CN103824055B - A kind of face identification method based on cascade neural network - Google Patents
A kind of face identification method based on cascade neural network Download PDFInfo
- Publication number
- CN103824055B CN103824055B CN201410053866.9A CN201410053866A CN103824055B CN 103824055 B CN103824055 B CN 103824055B CN 201410053866 A CN201410053866 A CN 201410053866A CN 103824055 B CN103824055 B CN 103824055B
- Authority
- CN
- China
- Prior art keywords
- picture
- face
- pair
- pictures
- neural network
- 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.)
- Active
Links
Abstract
The present invention relates to a kind of face identification method based on cascade neural network, its step includes:A) training face image set is established;B) key point of every face picture is detected;C) picture is normalized into the gray value picture of standard using key point information;D) face normal pictures are subjected to gray scale Data-Statistics normalization;E) deep neural network is trained, it is a pair of pictures that it, which is inputted, and the 0 of output represents to judge that this two pictures is that negative example is right, and 1 represents positive example pair;F) by all pictures to inputting neutral net so that all erroneous judgements are less than a given ratio into 1 negative example to shared ratio;G) it is repeated multiple times to use step e), f), terminate when overall network number reaches predetermined number;H) appoint the facial image new to a pair, be determined as negative example pair or positive example pair using obtained cascade neural network.The present invention can effectively improve the knowledge extracting rate of recognition of face, substantially reduce misclassification rate, and the reliability of safety defense monitoring system is greatly improved.
Description
Technical field
The invention belongs to image procossing and technical field of face recognition, and in particular to a kind of people based on cascade neural network
Face recognition method.
Background technology
Recognition of face(face recognition), that is, a pair of face pictures are given, whether to judge two face pictures
Belong to same person.Nucleus module of the recognition of face as human face analysis system, its performance determine system order of accuarcy and
Reliability.Its main performance index has two, knows extracting rate(Give a pair of pictures pair for belonging to same person, system correct judgment
Probability)And misclassification rate(Give a pair of pictures pair for being not belonging to same person, system misjudgment, it is believed that they belong to same
Personal probability).
Traditional face recognition algorithms can mainly be divided into two steps:Be first for any one face picture,
A characteristic vector is extracted using certain algorithm be used as the face picture by face critical point detection, after face alignment
Represent, then for inputting two face pictures to be compared, using caused by some machine learning or expert design phase
The similarity of characteristic vector corresponding to calculating two pictures like degree function, if what the similarity was previously set higher than some
Threshold value then judges that this comes from same person to picture, otherwise is judged as it not being same people.Wherein threshold value is usually in a checking
Experimental debugging obtains on the data set of scene.Conventional method has two primary limitations, and 1)The characteristic vector pickup of face picture is calculated
Method is usually engineer or is highly dependent on expertise, it is difficult to which automation produces a large amount of differentiation from sample learning
Characteristic vector used for follow-up similarity calculation system, 2)No matter which kind of method is used, single similarity calculation system is difficult to
Relatively good balance is being obtained between knowledge extracting rate and misclassification rate, practical application scene often requires that misclassification rate is very low(Less than thousand
/ mono- or a ten thousandth)In the case of know extracting rate and try one's best height, single similarity system is often difficult to accurate meet two
Balance between index.
The content of the invention
The present invention is carried in view of the above-mentioned problems, a kind of face identification method based on cascade neural network of proposition using cascade
The thinking risen reduces misclassification rate step by step, and per one-level neutral net for the different sample space of difficulty, so that adaptively
The face picture for learning differentiation represents to adapt to the task at this grade of networking, and the knowledge that can effectively improve recognition of face goes out
Rate, while misclassification rate is substantially reduced, greatly improve the reliability of safety defense monitoring system.
