CN107203752A - A kind of combined depth study and the face identification method of the norm constraint of feature two - Google Patents
A kind of combined depth study and the face identification method of the norm constraint of feature two Download PDFInfo
- Publication number
- CN107203752A CN107203752A CN201710377236.0A CN201710377236A CN107203752A CN 107203752 A CN107203752 A CN 107203752A CN 201710377236 A CN201710377236 A CN 201710377236A CN 107203752 A CN107203752 A CN 107203752A
- Authority
- CN
- China
- Prior art keywords
- layer
- face
- feature
- identification method
- mrow
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of study of combined depth and the face identification method of the norm constraint of feature two, comprise the following steps:Sample collection;Sample preprocessing;Model training is carried out using residual error convolutional neural networks, network structure is:PReLU activation primitives are used after every layer of convolutional layer, pond layer uses and connects a series of residual error Internets behind maxpooling, each pond layer;Connect two norm constraints layer and normalization layer respectively after last layer of network is connected entirely;Softmax graders are connected after normalization layer;Face is identified, and exports the classification information of face in real time.The inventive method has two very important advantages:One be equal to the attention rate of all samples;Two be in normalization space, the distance between sample characteristics by forcing identical category closer to and it is different classes of between the distance between sample characteristics farther classification capacity increased with this.
Description
Technical field
The present invention relates to computer vision technique, machine learning, deep learning, convolutional neural networks and artificial intelligence,
More particularly to a kind of combined depth study and the face identification method of the norm constraint of feature two.
Background technology
With the further investigation and more exploitations of artificial intelligence technology, except what is played Weiqi, sound is distinguished, people is distinguished
Face has been occurred in that.Recognition of face is field relatively more fiery before a project, and when big data is combined with recognition of face, possibility is a lot
Loss is just avoided that.This concept of recognition of face is always the focus that global technology Trends are discussed in recent years, at home
Also increasing to use, such as brush face is checked card, and brush face APP etc. is the embodiment of this technology in fact.Once just there is scholar to refer to
Go out, following most Utopian artificial intelligence should be realized by face recognition technology, and these, lifted to a certain extent
The living standard of the mankind.In fact, there are many industries at present in experiment artificial intelligence brush face.In terms of traffic, West Beijing visitor
Stand etc. exemplified by other national key cities, " the brush face " for realizing recognition of face this year on a large scale first enters the station, remove tradition logical
Road, the unmanned checking gate of the newly-increased multi-section that enters the station, passes through resident's China second-generation identity card verification, you can completion, which is entered the station, waits.In authenticating party
Face, domestic emerging artificial intelligence company is proposed the technology of photo array on face and identity card.By to status of taxpayers
Card and figure information are acquired, compared, and effectively can go out some details instead of eye recognition.
Existing face recognition algorithms to human face expression, posture, block and the robustness such as light is not high, be also exactly serious
Dependent on training data.It would therefore be desirable to design a kind of efficient, high-accuracy, low face for relying on training data
Recognizer.
The content of the invention
In order to overcome the disadvantages mentioned above of prior art, the present invention proposes a kind of combined depth study and the norm of feature two about
The face identification method of beam, it is intended to improve the robustness of recognition of face, dependence of the reduction human face recognition model to training data.
The technical solution adopted for the present invention to solve the technical problems is:A kind of combined depth study and the norm of feature two are about
The face identification method of beam, comprises the following steps:
Step 1: sample collection:The face database for only including human face region is collected, and marks the mark of each class face
Label;
Step 2: sample preprocessing:After facial image size normalization, illumination is done using histogram equalization and located in advance
Reason;Then image whitening processing is done;
Step 3: carrying out model training using residual error convolutional neural networks, network structure is:Used after every layer of convolutional layer
PReLU activation primitives, pond layer uses and connects a series of residual error Internets behind maxpooling, each pond layer;In network
Last layer full connection after connect respectively two norm constraints layer and normalization layer;Softmax points of the connection after normalization layer
Class device;
Step 4: face is identified, and the classification information of face is exported in real time.
