CN105701464A - Method of determining face detection false detection and key point positioning accuracy - Google Patents
Method of determining face detection false detection and key point positioning accuracy Download PDFInfo
- Publication number
- CN105701464A CN105701464A CN201610017969.9A CN201610017969A CN105701464A CN 105701464 A CN105701464 A CN 105701464A CN 201610017969 A CN201610017969 A CN 201610017969A CN 105701464 A CN105701464 A CN 105701464A
- Authority
- CN
- China
- Prior art keywords
- key point
- face
- face datection
- location
- sample
- 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
- G06V40/161—Detection; Localisation; Normalisation
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)
Abstract
The method discloses a method of determining face detection false detection and key point positioning accuracy. The method comprises the following steps of carrying out face detection; carrying out N key point positioning on a detected face and calling M binary classifiers to determine the number of inaccurate points in M key points; and designing a method to determine whether the face detection is false detection and the key point positioning is accurate. The method has advantages that in a face key point positioning link of a whole face identification system, some important key point positioning accuracy is determined so as to carry out backstepping on previous link face detection and determine whether the detection is the false detection; and purposes of determining whether the face detection is the false detection and simultaneously determining the key point positioning accuracy are reached, and double benefits are possessed.
Description
Technical field
The invention belongs to technical field of face recognition, it is proposed that a kind of method judging Face datection flase drop and key point positional accuracy。
Background technology
Current face recognition technology generally comprises several step: Face datection, face key point location, face characteristic are extracted, face alignment。Wherein Face datection and face key point location are most basic links, and Face datection refers to the image given for a width, it is determined that wherein whether containing face, if had, returning the position of all human face regions, size and attitude；Face key point location refers on Face datection basis, orients the positions such as the eyes of face, eyebrow, nose, face, profile further。The effect quality of these two technology directly affects follow-up face characteristic and extracts the effect with face alignment。These two technology are through development for many years at present; accuracy rate has been obtained for large increase; but in realistic individual face identification product; owing to scene is complex; even if current FA Face datection also often there will be flase drop phenomenon; it is demarcated as human face region, as being face by region labelings such as trees, building, vehicles by non-face image-region；And reasonable face key point location algorithm is as illumination, attitude and expression shape change, occurring positioning inaccurate situation, in the eyes of the face namely oriented, eyebrow, nose, face, the coordinate position of profile and facial image, there is bigger deviation real position。
In systems in practice, if there is Face datection flase drop, then also can there is bigger mistake based on face key point thereon location, then extract at follow-up face characteristic, face alignment arises that serious problems, also increase the burden that system is run simultaneously, additionally when more serious mistake occurs above two links, enter back into follow-up link less than too big meaning。This phenomenon is even more serious in the recognition of face scene coordinated without user, in safety monitoring scene, scene is complicated and changeable, user mismatches, there will be substantial amounts of Face datection flase drop and the inaccurate phenomenon of facial modeling, at this moment, if system can not judge Face datection flase drop or face key point Wrong localization, non-face so in a large number or poor quality picture (can accurately detect face, but facial modeling is forbidden) will enter in follow-up face alignment link, the burden of system will become very big, and recognition effect is had a greatly reduced quality。
One correct logic should be, position link at Face datection and face key point to check on, method can be had to judge, and whether Face datection whether flase drop, face key location be accurate, the qualified just follow-up link of entrance, do so can eliminate some insignificant identifications, improve the utilization rate of system, also improve the discrimination of face identification system to a certain extent。Therefore, design an algorithm and can interpolate that whether Face datection judges by accident and whether face key point location is accurate, for a practical face identification system, it appears be highly desirable to。
