longitudinal classifier and a right classifier; carrying out multi-dimension two-dimensional sliding window searching on an image including human face by using the human face classifier so as to determine the position of the human face; and then carrying out cross searching on the human face region to obtain contour rectangles of the left eye and the right eye; and calculating centers of the contour rectangles of the left eye and the right eye, wherein the two center points are the central positions of the left eye and the right eye. According to the invention, the human eye searching is improved from two-dimensional sliding window searching to cross searching of multiple one-dimensional sliding windows so that the accuracy in detecting the human eye can be largely improved; meanwhile, the searching range is narrowed and the calculating quantity is reduced.
Description
Technical field
The present invention relates to mode identification technology, specifically be based on the human eye cross localization method of GentleBoost machine learning.
Background technology
The human eye location is meant at a width of cloth and has comprised in the image of people's face, automatically human eye is detected, and provides the coordinate of left eye and right eye center.Accurate human eye location is a committed step of carrying out recognition of face, and the degree of accuracy of human eye location directly has influence on the accuracy of recognition of face.At present; Existing method about the human eye location is based on top-down searching method of refining step by step mostly, promptly at first obtains the sorter of people's face and human eye through the means (like Adaboost algorithm, SVM or the like) of machine learning; The search of then image that has comprised people's face being refined step by step; Find out the zone that has people's face in the image earlier, in the zone that has people's face, carry out more fine-grained search again, thereby confirm the exact position of human eye.
Sorter training method based on the AdaBoost machine learning algorithm is the common method of training classifier in the present human eye detection.The Adaboost machine learning algorithm is to solve two types of partition problems effective ways of (as confirming whether a width of cloth picture is people's face picture); This method is screened according to the characteristic set (like the Haar characteristic set) of given sample set (comprising positive sample and negative sample) to such sample, and then the characteristic that filters out is constituted a sorter through weighting.The sorter that is obtained by this method possesses higher accuracy and robustness, is widely used in the various PRSs, like person detecting, the detection of people's face, human eye detection system or the like.
Be that the input picture is at first through once multiple dimensioned two-dimentional moving window search as the top-down searching method of the elite step by step step 1; Confirm people's face region; And then at the half of search in the left side of people's face region left eye; At the half of again search right eye of people's face region, searching method all is two-dimentional moving window method.Existing human eye localization method is in the search procedure of carrying out people's face and human eye; All be that the mode with two-dimentional moving window is searched on different yardsticks; This method need be slided zone to be detected two dimensions; Need test a large amount of rectangular areas, to confirm whether it is people's face or human eye.This search procedure needs bigger operand; Simultaneously because near the target area (people's face or human eye area); Usually can detect the position of a plurality of doubtful people's faces or human eye, these class methods generally average the position of these doubtful people's faces, thereby obtain final people's face or position of human eye; The testing result that obtains like this comprises bigger stochastic error, can produce bigger influence to follow-up face recognition process.
In order to reduce the stochastic error that produces in the human eye position fixing process as far as possible, the present invention proposes the cross localization method of human eye, and the hunting zone in the human eye position fixing process is limited; Can dwindle the hunting zone on the one hand; Reduce calculated amount, reduce stochastic error on the other hand, improve accuracy.
Summary of the invention
The technical matters that the present invention solved is to provide the human eye cross localization method based on the GentleBoost machine learning, to solve the shortcoming in the above-mentioned background technology.
The technical matters that the present invention solved adopts following technical scheme to realize:
Human eye cross localization method based on the GentleBoost machine learning may further comprise the steps:
The first step: be the basis with the GentleBoost machine learning algorithm; Obtain six sorters through the sorter training process: people's face sorter, eyes sorter, left eye horizontal sorter, right eye horizontal sorter, vertical sorter of left eye and the vertical sorter of right eye;
Second step: end user's face sorter carries out multiple dimensioned two-dimentional moving window search to the image that comprises people's face, confirms the position of people's face; And then, human face region is carried out the right-angled intersection search, obtain the gabarit rectangle of left eye and right eye;
The 3rd step: calculate the center of left eye and right eye gabarit rectangle, these two centers that central point is exactly left eye and right eye.