Compared with prior art, the positive effect of the present invention is:The inventive method has two very important advantages:One
It is equal to the attention rate of all samples to be;Two be in normalization space, by the sample characteristics for forcing identical category
The distance between closer to and it is different classes of between the distance between sample characteristics farther classification capacity increased with this.Using this
Inventive method can improve recognition of face to human face expression, posture, block and light etc. robustness.
Brief description of the drawings
Examples of the present invention will be described by way of reference to the accompanying drawings, wherein:
Fig. 1 is recognition of face network structure;
Fig. 2 is residual error Internet;
Fig. 3 is the structure chart of two norm constraints layer and normalization layer;
Fig. 4 is face recognition algorithms model training flow chart.
Fig. 5 is face prediction flow chart.
Embodiment
A kind of combined depth study and the face identification method of the norm constraint of feature two, including following content:
First, facial image is pre-processed:For pre-processing facial image, including following three part:
(1) image size normalization, by all image scalings to same size;
(2) illumination pretreatment, reduces the influence of illumination:Using histogram equalization, first by tri- channel separations of RGB,
Then the image of three passages is equalized respectively, finally again by the triple channel image co-registration after equalization.
(3) image whitening processing, the distribution of uniform data:The pixel value of tri- passages of RGB is all subtracted 127.5, then
Again divided by 128.
2nd, human face recognition model is designed:For the model for the deep neural network for designing identification face, including following step
Suddenly:
(1) residual error network structure is designed
It is exactly that on the feedforward convolutional network of standard, plus one is jumped and bypasses the connection of some layers.Often bypass one layer of just production
The residual error of a raw residual error structure (residual block), convolutional layer prediction plus input tensor, as shown in Figure 2, so may be used
To accelerate the convergence of network, over-fitting is prevented.
(2) two norm constraint structures are designed
The formula of two norm constraints layer is as follows:
Wherein, X is the output of last layer of full articulamentum.
The output vector of two norm constraints layer is a unit vector, can so ensure all features in a fixation
Sphere space in, increase feature ga s safety degree.
(3) design normalization structure sheaf
The factor alpha for normalizing layer can be a variable, constantly update in the training process or a constant.
If α is a constant, α minimum value is:
Wherein C is the categorical measure of the database of training, and p is the average softmax probability of correctly one feature of classification, is led to
Often take 0.9.
(4) object function of planned network
Object function is as follows:
Wherein, xiFor an image in input minimum batch, M is minimum batch size, yiFor xiLabel, f
(xi) be network layer second from the bottom characteristic vector, C is the quantity of classification, and W is the mapping coefficient of respective layer, b for it is corresponding partially
Shifting value, γ is a weight parameter.
(5) combine each mixed-media network modules mixed-media, constitute human face recognition model
Face recognition algorithms are used for recognition of face using a residual error convolutional neural networks, and structure is as shown in Figure 1;Every layer
PReLU activation primitives are used after convolutional layer, pond layer uses and connects a series of residual errors behind maxpooling, each pond layer
Internet (as shown in Figure 2);Connect two norm constraints layer and normalization layer respectively after last layer of network is connected entirely (such as
Shown in accompanying drawing 3);Softmax graders are connected after normalization layer.
3rd, human face recognition model is trained
(1) all data are divided into some groups of data, every group includes 128 face samples;
(2) maximum frequency of training or minimal error are set;
(3) one group of data is fully entered in the network of whole recognition of face (such as in accompanying drawing 1), obtains corresponding
Error;
(4) error is returned by network one-level one-level again, while updating network parameter;
(5) (three) and (four) step are repeated until reaching maximum frequency of training or minimum training error.
4th, face is predicted:For recognition of face, the classification information of face is exported.
The workflow of this system is illustrated below according to the mode of operation of system.
(1) training mode of the invention
Train flow as shown in Figure 4, specific steps are described as follows:
(1) sample collection
Face database is collected, human face region is only included;The label of each class face of handmarking, if any N class people, is then marked
Sign and arrive N-1 for 0.
(2) sample preprocessing
Facial image size normalization;Illumination pretreatment is done using histogram equalization;And whitening processing is done, by RGB tri-
The pixel value of individual passage all subtracts 127.5, then again divided by 128.(3) model training
Embodiment 1
The instruction of human face recognition model is illustrated exemplified by training a face recognition algorithms model comprising 10000 people
Practice process.