At present in disclosed document, it is all the precision said and how to improve Face datection and the accuracy rate improving face key point location, but when background condition is complicated, Face datection still there will be flase drop phenomenon, and also can there is the inaccurate situation in location in face key point location, and these documents are all solve Face datection and face key point orientation problem in isolation, finally whether accurate to judge its Face datection whether flase drop or face key point location also without providing a method。
Summary of the invention
The technical problem to be solved is, a kind of method judging Face datection flase drop and key point positional accuracy is provided, Face datection whether flase drop can be pushed away by judging that accuracy that face key point positions is counter, there is the effect killed two birds with one stone, can decide whether to carry out follow-up face alignment link according to the Face detection accuracy obtained, thus alleviating the burden of system simultaneously。
For solving above-mentioned technical problem, a kind of method judging Face datection flase drop and key point positional accuracy provided by the invention, including step:
A. Face datection；
B. on the human face region detected, carry out the location of N number of key point；
C. in N number of key point, investigate wherein M primary focus, call M two graders and judge the number of inaccurate point, wherein M≤N in this M key point；
If being forbidden≤the a*M that counts, then judge that Face datection and critical point detection are all very reliable, continue follow-up link；
If being forbidden >=the b*M that counts, then judging that Face datection there occurs erroneous judgement, key point location is meaningless, stops follow-up link；
If <being forbidden to count, <b*M then judges that Face datection is reliable, but Partial key point location is forbidden, and decides whether to continue follow-up link according to the situation of location of mistake a*M；
In step (c), two classifier training steps include:
(c.1) generate positive sample, choose the K artificial face key point sample demarcated opened, to every pictures centered by certain key point, extract a local feature as positive sample in a zonule about；Random offset 1-2 pixel around real key point, the key point after skew, it is also considered as correct key point, then centered by the pixel after skew, extracts local feature as positive sample again, wherein K >=10000；
(c.2) negative sample is generated, with positive sample with in the face key point sample of a collection of artificial demarcation, random offset 5-10 pixel around real key point, key point after skew, can be used to the inaccurate point of simulator locating, then centered by the pixel after skew, extract local feature as negative sample again；
(c.3) training two classification, for each key point according to (c.1) and (c.2) method, generates large number of positive sample and negative sample, now just can train two graders。
Further, in step (c), a is 0.05-0.15, b is 0.35-0.45。
Further, in step (c.3), two classification adopt SVM classifier。
After adopting above structure, the method judging Face datection flase drop and key point positional accuracy of the present invention is compared with prior art, have the advantage that in the face key point location link in whole face identification system, judge some important key point accurate positioning degree, the anti-link Face datection pushed away before it whether flase drop, thus having reached both to may determine that Face datection whether flase drop, the purpose of key point positional accuracy can also be judged simultaneously, kill two birds with one stone。
Accompanying drawing explanation
Fig. 1 is flow chart of the present invention；
Fig. 2 is the present invention two classification based training flow chart。
Detailed description of the invention
In conjunction with accompanying drawing, the invention will be further described by the examples below。
As it is shown in figure 1, the present embodiment provide judge Face datection flase drop and the method for key point positional accuracy, including step: first carry out Face datection, afterwards near certain key point of face extraction one local feature, this feature can meet a distribution；And extract a same local feature deviateing the bigger position of this key point, then the distribution of this local feature has certain discrimination with the former, so for this key point, it is possible to design two graders distinguish this key point position whether accurate。According to such thinking, multiple main key points in face key point can be chosen, each key point separately designs two graders, it may determine that the whether accurate positioning of these key points, situation further according to these key point positional accuracies, carry out comprehensive descision, counter can also push away whether Face datection is flase drop simultaneously。