Among the present invention, said sorter training process may further comprise the steps:
The first step: sample is demarcated substep: choose the suitable sample set merging that comprises facial image the manual work demarcation is carried out in the position of people's face in this set and human eye, to obtain the coordinate position of human eye in the sample image;
Second step: training sample generates substep:, sample image is cut apart and size normalization composition and classification device training sample according to the calibration result of sample set;
The 3rd step: be the basis with the GentleBoost machine learning algorithm; With the training sample that forms serves as that input is trained; Obtain six sorters: people's face sorter, eyes sorter, left eye horizontal sorter, right eye horizontal sorter, the vertical sorter of left eye and the vertical sorter of right eye.
Among the present invention, said right-angled intersection search procedure may further comprise the steps:
The first step: use the eyes sorter that human face region is carried out one dimension moving window search from top to bottom, to confirm the position of eyes;
Second step: in detected eyes zone; Use left eye horizontal classification device to carry out from a left side and the search of right one dimension moving window in its left-half; To confirm the horizontal level of left eye; Use right eye horizontal classification device to carry out from a left side and the search of right one dimension moving window, to confirm the horizontal level of right eye at its right half part;
The 3rd step: in the scope with the horizontal level of left eye and right eye, use right and left eyes vertical classification device to carry out one dimension moving window search from top to bottom respectively, with the upright position of definite left eye and right eye.
Beneficial effect
The search that the present invention is directed to human eye is a plurality of one dimension moving window search of right-angled intersection mode by two-dimentional moving window search for improved, can significantly improve the degree of accuracy of human eye detection, dwindles the hunting zone simultaneously, reduces calculated amount.The method of the invention is that the modified version of Adaboost algorithm has higher speed of convergence and robustness.
Embodiment
For technological means, creation characteristic that the present invention is realized, reach purpose and effect and be easy to understand and understand, below in conjunction with concrete diagram, further set forth the present invention.
Practical implementation based on the human eye cross localization method of GentleBoost machine learning has two committed steps; One is the training of sorter, and this comprises six sorters: people's face sorter, eyes sorter, left eye horizontal classification device, right eye horizontal classification device, left eye vertical classification device and right eye vertical classification device; Another is the cross search procedure.Carry out detailed description in the face of these two committed steps down.
In carrying out the sorter training process; At first need carry out manual work to the image pattern that comprises people's face demarcates; So that the sample that needs in the generation training process; In carrying out the artificial calibration process of facial image sample,, need to demarcate three points: left eye angle, right eye angle and pupil center for each eye.If their coordinates in image are respectively (x
l, y
l), (x
r, y
r), and (x
p, y
p), the centre coordinate (x of eyes so
c, y
c) can calculate through following formula:
Successfully carrying out after image pattern demarcates, also just obtaining the centre coordinate position of left eye and right eye in the image pattern, so just can train cutting apart and normalization with positive sample.As shown in Figure 2, be people's face, eyes, the simple eye dividing method that reaches vertical simple eye positive sample of level, its median ocellus sample is example with the left eye, right eye is the mirror image of left eye.If two oculocentric distances, then can calculate positive sample size of people's face and the positive sample size of cutting apart of eyes; The simple eye sample size of level is d * d; After vertical simple eye sample size was cut apart completion for
, owing to varying in size of people face part in every pictures, the positive size that obtains also had nothing in common with each other; Therefore; Need zoom to unified size to size normalization, zoom to 22 * 22 by the face sample image as saying everyone; The eyes sample image zooms to 32 * 8, the image zoom to 16 of the simple eye sample of level * 16.
In the training of sorter, not only to use positive sample, also need suitable negative sample.When carrying out the negative sample generation,, take to obtain from the mode that the background parts beyond people's face extracts at random for people's face negative sample; For the eyes negative sample, then in human face region, vertically extract at random, for the simple eye sample of level, then along continuous straight runs extracts at random in the eyes zone, for vertical simple eye negative sample, then in vertical simple eye zone, vertically extracts at random.
After sample was ready, the sorter that can carry out being the basis with the GentleBoost machine learning algorithm was trained, and used the haar characteristic as Weak Classifier in the training process.