(1) human face data for including 10000 people, the label of artificial demarcation face, from 0 to 9999 are collected.
(2) sample preprocessing:
1) facial image size normalization;
2) histogram equalization
3) image whitening processing
(3) model training;
(2) predictive mode of the invention
Flow chart as shown in Figure 5, gives a facial image, size normalization processing, illumination is done to this facial image
Pretreatment and whitening processing, are then input in whole identification network, network obtains 1x10000 vector, finds out this
Position in vector where maximum is the classification belonging to the face.
Embodiment 2
(1) model is deployed on the PC of windows systems;
(2) source of input picture is done with camera;
(3) program can export the classification information of face in real time;
In summary, the present invention trains human face recognition model to realize by combined depth study and the norm constraint of feature two
A kind of high efficiency and high-precision real-time face know algorithm.
Claims (9)
1. a kind of combined depth study and the face identification method of the norm constraint of feature two, it is characterised in that:Comprise the following steps:
Step 1: sample collection:The face database for only including human face region is collected, and marks the label of each class face;
Step 2: sample preprocessing:After facial image size normalization, illumination pretreatment is done using histogram equalization;So
After do image whitening processing;
Step 3: carrying out model training using residual error convolutional neural networks, network structure is:Used after every layer of convolutional layer
PReLU activation primitives, pond layer uses and connects a series of residual error Internets behind maxpooling, each pond layer;In network
Last layer full connection after connect respectively two norm constraints layer and normalization layer;Softmax points of the connection after normalization layer
Class device;
Step 4: face is identified, and the classification information of face is exported in real time.
2. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The step of carrying out model training using residual error convolutional neural networks described in step 3 is as follows:
(1) all data are divided into some groups of data, every group includes 128 face samples;
(2) maximum frequency of training or minimum training error are set;
(3) one group of data is fully entered in residual error convolutional neural networks, obtains corresponding error;
(4) error is returned by network one-level one-level again, while updating network parameter;
(5) (three) and (four) step are repeated until reaching maximum frequency of training or minimum training error.
3. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The structure of the residual error Internet is:On the feedforward convolutional network of standard, plus a jump bypasses the company of some layers
Connect, often bypass one layer and produce a residual error structure.
4. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The output vector of the two norm constraints layer is a unit vector.
5. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The factor alpha of the normalization layer is a variable constantly updated an in the training process either constant.
6. a kind of combined depth study according to claim 5 and the face identification method of the norm constraint of feature two, it is special
Levy and be:When the factor alpha of the normalization layer is constant, its minimum value is:
<mrow>
<msub>
<mi>a</mi>
<mrow>
<mi>l</mi>
<mi>o</mi>
<mi>w</mi>
</mrow>
</msub>
<mo>=</mo>
<mi>l</mi>
<mi>o</mi>
<mi>g</mi>
<mfrac>
<mrow>
<mi>p</mi>
<mrow>
<mo>(</mo>
<mi>C</mi>
<mo>-</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mn>1</mn>
<mo>-</mo>
<mi>p</mi>
</mrow>
</mfrac>
</mrow>
Wherein C is the categorical measure of the database of training, and p is the average softmax probability of correctly one feature of classification.
7. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The object function of the residual error convolutional neural networks is:
minimize:
subject to:
Wherein, xiFor an image in input minimum batch, M is minimum batch size, yiFor xiLabel, f (xi) be
The characteristic vector of network layer second from the bottom, C is the categorical measure of the database of training, and W is the mapping coefficient of respective layer, and b is phase
The deviant answered, γ is weight parameter.
8. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The method of illumination pretreatment described in step 2 is:Using histogram equalization, first by tri- channel separations of RGB, so
The image of three passages is equalized respectively afterwards, finally again by the triple channel image co-registration after equalization.