The span of a and b is the empirical value obtained according to practical situation, in the present embodiment, a=0.1, b=0.4, judged result can be made more accurate, if key point location algorithm orients N number of key point, therefrom choose M key point (M≤N, generally can take M=N, concrete M is how many, it is typically based on practical situations, if follow-up recognizer needs all key points will be accurate, then M=N, if the accuracy requirement of some point is not as high by follow-up recognizer, then some point can train two graders, now M < N), M two graders of training, this M grader can input M result, result is that this point location of 1 expression is accurate,-1 represents that this point location is inaccurate, now there is following three situation:
(1) some result only few in classification results is shown as-1, number on schedule is not less than or equal in 0.1*M(such as 25 classification results, at most only two points are forbidden), now can be seen that inaccurate point is considerably less, this positioning result is very reliable, the previous link Face datection simultaneously also demonstrating this location be also it can trust that。
(2) classification results there is most of result be shown as-1, forbidden to count out be more than or equal in 0.4*M(such as 25 classification results, have more than 10 points to be forbidden), now represent that inaccurate point is very many, the key point of this location has serious problems, but owing to the accuracy of current various face key point location algorithms is significantly high, so this is an exception。Can counter releasing, the root that location is forbidden not is the problem in key point location own, but owing to Face datection there occurs erroneous judgement, because Face datection outputs non-face region and positions in link to face key point, so that substantial amounts of key point location is inaccurate。Through substantial amounts of test, also demonstrate such rule。
(3) if there being least a portion of result to be shown as-1 in classification results, forbidden to count out (in such as 25 classification results, has 5 points to be forbidden) between 0.1*M and 0.4*M, and specifically which point is inaccurate, it is also possible to directly observed by grader。Situation now is usually: Face datection is correct, but face key point is positioned at partial dot location and is forbidden。Because finding certain law in testing, when there is bigger illumination attitude expression shape change, the key point being likely to have part difficulty to position may be forbidden, but inaccurate number also controls at less than 1/3rd of sum, therefore there is such situation, may determine that Face datection is no problem, but enter into facial modeling link and there occurs mistake。
When there is above-mentioned (1) situation, it was shown that Face datection and key point location are all correct, it is possible to carry out the links such as subsequent characteristics extraction, face alignment；When there is (2nd) kind situation, representing that Face datection there occurs serious flase drop, key point location is naturally also not reliable, and this situation generally stops follow-up link at once；When (3rd) kind situation occurs, it was shown that Face datection is calibrated, but key point location has partial dot problematic, if enters follow-up link, whether having a strong impact on carryover effect according to inaccurate point judges。
As in figure 2 it is shown, include for certain two classifier training step:
(1) positive sample is generated。Select the artificial face key point sample demarcated of a collection of standard, demarcate picture for such as 10000, to every pictures centered by certain key point, extract a local feature in a zonule about, such as sift describes son, it is possible to obtain 10000 positive samples。So that grader has more robustness, can around real key point random offset 1-2 pixel, key point after skew, it is also assumed that correct key point, then extracting local feature again centered by the pixel after skew as positive sample, every pictures offsets 4 times, the positive sample of regeneration, the positive sample number so generated can increase by 40000,10000 before adding, it is possible to reach 50000 positive samples。
(2) negative sample is generated。At above-mentioned and positive sample with in the face key point sample of a collection of artificial demarcation, demarcate sample for such as 10000, random offset 5-10 pixel around real key point, key point after skew, can be used to the inaccurate point of simulator locating, then extracting local feature as negative sample centered by the pixel after skew again, every pictures offsets 5 times, so can generate 50000 negative samples。
(3) training two classification。For each key point, all according to the method generating positive sample and negative sample, generating large number of positive sample and negative sample, now can train two graders, two classification adopt SVM classifier。
A grader will be trained for each key point。If key point has N number of, then can train at most N number of two graders。But in use, it may not be necessary to all of key point all trains grader, as long as choosing M therein to compare the position having representative, as long as the speed of detection can be accelerated。