After 6 sorters are trained completion respectively, promptly can be used for the detection of human eye.The right-angled intersection search is used for accurately locating the central point of eyes; Search procedure is divided into four levels: first level is the two-dimension human face range searching of coarseness; Carry out multiple dimensioned two-dimentional moving window search through end user's face sorter on entire image, finally confirm the zone at people's face place; Second level uses the eyes sorter to carry out the search of one dimension moving window in people's face region from top to bottom, finally confirms the zone at eyes place; The 3rd level is from a left side and the simple eye sorter of right usage level carries out the search of one dimension moving window, the horizontal level of finally definite left eye and right eye in the zone at eyes place; The 4th level is in the zone that the simple eye sorter search of level obtains, and uses vertical simple eye sorter to carry out the one dimension moving window search of vertical direction, finally obtains the upright position of left eye and right eye.
Human eye cross localization method can also be carried out some improvement according to actual operating position.Provide wherein a kind of improvement design commonly used below: comprise the human eye cross localization method that glasses are handled.The necessary in actual use correctly personage of identification tape clear ophthalmic glasses of face identification system, this location algorithm that will ask for help eye not only can be handled the situation of not wearing glasses, and also must can handle the situation of wearing glasses.Because it is significantly different that the eyes of people's face of wearing glasses part and the eyes of not wearing glasses partly have, therefore must be directed against the human eye sorter that the sample training of wearing glasses, comprise horizontal classification device and vertical classification device.In addition, also must have a kind of means be used to judge current detection to people's face whether be people's face of wearing glasses.
Address these problems; Must increase the quantity of sorter, increase to 12: be i.e. people's face sorter, eyes sorter, the eyes sorter of wearing glasses, the identification and classification device of wearing glasses, left eye horizontal classification device, the left eye horizontal classification device of wearing glasses, right eye horizontal classification device, the right eye horizontal classification device of wearing glasses, left eye vertical classification device, wear glasses left eye vertical classification device, right eye vertical classification device and the right eye vertical classification device of wearing glasses through the sorter that the GentleBoost machine learning algorithm trains.Its method is: at first input picture is through multiple dimensioned two-dimentional moving window search; Obtain the position of people face part in image; Use the eyes sorter and the eyes sorter of wearing glasses that human face region is carried out the search of vertical one-dimensional moving window then respectively; Can obtain two eyes position Detection results so respectively; Then use the identification and classification device of wearing glasses and two eyes position Detection results are judged select correct eyes position, and whether the people's face in definite image is people's face of wearing glasses; If last this people's face is worn glasses; Then use the human eye horizontal classification device of wearing glasses the search of one dimension moving window to be carried out in the eyes region with the vertical classification device; If this people's face is not worn glasses; Then use the human eye horizontal classification device of not wearing glasses the search of one dimension moving window to be carried out in the eyes region, finally obtain the exact position of eyes with the vertical classification device.
More than show and described ultimate principle of the present invention and principal character and advantage of the present invention; The technician of the industry should understand; The present invention is not restricted to the described embodiments; That describes in the foregoing description and the instructions just explains principle of the present invention, and under the prerequisite that does not break away from spirit and scope of the invention, the present invention also has various changes and modifications; These variations and improvement all fall in the scope of the invention that requires protection, and the present invention requires protection domain to be defined by appending claims and equivalent thereof.
Claims (3)
1. based on the human eye cross localization method of GentleBoost machine learning, it is characterized in that, may further comprise the steps:
The first step: be the basis with the GentleBoost machine learning algorithm; Obtain six sorters through the sorter training process: people's face sorter, eyes sorter, left eye horizontal sorter, right eye horizontal sorter, vertical sorter of left eye and the vertical sorter of right eye;
Second step: end user's face sorter carries out multiple dimensioned two-dimentional moving window search to the image that comprises people's face, confirms the position of people's face; And then, human face region is carried out the right-angled intersection search, obtain the gabarit rectangle of left eye and right eye;
The 3rd step: calculate the center of left eye and right eye gabarit rectangle, these two centers that central point is exactly left eye and right eye.