The technical solution adopted by the present invention is as follows:
A kind of face identification method based on cascade neural network, its step include:
A) training face image set is established, it includes the picture of multiple different peoples, and wherein everyone face picture has
Multiple, these pictures may finally produce the picture pair of a large amount of same people(It is hereinafter positive example pair)With the picture of different people
It is right(It is hereinafter right for negative example);
B) using the key point of every face picture in detection a) of any one face critical point detection method;
C) using the key point positional information obtained in b), the picture in a) is normalized into the gray value figure of normal size
Piece;
D) the face normal pictures obtained in c) are subjected to gray scale Data-Statistics normalization;
E) all one deep neural networks of training picture training are used, it is a pair of pictures that it, which is inputted, and target output is
One 0-1 value, 0 represents to judge that this two pictures is that negative example is right, and 1 represents to be judged as positive example pair;
F) neutral net for drawing all training pictures to training in being input to e), passes through last layer of debugging network
Threshold value cause all erroneous judgements to be less than a given ratio to shared ratio into 1 negative example;
G) repeated multiple times using e), f) in step, unique difference is only to select in upper level network to be judged mistake every time
Negative example to being added to training set, positive example is to still using all samples as in first order network, when overall network number
Terminate when reaching predetermined number, the cascade neural network trained;
H) appoint the facial image new to a pair, use b-d) in the same pre-treatment step become alignment and normalizing in a pair
The picture pair of change, the cascade neural network trained in then inputting in order g), if in any one middle neutral net,
The picture is right to being judged to be broken into negative example, then system finally judges that the face picture is otherwise positive example pair to right for negative example.
Further, step c) carries out picture normalization using the face alignment algorithm based on affine transformation of standard.
Further, when step d) carries out the statistics normalization, pixel average 0, variance 1.
Further, two pictures of the pair of pictures of step e) from same person, or it is each from two people
One pictures.
Further, all it is made up of inside every one-level deep neural network four parts:Convolutional layer, maximum sample level, entirely
Articulamentum and prediction interval Pre, do the convolution taken turns more and sampling to input picture successively, eventually pass through the ratio that prediction interval exports 0 or 1
To result.
Further, the one given ratios of step f) are r, and r is between 0.1-0.5.
Further, terminate in step g) when overall network number reaches predetermined K, wherein K=3~5.
Relative to conventional method, contribution of the invention and beneficial effect are:
1)It is proposed a kind of multilayer cascading neural network architecture for high reliability application scenarios demand.In the structure shown here,
Face picture to be compared is to handled by the heterogeneous networks according to its complexity quilt " as thick to essence ".It is high for required precision
Face recognition application scene, single model will reach extremely low misclassification rate(Such as below one thousandth or a ten thousandth)It certainly will want
Reduce and know extracting rate, " can be divided and rule ", be absorbed in per primary network station before detecting by the way of the work of multiple series networks
The negative example that layer network cannot be distinguished by is right, and " passing through " ratio of negative example pair is successively reduced to below given threshold value, so as to carry significantly
The high performance of total system;
2)Propose a kind of algorithm of adaptive learning neutral net.This method can each nerve net in change system automatically
The sample distribution that network uses when learning(Negative example is to being increasingly difficult to)So as to select the network model for being adapted to the distribution, so as to
The adaptive limitation for generating multiple neutral nets, breaching traditional single network.
Brief description of the drawings
Fig. 1 is the step flow chart of the face identification method based on cascade neural network of the present invention;
Fig. 2 is individual depths neutral net schematic diagram in embodiment.
Embodiment
Below by specific embodiments and the drawings, the present invention will be further described.
The face identification method based on cascade neural network of the present invention, its idiographic flow is as shown in figure 1, to it specifically
It is bright as follows:
A) training face image set A is established, it includes N number of difference(Identity)People picture, wherein everyone people
Face picture has M(N can be 10000 or so, M50 or so in practice), these pictures may finally produce M* (M-1)/2*N
The picture pair of individual same people(It is hereinafter positive example pair), the picture pair of the different people of N* (N-1)/2*M*M(It is hereinafter negative
Example is right);
B) using the key point of every face picture in detection a) of any one face critical point detection algorithm;
This step can be detected using any one existing face critical point detection method.Face key point at present
Detection algorithm can be largely classified into two classes:The first kind is using each key point as independent part, and each key point is according to it
Local feature individually trains detector;Second class puts all key points training together, and emphasis considers the phase between key point
To position relationship, a globally optimal solution is finally obtained.
C) the face alignment algorithm based on affine transformation of standard is used, using the key point positional information obtained in b),
Picture in a) is normalized into 60 × 60 gray value picture;
D) the face normal pictures obtained in c) are subjected to statistics normalization so that pixel average 0, variance 1;
E) all one deep neural networks of training picture training are used, its input is a pair of pictures(It may be from same
Two pictures of one people, it is also possible to each pictures from two people), target output is a 0-1 value, and 0 represents to judge
This two pictures is that negative example is right, and 1 represents to be judged as positive example pair;
As shown in Fig. 2 all it is made up of inside per one-level deep neural network four parts:Convolutional layer(Con), maximum sampling
Layer(Mp), full articulamentum(Fuck)And prediction interval(Pre).According to sequential organization as shown in Figure 2, input picture is done successively more
The convolution of wheel and sampling, eventually pass through the comparison result that prediction interval exports 0 or 1.