Using method simplified summary is exactly, and after face key point location algorithm completes, for certain key point, only centered by this point, need to extract the local feature the same with the training stage, then calls two graders of correspondence to judge that it belongs to positive class still negative class on earth。
In the present embodiment, it is necessary to the face key point sample of substantial amounts of manual mark accurately, multiple two classification are trained to judge the order of accuarcy that face key point positions exactly。Namely the characteristic point each needs paid close attention to trains a model。It is the location quality situation extremely accurate reflecting face key point that so multiple models integrate the judged result obtained。
Claims (3)
1. the method judging Face datection flase drop and key point positional accuracy, it is characterised in that include step:
Face datection；
The face detected carries out the location of N number of key point；
Call M two graders and judge the number of inaccurate point, wherein M≤N in M key point；
If being forbidden≤the a*M that counts, then judge that Face datection and critical point detection are all very reliable, continue follow-up link；
If being forbidden >=the b*M that counts, then judging that Face datection there occurs erroneous judgement, key point location is meaningless, stops follow-up link；
If <being forbidden to count, <b*M then judges that Face datection is reliable, but Partial key point location is forbidden, and decides whether to continue follow-up link according to the situation of location of mistake a*M；
In step (c), two classifier training steps include:
(c.1) generate positive sample, choose the K artificial face key point sample demarcated opened, to every pictures centered by certain key point, extract a local feature as positive sample in a zonule about；Random offset 1-2 pixel around real key point, the key point after skew, it is also considered as correct key point, then centered by the pixel after skew, extracts local feature as positive sample again, wherein K >=10000；
(c.2) negative sample is generated, with positive sample with in the face key point sample of a collection of artificial demarcation, random offset 5-10 pixel around real key point, key point after skew, can be used to the inaccurate point of simulator locating, then centered by the pixel after skew, extract local feature as negative sample again；
(c.3) training two classification, for each key point according to (c.1) and (c.2) method, generates large number of positive sample and negative sample, now just can train two graders。
2. the method for judgement Face datection flase drop according to claim 1 and key point positional accuracy, it is characterised in that: in step (c), a is 0.05-0.15, b is 0.35-0.45。
3. the method for judgement Face datection flase drop according to claim 1 and key point positional accuracy, it is characterised in that: in step (c.3), two classification adopt SVM classifier。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610017969.9A CN105701464A (en) | 2016-01-13 | 2016-01-13 | Method of determining face detection false detection and key point positioning accuracy |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610017969.9A CN105701464A (en) | 2016-01-13 | 2016-01-13 | Method of determining face detection false detection and key point positioning accuracy |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105701464A true CN105701464A (en) | 2016-06-22 |
Family
ID=56226202
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610017969.9A Pending CN105701464A (en) | 2016-01-13 | 2016-01-13 | Method of determining face detection false detection and key point positioning accuracy |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105701464A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679504A (en) * | 2017-10-13 | 2018-02-09 | 北京奇虎科技有限公司 | Face identification method, device, equipment and storage medium based on camera scene |
| CN108197593A (en) * | 2018-01-23 | 2018-06-22 | 深圳极视角科技有限公司 | More size face's expression recognition methods and device based on three-point positioning method |
| CN108470322A (en) * | 2018-03-09 | 2018-08-31 | 北京小米移动软件有限公司 | Handle the method, apparatus and readable storage medium storing program for executing of facial image |
| CN110298291A (en) * | 2019-06-25 | 2019-10-01 | 吉林大学 | Ox face and ox face critical point detection method based on Mask-RCNN |
| CN111368792A (en) * | 2020-03-18 | 2020-07-03 | 北京奇艺世纪科技有限公司 | Characteristic point mark injection molding type training method and device, electronic equipment and storage medium |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2004072802A2 (en) * | 2003-02-11 | 2004-08-26 | New Jersey Institute Of Technology | Face detection method and apparatus |
| CN1808465A (en) * | 2005-01-21 | 2006-07-26 | 中国科学院计算技术研究所 | Evaluation method and system for face detection system |
| CN102831388A (en) * | 2012-05-23 | 2012-12-19 | 上海交通大学 | Method and system for detecting real-time characteristic point based on expanded active shape model |
| CN102867174A (en) * | 2012-08-30 | 2013-01-09 | 中国科学技术大学 | Method and device for positioning human face features |