2. the human eye cross localization method based on the GentleBoost machine learning according to claim 1 is characterized in that said sorter training process may further comprise the steps:
The first step: sample is demarcated substep: choose the suitable sample set merging that comprises facial image the manual work demarcation is carried out in the position of people's face in this set and human eye, to obtain the coordinate position of human eye in the sample image;
Second step: training sample generates substep:, sample image is cut apart and size normalization composition and classification device training sample according to the calibration result of sample set;
The 3rd step: be the basis with the GentleBoost machine learning algorithm; With the training sample that forms serves as that input is trained; Obtain six sorters: people's face sorter, eyes sorter, left eye horizontal sorter, right eye horizontal sorter, the vertical sorter of left eye and the vertical sorter of right eye.
3. the human eye cross localization method based on the GentleBoost machine learning according to claim 1 is characterized in that said right-angled intersection search procedure may further comprise the steps:
The first step: use the eyes sorter that human face region is carried out one dimension moving window search from top to bottom, to confirm the position of eyes;
Second step: in detected eyes zone; Use left eye horizontal classification device to carry out from a left side and the search of right one dimension moving window in its left-half; To confirm the horizontal level of left eye; Use right eye horizontal classification device to carry out from a left side and the search of right one dimension moving window, to confirm the horizontal level of right eye at its right half part;
The 3rd step: in the scope with the horizontal level of left eye and right eye, use right and left eyes vertical classification device to carry out one dimension moving window search from top to bottom respectively, with the upright position of definite left eye and right eye.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110243778 CN102426644B (en) | 2011-08-24 | 2011-08-24 | Human eye cross localization method based on GentleBoost machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110243778 CN102426644B (en) | 2011-08-24 | 2011-08-24 | Human eye cross localization method based on GentleBoost machine learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102426644A true CN102426644A (en) | 2012-04-25 |
| CN102426644B CN102426644B (en) | 2013-07-10 |
Family
ID=45960623
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201110243778 Expired - Fee Related CN102426644B (en) | 2011-08-24 | 2011-08-24 | Human eye cross localization method based on GentleBoost machine learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102426644B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103678984A (en) * | 2013-12-20 | 2014-03-26 | 湖北微模式科技发展有限公司 | Method for achieving user authentication by utilizing camera |
| CN106709420A (en) * | 2016-11-21 | 2017-05-24 | 厦门瑞为信息技术有限公司 | Method for monitoring driving behaviors of driver of commercial vehicle |
| CN106991363A (en) * | 2016-01-21 | 2017-07-28 | 北京三星通信技术研究有限公司 | A kind of method and apparatus of Face datection |
| CN109994202A (en) * | 2019-03-22 | 2019-07-09 | 华南理工大学 | A method of the face based on deep learning generates prescriptions of traditional Chinese medicine |
| CN110705468A (en) * | 2019-09-30 | 2020-01-17 | 四川大学 | Eye movement range identification method and system based on image analysis |
| CN112825115A (en) * | 2019-11-20 | 2021-05-21 | 北京眼神智能科技有限公司 | Monocular image-based glasses detection method and device, storage medium and equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101196998A (en) * | 2007-12-29 | 2008-06-11 | 北京中星微电子有限公司 | Eye locating method and device |
| CN101984453A (en) * | 2010-11-02 | 2011-03-09 | 中国科学技术大学 | Human eye recognition system and method |
-
2011
- 2011-08-24 CN CN 201110243778 patent/CN102426644B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101196998A (en) * | 2007-12-29 | 2008-06-11 | 北京中星微电子有限公司 | Eye locating method and device |
| CN101984453A (en) * | 2010-11-02 | 2011-03-09 | 中国科学技术大学 | Human eye recognition system and method |
Non-Patent Citations (1)
| Title |
|---|
| 王基帆等: "基于Adaboost算法的人眼定位", 《现代计算机》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103678984A (en) * | 2013-12-20 | 2014-03-26 | 湖北微模式科技发展有限公司 | Method for achieving user authentication by utilizing camera |