F) neutral net for drawing all training pictures to training in being input to e), passes through last layer of debugging network
Threshold value cause all erroneous judgements to be less than a given ratio to shared ratio into 1 negative example(Such as 0.1), such our institutes
The negative example having is to just divide into two parts:It is right by the judicious negative example of network in e)(Account for 90%)With the negative example for being judged mistake
It is right(Account for 10%);
G) repeated multiple times using e), f) in step, unique difference is only to select in upper level network to be judged mistake every time
Negative example to being added to training set, positive example is to still using all samples as in first order network, when overall network number
Terminate when reaching predetermined K(K=3~5), so as to the cascade neural network trained;
H) threshold value caused by the debugging of last layer in K obtained deep neural network and respective step f) will be trained to be stored in
File, as the algorithm parameter trained;
I) appoint the facial image new to a pair, use b-d) in the same pre-treatment step become in a pair 60 × 60 alignment
And normalized picture is to K neutral net being trained in inputting in order h), if in any one middle nerve net
Network, the picture is to being judged to be broken into 0(Negative example is right), then system finally judge that the face picture is otherwise positive example pair to right for negative example.
The above method proposed by the present invention is directed to low misclassification rate, and height knows extracting rate requirement and devises a kind of cascade neural network knot
Structure.In the structure shown here, face picture to be compared is to handled by the heterogeneous networks according to its complexity quilt " as thick to essence ".It is more
Individual series network work, sample space " is divided and rule ", is absorbed in what layer network before detecting cannot be distinguished by per primary network station
Negative example is right, " passing through " ratio of negative example pair is successively reduced to below given threshold value, so as to substantially increase the property of total system
Energy.
In the above-mentioned methods, the present invention proposes a kind of algorithm of adaptive learning neutral net, in the algorithm simultaneously
The sample distribution used during each neural network learning can change automatically, so as to which algorithm can be trained adaptively properly
Network model, breach the functional limitation of traditional single network.
Based on above reason, the present invention is effectively improved the performance of face identification system.As shown in table 1, give test/
The ratio of picture sample is trained, the present invention will be far below conventional method in the core system index of recognition of face.
The Experimental comparison results of the present invention of table 1. and conventional method
Traditional single network method | The inventive method | |
Misclassification rate | 1.3% | 0.1% |
Know extracting rate | 54.8% | 62.2% |
The above embodiments are merely illustrative of the technical solutions of the present invention rather than is limited, the ordinary skill of this area
Technical scheme can be modified by personnel or equivalent substitution, without departing from the spirit and scope of the present invention, this
The protection domain of invention should be to be defined described in claim.
Claims (7)
1. a kind of face identification method based on cascade neural network, its step include:
A) training face image set is established, it includes the picture of multiple different peoples, and wherein everyone face picture has multiple,
By the picture of caused a large amount of same people to being referred to as positive example pair, the picture of different people is right to being referred to as negative example;
B) using the key point of every face picture in detection a) of face critical point detection method;
C) using the key point positional information obtained in b), the picture in a) is normalized into the gray value picture of normal size;
D) the face normal pictures obtained in c) are subjected to gray scale Data-Statistics normalization;
E) all deep neural networks of training picture training one are used, it is a pair of pictures that it, which is inputted, and target output is a 0-1
Value, 0 represents to judge that this two pictures is that negative example is right, and 1 represents to be judged as positive example pair;
F) neutral net for drawing all training pictures to training in being input to e), passes through the threshold of last layer of debugging network
Value causes all erroneous judgements to be less than a given ratio to shared ratio into 1 negative example;
G) repeated multiple times using e), f) in step, unique difference is only to select in upper level network to be judged mistake every time
Negative example is to being added to training set, and positive example is to still using all samples as in first order network, when overall network number reaches
Terminate during to predetermined number, the cascade neural network trained is each to cause in the cascading neural network architecture
Level network, which is absorbed in, detects that the negative example that preceding layer network cannot be distinguished by is right, so that by face picture to be compared to according to its difficulty or ease
Degree is by as handled by the thick heterogeneous networks to essence;
H) appoint the facial image new to a pair, use b)-d) in step be processed into a pair of alignment and normalized picture pair,
Then the cascade neural network trained in inputting in order g), if in middle any one neutral net picture to being judged to
It is right to be broken into negative example, then system finally judges that the face picture is otherwise positive example pair to right for negative example.