9. a kind of combined depth study according to claim 1 and the face identification method of the norm constraint of feature two, it is special
Levy and be:The method of step 2 described image whitening processing is:The pixel value of tri- passages of RGB is all subtracted 127.5, Ran Houzai
Divided by 128.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710377236.0A CN107203752A (en) | 2017-05-25 | 2017-05-25 | A kind of combined depth study and the face identification method of the norm constraint of feature two |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710377236.0A CN107203752A (en) | 2017-05-25 | 2017-05-25 | A kind of combined depth study and the face identification method of the norm constraint of feature two |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107203752A true CN107203752A (en) | 2017-09-26 |
Family
ID=59906047
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710377236.0A Pending CN107203752A (en) | 2017-05-25 | 2017-05-25 | A kind of combined depth study and the face identification method of the norm constraint of feature two |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107203752A (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107886062A (en) * | 2017-11-03 | 2018-04-06 | 北京达佳互联信息技术有限公司 | Image processing method, system and server |
CN108319943A (en) * | 2018-04-25 | 2018-07-24 | 北京优创新港科技股份有限公司 | A method of human face recognition model performance under the conditions of raising is worn glasses |
CN108509862A (en) * | 2018-03-09 | 2018-09-07 | 华南理工大学 | Anti- angle and the fast human face recognition for blocking interference |
CN108573243A (en) * | 2018-04-27 | 2018-09-25 | 上海敏识网络科技有限公司 | A kind of comparison method of the low quality face based on depth convolutional neural networks |
CN108734290A (en) * | 2018-05-16 | 2018-11-02 | 湖北工业大学 | It is a kind of based on the convolutional neural networks construction method of attention mechanism and application |
CN108898105A (en) * | 2018-06-29 | 2018-11-27 | 成都大学 | It is a kind of based on depth characteristic and it is sparse compression classification face identification method |
CN108961366A (en) * | 2018-06-06 | 2018-12-07 | 大连大学 | Based on convolution self-encoding encoder and manifold learning human motion edit methods |
CN110163032A (en) * | 2018-02-13 | 2019-08-23 | 浙江宇视科技有限公司 | A kind of method for detecting human face and device |
CN110781866A (en) * | 2019-11-08 | 2020-02-11 | 成都大熊猫繁育研究基地 | Panda face image gender identification method and device based on deep learning |
CN111738927A (en) * | 2020-03-23 | 2020-10-02 | 阳光暖果(北京)科技发展有限公司 | Face recognition feature enhancement and denoising method and system based on histogram equalization |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105160400A (en) * | 2015-09-08 | 2015-12-16 | 西安交通大学 | L21 norm based method for improving convolutional neural network generalization capability |
CN105894050A (en) * | 2016-06-01 | 2016-08-24 | 北京联合大学 | Multi-task learning based method for recognizing race and gender through human face image |
CN105989368A (en) * | 2015-02-13 | 2016-10-05 | 展讯通信(天津)有限公司 | Target detection method and apparatus, and mobile terminal |
-
2017
- 2017-05-25 CN CN201710377236.0A patent/CN107203752A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105989368A (en) * | 2015-02-13 | 2016-10-05 | 展讯通信(天津)有限公司 | Target detection method and apparatus, and mobile terminal |
CN105160400A (en) * | 2015-09-08 | 2015-12-16 | 西安交通大学 | L21 norm based method for improving convolutional neural network generalization capability |
CN105894050A (en) * | 2016-06-01 | 2016-08-24 | 北京联合大学 | Multi-task learning based method for recognizing race and gender through human face image |
Non-Patent Citations (3)
Title |
---|
KAIMING HE等: "Deep Residual Learning for Image Recognition", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
RAJEEV RANJAN等: "L2-constrained Softmax Loss for Discriminative Face Verification", 《HTTPS://ARXIV.ORG/ABS/1703.09507V2》 * |
张学忠: "基于KPCA和LDA融合改进的人脸识别算法研究", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107886062A (en) * | 2017-11-03 | 2018-04-06 | 北京达佳互联信息技术有限公司 | Image processing method, system and server |
CN107886062B (en) * | 2017-11-03 | 2019-05-10 | 北京达佳互联信息技术有限公司 | Image processing method, system and server |
CN110163032A (en) * | 2018-02-13 | 2019-08-23 | 浙江宇视科技有限公司 | A kind of method for detecting human face and device |
CN108509862A (en) * | 2018-03-09 | 2018-09-07 | 华南理工大学 | Anti- angle and the fast human face recognition for blocking interference |