| CN106991363A (en) * | 2016-01-21 | 2017-07-28 | 北京三星通信技术研究有限公司 | A kind of method and apparatus of Face datection |
| CN106991363B (en) * | 2016-01-21 | 2021-02-09 | 北京三星通信技术研究有限公司 | Face detection method and device |
| CN106709420A (en) * | 2016-11-21 | 2017-05-24 | 厦门瑞为信息技术有限公司 | Method for monitoring driving behaviors of driver of commercial vehicle |
| CN106709420B (en) * | 2016-11-21 | 2020-07-10 | 厦门瑞为信息技术有限公司 | Method for monitoring driving behavior of commercial vehicle driver |
| CN109994202A (en) * | 2019-03-22 | 2019-07-09 | 华南理工大学 | A method of the face based on deep learning generates prescriptions of traditional Chinese medicine |
| CN110705468A (en) * | 2019-09-30 | 2020-01-17 | 四川大学 | Eye movement range identification method and system based on image analysis |
| CN110705468B (en) * | 2019-09-30 | 2022-08-30 | 四川大学 | Eye movement range identification method and system based on image analysis |
| CN112825115A (en) * | 2019-11-20 | 2021-05-21 | 北京眼神智能科技有限公司 | Monocular image-based glasses detection method and device, storage medium and equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102426644B (en) | 2013-07-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103093215B (en) | Human-eye positioning method and device | |
| CN102426644B (en) | Human eye cross localization method based on GentleBoost machine learning | |
| CN103632136B (en) | Human-eye positioning method and device | |
| CN100561503C (en) | A kind of people's face canthus and corners of the mouth location and method and the device followed the tracks of | |
| CN101840509B (en) | Measuring method for eye-observation visual angle and device thereof | |
| CN105759971A (en) | Method and system for automatically prompting distance from human eyes to screen | |
| CN106022304A (en) | Binocular camera-based real time human sitting posture condition detection method | |
| CN101984453B (en) | Human eye recognition system and method | |
| CN102902967A (en) | Method for positioning iris and pupil based on eye structure classification | |
| CN101923645A (en) | Iris splitting method suitable for low-quality iris image in complex application context | |
| CN105740779A (en) | Method and device for human face in-vivo detection | |
| CN103810491A (en) | Head posture estimation interest point detection method fusing depth and gray scale image characteristic points | |
| CN106203375A (en) | A kind of based on face in facial image with the pupil positioning method of human eye detection | |
| CN104050448A (en) | Human eye positioning method and device and human eye region positioning method and device | |
| CN101159018A (en) | Image characteristic points positioning method and device | |
| CN105138983B (en) | The pedestrian detection method divided based on weighting block model and selective search | |
| CN103390151B (en) | Method for detecting human face and device | |
| CN105512630B (en) | Human eye detection and localization method | |
| CN109670440A (en) | The recognition methods of giant panda face and device | |
| CN106503645A (en) | Monocular distance-finding method and system based on Android | |
| CN109766796A (en) | A kind of depth pedestrian detection method towards dense population | |
| CN108537787A (en) | A kind of quality judging method of facial image | |
| CN106570447A (en) | Face photo sunglass automatic removing method based on gray histogram matching | |
| CN104346621A (en) | Method and device for creating eye template as well as method and device for detecting eye state | |
| CN102129556B (en) | Judging method of definition of iris image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| ASS | Succession or assignment of patent right |
Owner name: HUNAN AIVAS SCIENCE + TECHNOLOGY CO., LTD. Free format text: FORMER OWNER: XIA DONG Effective date: 20130109 |
|
| C41 | Transfer of patent application or patent right or utility model | ||
| COR | Change of bibliographic data |
Free format text: CORRECT: ADDRESS; FROM: 410000 CHANGSHA, HUNAN PROVINCE TO: 410205 CHANGSHA, HUNAN PROVINCE |
|
| TA01 | Transfer of patent application right |
Effective date of registration: 20130109 Address after: 410205 Hunan province Changsha high tech Zone Lu Jing Road No. 2 building two floor productivity wealth center Applicant after: HUNAN LINGCHUANG INTELLIGENT SCIENCE & TECHNOLOGY CO., LTD. Address before: 410000 Hunan province Changsha Kaifu District, No. 10 century ship road trip spring Jinyuan 3 block J Applicant before: Xia Dong |
|
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20130710 Termination date: 20140824 |
|
| EXPY | Termination of patent right or utility model |