2. the method as described in claim 1, it is characterised in that:Step c) is alignd using the face based on affine transformation of standard
Algorithm carries out picture normalization.
3. the method as described in claim 1, it is characterised in that:When step d) carries out the statistics normalization, pixel average
For 0, variance 1.
4. the method as described in claim 1, it is characterised in that:Two figures of the pair of pictures of step e) from same person
Piece, or each pictures from two people.
5. the method as described in claim 1, it is characterised in that:All it is by four part groups inside per one-level deep neural network
Into:Convolutional layer, maximum sample level, full articulamentum and prediction interval Pre, the convolution taken turns more and sampling are done to input picture successively, most
0 or 1 comparison result is exported by prediction interval eventually.
6. the method as described in claim 1, it is characterised in that:Ratio given step f) described one is 0.1~0.5.
7. the method as described in claim 1, it is characterised in that:Terminate in step g) when overall network number reaches predetermined K,
Wherein K=3~5.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410053866.9A CN103824055B (en) | 2014-02-17 | 2014-02-17 | A kind of face identification method based on cascade neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410053866.9A CN103824055B (en) | 2014-02-17 | 2014-02-17 | A kind of face identification method based on cascade neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103824055A CN103824055A (en) | 2014-05-28 |
CN103824055B true CN103824055B (en) | 2018-03-02 |
Family
ID=50759107
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410053866.9A Active CN103824055B (en) | 2014-02-17 | 2014-02-17 | A kind of face identification method based on cascade neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103824055B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11853885B2 (en) | 2015-01-28 | 2023-12-26 | Google Llc | Image classification using batch normalization layers |
Families Citing this family (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015180101A1 (en) * | 2014-05-29 | 2015-12-03 | Beijing Kuangshi Technology Co., Ltd. | Compact face representation |
JP6345276B2 (en) * | 2014-06-16 | 2018-06-20 | ベイジン センスタイム テクノロジー デベロップメント カンパニー, リミテッド | Face authentication method and system |
US9715642B2 (en) | 2014-08-29 | 2017-07-25 | Google Inc. | Processing images using deep neural networks |
EP3158498A4 (en) * | 2014-11-15 | 2017-08-16 | Beijing Kuangshi Technology Co. Ltd. | Face detection using machine learning |
CN110874571B (en) * | 2015-01-19 | 2023-05-05 | 创新先进技术有限公司 | Training method and device of face recognition model |
CN106548127B (en) * | 2015-09-18 | 2022-11-04 | 松下电器(美国)知识产权公司 | Image recognition method |
US9953217B2 (en) * | 2015-11-30 | 2018-04-24 | International Business Machines Corporation | System and method for pose-aware feature learning |
CN105631406B (en) * | 2015-12-18 | 2020-07-10 | 北京小米移动软件有限公司 | Image recognition processing method and device |
CN105354565A (en) * | 2015-12-23 | 2016-02-24 | 北京市商汤科技开发有限公司 | Full convolution network based facial feature positioning and distinguishing method and system |
CN106548179B (en) * | 2016-09-29 | 2019-09-17 | 北京市商汤科技开发有限公司 | The detection method of object and dress ornament key point, device and electronic equipment |
CN108009560B (en) * | 2016-11-02 | 2021-05-11 | 广州图普网络科技有限公司 | Commodity image similarity category judgment method and device |
CN106599926A (en) * | 2016-12-20 | 2017-04-26 | 上海寒武纪信息科技有限公司 | Expression picture pushing method and system |
CN106909882A (en) * | 2017-01-16 | 2017-06-30 | 广东工业大学 | A kind of face identification system and method for being applied to security robot |
CN106951867B (en) * | 2017-03-22 | 2019-08-23 | 成都擎天树科技有限公司 | Face identification method, device, system and equipment based on convolutional neural networks |
CN108875502B (en) * | 2017-11-07 | 2021-11-16 | 北京旷视科技有限公司 | Face recognition method and device |
CN110163053B (en) | 2018-08-02 | 2021-07-13 | 腾讯科技(深圳)有限公司 | Method and device for generating negative sample for face recognition and computer equipment |
CN109543663B (en) * | 2018-12-28 | 2021-04-27 | 北京旷视科技有限公司 | Method, device and system for identifying identity of dog and storage medium |