| CN103514441A (en) * | 2013-09-21 | 2014-01-15 | 南京信息工程大学 | Facial feature point locating tracking method based on mobile platform |
| CN103679118A (en) * | 2012-09-07 | 2014-03-26 | 汉王科技股份有限公司 | Human face in-vivo detection method and system |
| CN104899563A (en) * | 2015-05-29 | 2015-09-09 | 深圳大学 | Two-dimensional face key feature point positioning method and system |
| CN105095827A (en) * | 2014-04-18 | 2015-11-25 | 汉王科技股份有限公司 | Facial expression recognition device and facial expression recognition method |
-
2016
- 2016-01-13 CN CN201610017969.9A patent/CN105701464A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2004072802A2 (en) * | 2003-02-11 | 2004-08-26 | New Jersey Institute Of Technology | Face detection method and apparatus |
| CN1808465A (en) * | 2005-01-21 | 2006-07-26 | 中国科学院计算技术研究所 | Evaluation method and system for face detection system |
| CN102831388A (en) * | 2012-05-23 | 2012-12-19 | 上海交通大学 | Method and system for detecting real-time characteristic point based on expanded active shape model |
| CN102867174A (en) * | 2012-08-30 | 2013-01-09 | 中国科学技术大学 | Method and device for positioning human face features |
| CN103679118A (en) * | 2012-09-07 | 2014-03-26 | 汉王科技股份有限公司 | Human face in-vivo detection method and system |
| CN103514441A (en) * | 2013-09-21 | 2014-01-15 | 南京信息工程大学 | Facial feature point locating tracking method based on mobile platform |
| CN105095827A (en) * | 2014-04-18 | 2015-11-25 | 汉王科技股份有限公司 | Facial expression recognition device and facial expression recognition method |
| CN104899563A (en) * | 2015-05-29 | 2015-09-09 | 深圳大学 | Two-dimensional face key feature point positioning method and system |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107679504A (en) * | 2017-10-13 | 2018-02-09 | 北京奇虎科技有限公司 | Face identification method, device, equipment and storage medium based on camera scene |
| CN108197593A (en) * | 2018-01-23 | 2018-06-22 | 深圳极视角科技有限公司 | More size face's expression recognition methods and device based on three-point positioning method |
| CN108197593B (en) * | 2018-01-23 | 2022-02-18 | 深圳极视角科技有限公司 | Multi-size facial expression recognition method and device based on three-point positioning method |
| CN108470322A (en) * | 2018-03-09 | 2018-08-31 | 北京小米移动软件有限公司 | Handle the method, apparatus and readable storage medium storing program for executing of facial image |
| CN110298291A (en) * | 2019-06-25 | 2019-10-01 | 吉林大学 | Ox face and ox face critical point detection method based on Mask-RCNN |
| CN111368792A (en) * | 2020-03-18 | 2020-07-03 | 北京奇艺世纪科技有限公司 | Characteristic point mark injection molding type training method and device, electronic equipment and storage medium |
| CN111368792B (en) * | 2020-03-18 | 2024-05-14 | 北京奇艺世纪科技有限公司 | Feature point labeling model training method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Qiao et al. | LGPMA: complicated table structure recognition with local and global pyramid mask alignment | |
| CN105701464A (en) | Method of determining face detection false detection and key point positioning accuracy | |
| CN111460962B (en) | Face recognition method and face recognition system for mask | |
| CN104298982B (en) | A kind of character recognition method and device | |
| CN109325538B (en) | Object detection method, device and computer-readable storage medium | |
| CN103049750B (en) | Character identifying method | |
| CN103473571B (en) | Human detection method | |
| CN106874894A (en) | A kind of human body target detection method based on the full convolutional neural networks in region | |
| CN109063625A (en) | A kind of face critical point detection method based on cascade deep network | |
| JP2017531262A (en) | Intelligent scoring method and system for descriptive problems | |
| CN106022317A (en) | Face identification method and apparatus | |
| CN107358223A (en) | A kind of Face datection and face alignment method based on yolo | |
| CN103778409A (en) | Human face identification method based on human face characteristic data mining and device | |
| CN105426870A (en) | Face key point positioning method and device | |
| CN102043953A (en) | Real-time-robust pedestrian detection method aiming at specific scene | |
| CN105426882B (en) | The method of human eye is quickly positioned in a kind of facial image | |
| EP3611690A1 (en) | Recognition device, recognition method, and recognition program | |
| CN109858372A (en) | A kind of lane class precision automatic Pilot structured data analysis method | |
| CN106778796A (en) | Human motion recognition method and system based on hybrid cooperative model training | |
| CN105138983B (en) | The pedestrian detection method divided based on weighting block model and selective search | |
| CN103473564A (en) | Front human face detection method based on sensitive area | |
| CN109522838A (en) | A kind of safety cap image recognition algorithm based on width study | |
| CN110796101A (en) | Face recognition method and system of embedded platform | |
| CN103544478A (en) | All-dimensional face detection method and system | |
| CN104700620A (en) | Traffic checkpoint-based method and device for recognizing fake-licensed vehicles |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20191220 |
|
| AD01 | Patent right deemed abandoned |