US11417147B2 (en) | 2018-03-09 | 2022-08-16 | South China University Of Technology | Angle interference resistant and occlusion interference resistant fast face recognition method |
CN108509862B (en) * | 2018-03-09 | 2022-03-25 | 华南理工大学 | Rapid face recognition method capable of resisting angle and shielding interference |
WO2019169942A1 (en) * | 2018-03-09 | 2019-09-12 | 华南理工大学 | Anti-angle and occlusion interference fast face recognition method |
CN108319943A (en) * | 2018-04-25 | 2018-07-24 | 北京优创新港科技股份有限公司 | A method of human face recognition model performance under the conditions of raising is worn glasses |
CN108573243A (en) * | 2018-04-27 | 2018-09-25 | 上海敏识网络科技有限公司 | A kind of comparison method of the low quality face based on depth convolutional neural networks |
CN108734290B (en) * | 2018-05-16 | 2021-05-18 | 湖北工业大学 | Convolutional neural network construction method based on attention mechanism and application |
CN108734290A (en) * | 2018-05-16 | 2018-11-02 | 湖北工业大学 | It is a kind of based on the convolutional neural networks construction method of attention mechanism and application |
CN108961366A (en) * | 2018-06-06 | 2018-12-07 | 大连大学 | Based on convolution self-encoding encoder and manifold learning human motion edit methods |
CN108898105A (en) * | 2018-06-29 | 2018-11-27 | 成都大学 | It is a kind of based on depth characteristic and it is sparse compression classification face identification method |
CN110781866A (en) * | 2019-11-08 | 2020-02-11 | 成都大熊猫繁育研究基地 | Panda face image gender identification method and device based on deep learning |
CN111738927A (en) * | 2020-03-23 | 2020-10-02 | 阳光暖果(北京)科技发展有限公司 | Face recognition feature enhancement and denoising method and system based on histogram equalization |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107203752A (en) | A kind of combined depth study and the face identification method of the norm constraint of feature two | |
WO2022160771A1 (en) | Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model | |
CN107886073B (en) | Fine-grained vehicle multi-attribute identification method based on convolutional neural network | |
CN103824054B (en) | A kind of face character recognition methods based on cascade deep neural network | |
CN110046671A (en) | A kind of file classification method based on capsule network | |
CN107316015A (en) | A kind of facial expression recognition method of high accuracy based on depth space-time characteristic | |
CN104866829A (en) | Cross-age face verify method based on characteristic learning | |
CN106803069A (en) | Crowd's level of happiness recognition methods based on deep learning | |
CN105447473A (en) | PCANet-CNN-based arbitrary attitude facial expression recognition method | |
CN110717481A (en) | Method for realizing face detection by using cascaded convolutional neural network | |
CN110120041A (en) | Pavement crack image detecting method | |
CN109492750B (en) | Zero sample image classification method based on convolutional neural network and factor space | |
CN112732921B (en) | False user comment detection method and system | |
CN110321862B (en) | Pedestrian re-identification method based on compact ternary loss | |
CN106709528A (en) | Method and device of vehicle reidentification based on multiple objective function deep learning | |
CN109815864A (en) | A kind of facial image age recognition methods based on transfer learning | |
Ocquaye et al. | Dual exclusive attentive transfer for unsupervised deep convolutional domain adaptation in speech emotion recognition | |
CN109165337A (en) | A kind of method and system of knowledge based map construction bidding field association analysis | |
CN104408470A (en) | Gender detection method based on average face preliminary learning | |
CN113011386B (en) | Expression recognition method and system based on equally divided characteristic graphs | |
CN105631477A (en) | Traffic sign recognition method based on extreme learning machine and self-adaptive lifting | |
CN104463243A (en) | Sex detection method based on average face features | |
CN108363962B (en) | Face detection method and system based on multi-level feature deep learning | |
CN110321936A (en) | A method of realizing that picture two is classified based on VGG16 and SVM | |
CN107545281B (en) | Single harmful gas infrared image classification and identification method based on deep learning |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170926 |
|
RJ01 | Rejection of invention patent application after publication |