CN110110673B (en) * | 2019-05-10 | 2020-11-27 | 杭州电子科技大学 | Face recognition method based on bidirectional 2DPCA and cascade forward neural network |
CN110414587A (en) * | 2019-07-23 | 2019-11-05 | 南京邮电大学 | Depth convolutional neural networks training method and system based on progressive learning |
CN110647948B (en) * | 2019-10-08 | 2023-04-21 | 南京大学 | Picture stitching detection method and system based on neural network |
CN113128263A (en) * | 2019-12-30 | 2021-07-16 | 深圳云天励飞技术有限公司 | Face recognition method and device |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101070442B1 (en) * | 2010-05-03 | 2011-10-05 | 주식회사 크라스아이디 | Face recognition system and method using multi-level face recognition |
CN102831396A (en) * | 2012-07-23 | 2012-12-19 | 常州蓝城信息科技有限公司 | Computer face recognition method |
-
2014
- 2014-02-17 CN CN201410053866.9A patent/CN103824055B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101070442B1 (en) * | 2010-05-03 | 2011-10-05 | 주식회사 크라스아이디 | Face recognition system and method using multi-level face recognition |
CN102831396A (en) * | 2012-07-23 | 2012-12-19 | 常州蓝城信息科技有限公司 | Computer face recognition method |
Non-Patent Citations (2)
Title |
---|
"基于旋转不变局部相位量化特征的人脸确认算法研究";高志升 等;《计算机应用研究》;20120131;第29卷(第1期);摘要,第353页第1栏第1段,第354页第3节,图1-2 * |
"基于级联神经网络的人脸检测方法的研究";陈泽宇,等;《红外与毫米波学报》;20000229;第19卷(第01期);第59页第3.1节,第60页第3.3节,第2.4节第1段 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11853885B2 (en) | 2015-01-28 | 2023-12-26 | Google Llc | Image classification using batch normalization layers |
Also Published As
Publication number | Publication date |
---|---|
CN103824055A (en) | 2014-05-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103824055B (en) | A kind of face identification method based on cascade neural network | |
CN108710892B (en) | Cooperative immune defense method for multiple anti-picture attacks | |
US10223582B2 (en) | Gait recognition method based on deep learning | |
CN112163638B (en) | Method, device, equipment and medium for defending image classification model back door attack | |
CN102890776B (en) | The method that expression figure explanation is transferred by facial expression | |
CN110610129A (en) | Deep learning face recognition system and method based on self-attention mechanism | |
CN107528823A (en) | A kind of network anomaly detection method based on improved K Means clustering algorithms | |
CN110851835A (en) | Image model detection method and device, electronic equipment and storage medium | |
CN105760835A (en) | Gait segmentation and gait recognition integrated method based on deep learning | |
CN106415594A (en) | A method and a system for face verification | |
CN106022317A (en) | Face identification method and apparatus | |
CN109740689B (en) | Method and system for screening error labeling data of image semantic segmentation | |
CN106919921B (en) | Gait recognition method and system combining subspace learning and tensor neural network | |
CN111783505A (en) | Method and device for identifying forged faces and computer-readable storage medium | |
CN108520215B (en) | Single-sample face recognition method based on multi-scale joint feature encoder | |
WO2020093303A1 (en) | Processing method and apparatus based on facial recognition, and device and readable storage medium | |
CN109387521A (en) | Image processing system | |
CN112632535A (en) | Attack detection method and device, electronic equipment and storage medium | |
Hou et al. | Mahalanobis distance based adversarial network for anomaly detection | |
CN109815946A (en) | Multithreading business license positioning identifying method based on intensive connection network | |
CN113076980A (en) | Out-of-distribution image detection method based on attention enhancement and input disturbance | |
CN109101984A (en) | A kind of image-recognizing method and device based on convolutional neural networks | |
CN108537206B (en) | Face verification method based on convolutional neural network | |
CN114445898B (en) | Face living body detection method, device, equipment, storage medium and program product | |
CN116366303A (en) | Network anomaly detection method, device, equipment and medium based on deep learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |