CN108921123A - A kind of face identification method based on double data enhancing - Google Patents
A kind of face identification method based on double data enhancing Download PDFInfo
- Publication number
- CN108921123A CN108921123A CN201810780758.XA CN201810780758A CN108921123A CN 108921123 A CN108921123 A CN 108921123A CN 201810780758 A CN201810780758 A CN 201810780758A CN 108921123 A CN108921123 A CN 108921123A
- Authority
- CN
- China
- Prior art keywords
- image
- sample
- face
- layer
- model
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of face identification method based on double data enhancing, the human face data collection enhanced by double data improves the accuracy rate of recognition of face.Including：S1, selection data set；S2, data set pretreatment；S3, the first tuple are according to enhancing：It builds and confrontation network model is maximumlly generated based on information, and complete to train；S4, the second tuple are according to enhancing：The face sample generated in S3 is translated, is rotated, is overturn and scaling processing；S5, face classification select convolutional neural networks model to be trained and identify the face sample generated in S4.
Description
Technical field
The invention belongs to machine learning field, in particular to a kind of face identification method based on double data enhancing.
Background technique
Supervised learning is a kind of common method in machine learning field, it refers to the number for respectively concentrating large-scale data
Accordingly and input of the corresponding label as certain mathematical model, being then trained to the model makes its study to given data
Feature, thus the process handled unknown data.According to the definition of supervised learning it is found that data set is to determine most
One of the key link of final cast performance superiority and inferiority.It increases income with the continuous development of depth learning technology and in different field number
According to the appearance of collection, supervised learning method has obtained more being widely applied.But the data set increased income at present still has one
It is a little insufficient.By taking human face data collection as an example, on the one hand there is a problem of that population sample is limited, on the other hand for individual human face sample
For, the characteristic attribute that it includes is extremely limited, while the collection of data set, arrangement, label be one expend very much the time with
The process of energy, so wanting to concentrate in a large-scale data by a series of spy of consecutive variations of all individual human face samples
Sign attribute, which is included, seems very difficult.Therefore, these problems existing for human face data collection limit to a certain extent at present
Popularization of the recognition of face based on supervised learning method under more application scenarios.
Summary of the invention
The present invention provides a kind of based on double data enhancing for some shortcomings existing for existing human face data collection
Face identification method, the human face data collection enhanced by double data, improves the accuracy rate of recognition of face.
The object of the present invention is achieved like this：
A kind of face identification method based on double data enhancing, includes the following steps：
S1, selection data set；
S2, data set pretreatment；
S3, the first tuple are according to enhancing：
It builds and confrontation network model is maximumlly generated based on information, confrontation network model is maximumlly generated based on information
Including generator G and arbiter D, generator G passes through the type and dimension of control latent variable c, to utilize random noise
Vector z is generated and the close generation sample G (z, c) of authentic specimen X probability density distribution, will generate sample G (z, c) and true sample
This X is respectively as the input of arbiter D, and arbiter D is responsible, and generation sample G (z, c) and authentic specimen X to input carry out "true"
With the judgement of "false", finally in maximumlly generating confrontation network model based on information, generator G and arbiter D are carried out not
Disconnected alternating training, generator G constantly promote the generative capacity of oneself, and arbiter D then continues to optimize the classification capacity of oneself, when
When arbiter D can not judge that some face sample comes from generation sample or authentic specimen X, which is completed；
S4, the second tuple are according to enhancing：
The face sample generated in S3 is translated, is rotated, is overturn and scaling processing；
S5, face classification
Convolutional neural networks model is selected to be trained and identify the face sample generated in S4.
Preferably, in S1, the open source human face data collection CelebA of Hong Kong Chinese University is selected, as data set.
Preferably, preprocess method is as follows in S2：The image concentrated to data is cut, and is normalized,
The pixel value of image is uniformly arrived between [- 1,1].
Preferably, it in S3, builds the method for maximumlly generating confrontation network model based on information and includes the following steps：
S31, Maker model G is built
The random noise vector for inputting generator G is divided into incompressible variable z, latent variable c, generator G utilize with
Machine noise vector generates the generation sample G (z, c) close with the authentic specimen probability density distribution of face in data set；
Information theory is introduced, indicates mutual between the output G (z, c) of latent variable c and generator G with I (c, G (c, z))
Information, mutual information between the two are as follows：
I (c, G (c, z))=H (c)-H (c | G (z, c))=H (G (z, c))-H (G (z, c) | c) (1)
As available from the above equation：When mono- timing of latent variable c, it is possible to reduce the uncertainty of output sample G (z, c), if latent
It is uncorrelated in variable c and output sample G (z, c), then I (c, G (c, z))=0；It is pre- in order to be obtained by control latent variable c
The output of phase maximizes mutual information item I (c, G (c, z)), and the objective function for being originally generated confrontation network is：
It introduces after mutual information item I (c, G (c, z)), the objective function that confrontation network is maximumlly generated based on information is：
It simultaneously in generator G, is up-sampled using warp lamination, and uses relu activation primitive, so that making an uproar
Sound gradually becomes high-resolution image；
S32, arbiter model D is built
Arbiter D is made of one four layers of convolutional neural networks, and authentic specimen X is judged as 1 as far as possible by its target,
It generates sample and is judged as 0, then, it is 0 that setting, which generates sample G (z, c), and authentic specimen X is 1, subsequently into carrying out in arbiter D
Training, finally obtains the classifier haveing excellent performance；
S33, selection optimizer
Select Adam as optimizer.
Preferably, in S31, the dimension for defining incompressible variable z is 128 dimensions, and the dimension of latent variable c is 58 dimensions, is dived
The discrete variable and 8 continuous variables tieed up in variable c by 5 10 are constituted, wherein the discrete variable of 5 10 dimensions is for controlling
The type of feature, 8 continuous variables are used for the consecutive variations of controlling feature.
Preferably, in S33, the inner parameter of Adam is set as：Learning rate lr=0.0005, β=0.5.
Preferably, in S4,
Shift method is：
The overall movement that image is carried out along X-axis or Y direction or simultaneously along X-axis, Y-axis, if certain is put to X-direction movement
Tx, the mobile ty of Y direction, (x, y) is coordinate before converting, and (X, Y) is coordinate after transformation, then the formula translated is：
Spinning solution is：
Image is using certain point as the center of circle, using the point and origin line as radius, rotates θ degree counterclockwise, if the point (x, y),
New position (X, Y), the then formula rotated are：
Method for turning is：
Image is using X-axis or Y-axis as to axis, acquired mirror image, if seat of certain point coordinate (x, y) along X-axis overturning, after conversion
Mark (X1, Y1)；It is overturn along Y-axis, the coordinate (X after conversion2, Y2), overturning expression formula is：
Zoom method is：
The scaling of image refers to that image is zoomed in or out with Y direction according to a certain percentage along the x axis, if figure
The coordinate of certain point is (x, y) as in, scales sx times in X-direction, and Y direction scales sy times, and transformed coordinate is (X, Y),
The formula then scaled is：
Preferably, in S5, convolutional neural networks model includes：
Input layer
In input layer, using the face sample image with continuous changing features generated in S4 as convolutional neural networks
Input, wherein 80% is used to train, 20% for testing；
Convolution-pond layer
The depth and width for adjusting convolutional layer make convolutional layer extract best, the different convolution kernel of facial image characteristic effect
For extracting different features；Shallow-layer convolutional layer is for extracting rudimentary semantic feature；Deep layer convolutional layer is advanced for extracting
Semantic feature selects the suitable convolution number of plies according to the complexity of image, and the new pixel exported in convolutional layer is by following
Formula is calculated：
Wherein, f () represents activation primitive,Some pixel value of one layer of characteristic image is represented,Represent convolution
Core, * represent convolution algorithm；In view of the output of this layer can be associated with upper one layer of multiple characteristic image, MjIt indicates to participate in operation
The subset of upper one layer of characteristic image；Bias term is represented, subscript l indicates l layers；
Pond layer is used to carry out dimension-reduction treatment, the basis of the eigenmatrix obtained after convolution algorithm to the feature of extraction
On, be added maximum pondization processing, pondization operate in each neuron correspond to the position each N × 1, formula in convolution and be：
Wherein, u (n, 1) is a window function of convolution operation, ajThe maximum value in correspondence image region；
SoftMax layers
SoftMax layers for being mapped as corresponding probability value, final choice probability value maximum institute for the output valve of pond layer
Result of the classification as category of model, it is assumed that input feature vector is denoted as x(i), sample label is denoted as y(i), composing training collection S=
{(x(1),y(1)),…,(x(m),y(m)), for given input x, its probability value p is estimated to each classification j using hypothesized model
(y=j | x), wherein assuming that function is：
Wherein, θ1,θ2,…,θkFor the model parameter that can learn,To normalize item, so that the sum of all probability
It is 1, to obtain cost function：
Wherein, 1 { } was an indicative function, and when the value in bracket is true, otherwise it is 0 that the result of function, which is 1,
It is described to assume that function is the popularization to logistic regression, therefore cost function is readily modified as：
Its partial derivative is asked for SoftMax cost function J (θ), obtains gradient formula：
For a vector, its first of elementIt is J (θ) to θjFirst of component partial derivative,
After obtaining the above solution partial derivative formula, cost function J (θ) is carried out using stochastic gradient descent algorithm minimum
Change, requires to be updated parameter in each iterative process：Finally realize
SoftMax returns disaggregated model.
By adopting the above-described technical solution, the present invention has the advantages that：
It has been built in the present invention and confrontation network, English abbreviation is maximumlly generated based on information：InfoGAN.Utilize the mould
Type can effectively control the advantage of the random noise vector of input, so that it is exported the face sample that we want, thus to existing
Human face data collection enhanced, effectively alleviate available data collection totality Finite Samples and individual human face sample included
Characteristic attribute it is insufficient the problems such as.
Existing condition generates confrontation network (CGAN) and is not belonging to the unsupervised model in complete meaning, because in the model
Label information c is introduced between the generating process of sample and differentiates process：1) random noise variable z and label information c are total to
With the input as generator G.Label information c is introduced, so that generator G is no longer randomly generated sample.But generate label
Information c specified sample；2) label information c has also been introduced in arbiter D training process, by raw data set and label information
C is stitched together, to improve the stability of arbiter training.Confrontation, which is maximized, the present invention is based on information generates network (i.e.
It InfoGAN) is a unsupervised learning model, it is trained using unknown sample (i.e. sample is free of label), is passed through
The type and dimension of latent variable c are controlled, and attempts to maximize latent variable c and generates the association between sample G (z, c)
Property, to achieve the purpose that control generates sample characteristics attribute change, which can generate a large amount of legacy data concentrations and be not present
Sample.
The present invention can effectively control the advantage of input using the model, it is made to export the face sample that we want, from
And to existing human face data collection carry out first weight data enhance, effectively alleviate available data collection totality Finite Samples with
And the individual human face sample characteristic attribute that is included it is insufficient the problems such as；We further locate enhanced data set simultaneously
Reason, including translation, rotation, scaling etc., realize the second tuple to human face data collection according to enhancing, the people obtained by dual enhancing
Face data set will enable the accuracy rate of the face identification method based on supervised learning further to get a promotion.By dual
Data enhancing processing, significantly increases the sample size of certain feature consecutive variations under the same face, for subsequent convolution mind
Effective, reliable data source is provided through network class model, and then improves recognition of face under continuous changing features
Accuracy rate.
Detailed description of the invention
Fig. 1 is maximumlly to generate confrontation network based on information to be used for the network frame that face characteristic generates；
Fig. 2 is the loss function curve maximumlly generated in confrontation model training process based on information；
Fig. 3 is InfoGAN-CNN mixed model general frame
Fig. 4 is InfoGAN-CNN mixed model inner parameter；
Fig. 5 is part sample in face generating process；
Fig. 6 is the part sample of generation after the completion of model training；
Fig. 7 is the facial image of the eyes size consecutive variations of synthesis；
Fig. 8 is the image of the facial orientation consecutive variations of synthesis；
Fig. 9 is the facial image of the nose size consecutive variations of synthesis；
Figure 10 is the facial image of the lip thickness consecutive variations of synthesis；
Figure 11 is the facial image of the hair style consecutive variations of synthesis；
Figure 12 is the gender variation facial image of synthesis；
Figure 13 is the facial image of the colour of skin consecutive variations of synthesis；
Figure 14 is the facial image of the color development consecutive variations of synthesis；
Figure 15 is the facial image of the mood consecutive variations of synthesis；
Figure 16 is the facial image of synthesis whether worn glasses；
Figure 17 is the facial image of the small consecutive variations of being bold of synthesis；
Figure 18 is the facial image of the beard consecutive variations of synthesis.
Specific embodiment
Referring to Fig. 1, Fig. 2, a kind of face identification method based on double data enhancing includes the following steps：
S1, the suitable data set of selection
By comparison, open source human face data collection -- the CelebA of Hong Kong Chinese University is selected in the present invention, as model
Trained data source.The data set contains more than 200,000 or so famous person's facial image and 40 face characteristic label, is
The feature that corresponding consecutive variations are generated in the present invention has established data basis.
S2, data set pretreatment
The original size of image is 178 × 218 in CelebA data set, this will carry out biggish calculating to the training band in later period
Burden, so we are cut to 32 × 32.They are normalized simultaneously, the pixel value of image is uniformly arrived [-
1,1] between.
S3, the first tuple are according to enhancing
It builds and confrontation network model (InfoGAN), model frame used in the present invention is maximumlly generated based on information
Frame is as shown in Figure 1.It includes generator G and arbiter D that confrontation network model is maximumlly generated based on information, and generator G passes through
The type (being divided into discrete variable and continuous variable) and dimension of latent variable c are controlled, to generate using random noise vector z
Generation sample ("false" sample) G (z, c) close with authentic specimen probability density distribution will generate sample G (z, c) and true sample
This ("true" sample) X respectively as the input of arbiter D, is trained it, it is made to obtain two classifiers.Two models
It carries out continuous alternately training, generator G and constantly promotes the generative capacity of oneself, the face sample of generation is more true to nature；And sentence
Other device D then continues to optimize the classification capacity of oneself, to obtain the excellent classifier of classification performance.Finally when two models
Ability it is all very strong when, i.e. arbiter D can not judge that some face sample comes from "false" data set G (z, c) still
When "true" sample set X, which is completed.
In S3, builds the method for maximumlly generating confrontation network model based on information and include the following steps：
S31, Maker model G is built
The uncontrollability of random noise vector in confrontation model (GAN) is originally generated and can not be explanatory to make up,
The noise vector of input is divided into two parts in the present invention：1) incompressible variable z；2) latent variable c.And define can not
The z-dimension of compression noise is 128 dimensions.The dimension of latent variable c is 58 dimensions, and the discrete variable tieed up including 5 10 is for controlling
The type of feature processed, 8 continuous variables are used for the consecutive variations of controlling feature.
In order to reach by effectively controlling input, thus obtain it is anticipated that image pattern purpose, i.e., so that
The correlation of height is established between latent variable c and the output G (z, c) for generating model, we introduce information theory.In information theory
In, I (X；Y the mutual information between X and Y) is represented.And I (c, G (c, z)) indicates latent variable c and generates model in the present invention
Output G (z, c) between mutual information.Mutual information between the two is as follows：
I (c, G (c, z))=H (c)-H (c | G (z, c))=H (G (z, c))-H (G (z, c) | c) (1)
As available from the above equation：I (c, G (c, z)) item represents when mono- timing of latent variable c, it is possible to reduce output sample G (z,
C) uncertainty.If latent variable c and output sample G (z, c) are uncorrelated, I (c, G (c, z))=0；In the present invention
In, it is desirable to by effectively controlling latent variable c, to obtain expected output, then need mutual information item I (c, G (c, z)) most
Bigization.Be originally generated confrontation network objective function be：
It introduces after mutual information item I (c, G (c, z)), the objective function that confrontation network is maximumlly generated based on information is：
Simultaneously in generator G, present invention uses warp laminations, are up-sampled, and used the activation of " relu "
Function, so that noise gradually becomes high-resolution image.
S32, arbiter model is built
Arbiter D is made of one four layers of convolutional neural networks, and "true" sample X is judged as 1 as far as possible by its target,
"false" sample be judged as 0. then we artificially give "false" sample G (z, c) labelled " 0 "；It is labelled to "true" sample X
" 1 ", subsequently into being trained in arbiter D.Finally obtain the classifier haveing excellent performance.
S33, the suitable optimizer of selection
Select suitable optimizer can not only training for promotion speed, also will affect the final performance of model.Our bases
In deep learning tool -- the TensorFlow of Google's open source, builds and confrontation network is maximumlly generated based on information
(InfoGAN).A variety of optimizers are integrated in TensorFlow for us：Adam,SGD,RMSprop,Adagrad,
Adadelta. by comparing, Adam optimizer has been selected in the present invention, its inner parameter is set as；Learning rate lr=
0.0005, β=0.5.
S4, the second tuple are according to enhancing：
Although some feature maximumlly generated under the confrontation available same face of network based on information is continuous
Multiple samples of variation, but since there are certain unstability in training for confrontation generation network, so the sample generated
More, the sample generated below may can deviate from original facial image, it becomes difficult to it recognizes, it at this time cannot be again by generating sample
This mode improves the accuracy rate of the recognition of face based on supervised learning.Then pass through the face sample to generation in the present invention
This such as is translated, is rotated, overturn and is scaled at processing, that is, carries out the second tuple according to enhancing, to further enhance face sample
This quantity provides reliable sample for subsequent disaggregated model, prevents model over-fitting.
S41, translation
The overall movement that image is carried out along X-axis or Y direction (or both simultaneously).If certain is put to X-direction mobile tx, Y
Axis direction moves ty, and (x, y) is coordinate before converting, and (X, Y) is coordinate after transformation.The formula then translated is：
S42, rotation
Image is using certain point as the center of circle, using the point and origin line as radius, rotates θ degree counterclockwise.If the point (x, y),
New position (X, Y), the then formula rotated are：
S43, overturning
Image is using X-axis or Y-axis as to axis, acquired mirror image.If certain point coordinate (x, y) is overturn along X-axis, the seat after conversion
Mark (X1, Y1)；It is overturn along Y-axis, the coordinate (X after conversion2, Y2).Overturning expression formula is：
S44, scaling
The scaling of image refers to that image is zoomed in or out with Y direction according to a certain percentage along the x axis.If figure
The coordinate of certain point is (x, y) as in, scales sx times in X-direction, Y direction scales sy times, and transformed coordinate is (X, Y).
The formula then scaled is：
S5, convolutional neural networks (CNN) --- face classification
In field of image recognition, convolutional neural networks show powerful competitiveness.Wherein, LeNet-5, AlexNet,
VGG, GooleNet etc. are the models in convolutional neural networks developing history with milestone significance.InfoGAN mould in the present invention
The face characteristic that type generates is consecutive variations, but these features and uncomplicated, so we have selected simple convolution mind
It is trained and identifies through network model.
Input layer
In input layer, the facial image with continuous changing features that we are 32 × 32 using the size of generation is as convolution
The input of neural network.Wherein 80% for training, and 20% for testing.
Convolution-pond layer
Compared with traditional full articulamentum, convolutional layer relies on the machine of its unique local receptor field and globally shared weight
System, can effectively reduce training parameter, the efficiency of training for promotion.The target of convolutional layer is extraction facial image feature, so
We are mainly adjusted to obtain optimal effect from the depth of convolutional layer and width.On the one hand, different convolution kernels is used
In extracting different features, so selecting appropriate number of convolution kernel particularly significant；On the other hand, shallow-layer convolutional layer is for extracting
Low-level features；Deep layer convolutional layer is suitable therefore, it is necessary to be selected according to the complexity of image for extracting advanced semantic feature
The convolution number of plies.The new pixel exported in convolutional layer can be calculated by formula (6)：
Wherein, f () represents activation primitive,Some pixel value of one layer of characteristic image is represented,Represent convolution
Core, * represent convolution algorithm；In view of the output of this layer can be associated with upper one layer of multiple characteristic image, MjIt indicates to participate in operation
The subset of upper one layer of characteristic image；Bias term is represented, subscript l indicates l layers.
The target of pond layer is not to extract face characteristic, but further carry out dimension-reduction treatment to the feature of extraction.?
On the basis of the eigenmatrix obtained after convolution algorithm, we joined maximum pondization processing, i.e., in one 2 × 2 submatrix
In, the matrix is replaced using wherein maximum value, to achieve the purpose that certain prominent feature and reduce data dimension.Chi Hua
Each neuron corresponds to the position each N × 1 in convolution in operation.Its formula is：
Wherein, u (n, 1) is a window function of convolution operation, ajThe maximum value in correspondence image region.
SoftMax
SoftMax regression model is expansion of the Logic Regression Models in more classification problems.In the present invention, we will
Characteristic attribute in CelebA data set is divided into 3 levels, and 4 facial images are selected in the feature of each level and are carried out
Test.The SoftMax layers of the last layer as convolutional neural networks model, its effect are to be mapped as the output valve of preceding layer
Corresponding probability value, result of the classification as category of model where final choice probability value is maximum.Assuming that input feature vector is denoted as x(i), sample label is denoted as y(i)(y(i)Become 0,1,2 three classes after vector coding), thus constitute training set S={ (x(1),y(1)),…,(x(m),y(m))}.For given input x, its probability value p (y is estimated to each classification j using hypothesized model
=j | x), wherein assuming that function is：
Wherein, θ1,θ2,…,θkFor the model parameter that can learn,To normalize item, so that the sum of all probability
It is 1, to obtain cost function
Wherein, 1 { } was an indicative function, and when the value in bracket is true, otherwise it is 0 that the result of function, which is 1,.
Formula (8) is the popularization to logistic regression, therefore cost function is readily modified as：
Its partial derivative is asked for SoftMax cost function J (θ), obtains gradient formula：
For a vector, its first of elementIt is J (θ) to θjFirst of component partial derivative.
After obtaining the above solution partial derivative formula, cost function J (θ) is carried out using stochastic gradient descent algorithm minimum
Change.It requires to be updated parameter in each iterative process：Finally realize
SoftMax returns disaggregated model.
In order to complete human face segmentation and recognition of face task, we have proposed the whole frames of InfoGAN-CNN mixed model
Frame is as shown in figure 3, detailed model inner parameter is as shown in Figure 4.
Referring to Fig. 5-Figure 18, for based on the facial image in the continuous feature generation of face for generating confrontation network.
Finally, it is stated that preferred embodiment above is only used to illustrate the technical scheme of the present invention and not to limit it, although logical
It crosses above preferred embodiment the present invention is described in detail, however, those skilled in the art should understand that, can be
Various changes are made to it in form and in details, without departing from claims of the present invention limited range.
Claims (8)
1. a kind of face identification method based on double data enhancing, it is characterised in that：Include the following steps：
S1, selection data set；
S2, data set pretreatment；
S3, the first tuple are according to enhancing：
It builds and confrontation network model is maximumlly generated based on information, confrontation network model is maximumlly generated based on information includes
Generator G and arbiter D, generator G pass through the type and dimension of control latent variable c, to utilize random noise vector z
Generation and the close generation sample G (z, c) of authentic specimen X probability density distribution will generate sample G (z, c) and authentic specimen X points
Input not as arbiter D, arbiter D be responsible for carrying out the generation sample G (z, c) and authentic specimen X of input "true" and
The judgement of "false", finally in maximumlly generating confrontation network model based on information, generator G and arbiter D are carried out constantly
Alternating training, generator G constantly promotes the generative capacity of oneself, and arbiter D then continues to optimize the classification capacity of oneself, when sentencing
When other device D can not judge that some face sample comes from generation sample or authentic specimen X, which is completed；
S4, the second tuple are according to enhancing：
The face sample generated in S3 is translated, is rotated, is overturn and scaling processing；
S5, face classification
Convolutional neural networks model is selected to be trained and identify the face sample generated in S4.
2. a kind of face identification method based on double data enhancing according to claim 1, it is characterised in that：In S1,
The open source human face data collection CelebA for selecting Hong Kong Chinese University, as data set.
3. a kind of face identification method based on double data enhancing according to claim 1, it is characterised in that：It is pre- in S2
Processing method is as follows：The image concentrated to data is cut, and is normalized, the pixel value of image is uniformly arrived [-
1,1] between.
4. a kind of face identification method based on double data enhancing according to claim 1, which is characterized in that in S3,
The method for maximumlly generating confrontation network model based on information is built to include the following steps：
S31, Maker model G is built
The random noise vector for inputting generator G is divided into incompressible variable z, latent variable c, generator G is utilized and made an uproar at random
Sound vector generates the generation sample G (z, c) close with the authentic specimen probability density distribution of face in data set；
Information theory is introduced, the mutual information between the output G (z, c) of latent variable c and generator G is indicated with I (c, G (c, z)),
Mutual information between the two is as follows：
I (c, G (c, z))=H (c)-H (c | G (z, c))=H (G (z, c))-H (G (z, c) | c) (1)
As available from the above equation：When mono- timing of latent variable c, it is possible to reduce the uncertainty of output sample G (z, c), if potential change
It measures c and output sample G (z, c) is uncorrelated, then I (c, G (c, z))=0；Expected from being obtained by control latent variable c
Output maximizes mutual information item I (c, G (c, z)), and the objective function for being originally generated confrontation network is：
It introduces after mutual information item I (c, G (c, z)), the objective function that confrontation network is maximumlly generated based on information is：
Simultaneously in generator G, up-sampled using warp lamination, and use relu activation primitive so that noise by
Fade to high-resolution image；
S32, arbiter model D is built
Arbiter D is made of one four layers of convolutional neural networks, and authentic specimen X is judged as 1 as far as possible by its target, is generated
Sample is judged as 0, and then, it is 0 that setting, which generates sample G (z, c), and authentic specimen X is 1, subsequently into being instructed in arbiter D
Practice, finally obtains the classifier haveing excellent performance；
S33, selection optimizer
Select Adam as optimizer.
5. a kind of face identification method based on double data enhancing according to claim 4, which is characterized in that in S31,
The dimension for defining incompressible variable z is 128 dimensions, and the dimension of latent variable c is 58 dimensions, and latent variable c is tieed up discrete by 5 10
Variable and 8 continuous variables are constituted, wherein the discrete variable of 5 10 dimensions is used for the type of controlling feature, 8 continuous variables
Consecutive variations for controlling feature.
6. a kind of face identification method based on double data enhancing according to claim 4, which is characterized in that in S33,
The inner parameter of Adam is set as：Learning rate lr=0.0005, β=0.5.
7. a kind of face identification method based on double data enhancing according to claim 1, which is characterized in that in S4,
Shift method is：
The overall movement that image is carried out along X-axis or Y direction or simultaneously along X-axis, Y-axis, if certain is put to the mobile tx of X-direction, Y-axis
Ty is moved in direction, and (x, y) is coordinate before converting, and (X, Y) is coordinate after transformation, then the formula translated is：
Spinning solution is：
Image is using certain point as the center of circle, using the point and origin line as radius, rotates θ degree counterclockwise, if the point (x, y), new
Position (X, Y), the then formula rotated are：
Method for turning is：
Image is using X-axis or Y-axis as to axis, acquired mirror image, if coordinate of certain point coordinate (x, y) along X-axis overturning, after conversion
(X1, Y1)；It is overturn along Y-axis, the coordinate (X after conversion2, Y2), overturning expression formula is：
Zoom method is：
The scaling of image refers to that image is zoomed in or out with Y direction according to a certain percentage along the x axis, if in image
The coordinate of certain point is (x, y), scales sx times in X-direction, Y direction scales sy times, and transformed coordinate is (X, Y), then contracts
The formula put is：
8. a kind of face identification method based on double data enhancing according to claim 1, which is characterized in that in S5,
Convolutional neural networks model includes：
Input layer
In input layer, using the face sample image with continuous changing features generated in S4 as the defeated of convolutional neural networks
Enter, wherein 80% is used to train, 20% for testing；
Convolution-pond layer
The depth and width for adjusting convolutional layer, extracting convolutional layer, facial image characteristic effect is best, and different convolution kernels are used for
Extract different features；Shallow-layer convolutional layer is for extracting rudimentary semantic feature；Deep layer convolutional layer is for extracting advanced semanteme
Feature selects the suitable convolution number of plies according to the complexity of image, and the new pixel exported in convolutional layer is by following formula
It is calculated：
Wherein, f () represents activation primitive,Some pixel value of one layer of characteristic image is represented,Represent convolution kernel, * generation
Table convolution algorithm；In view of the output of this layer can be associated with upper one layer of multiple characteristic image, MjIndicate upper one layer of participation operation
Characteristic image subset；Bias term is represented, subscript l indicates l layers；
Pond layer is used to carry out the feature of extraction dimension-reduction treatment, on the basis of the eigenmatrix obtained after convolution algorithm, adds
Enter the processing of maximum pondization, pondization operate in each neuron correspond to the position each N × 1, formula in convolution and be：
Wherein, u (n, 1) is a window function of convolution operation, ajThe maximum value in correspondence image region；
SoftMax layers
SoftMax layers are used to for the output valve of pond layer being mapped as corresponding probability value, where final choice probability value is maximum
Result of the classification as category of model, it is assumed that input feature vector is denoted as x(i), sample label is denoted as y(i), composing training collection S={ (x(1),y(1)),…,(x(m),y(m)), for given input x, its probability value p (y is estimated to each classification j using hypothesized model
=j | x), wherein assuming that function is：
Wherein, θ1,θ2,…,θkFor the model parameter that can learn,To normalize item, so that the sum of all probability are 1,
To obtain cost function：
Wherein, 1 { } was an indicative function, and when the value in bracket is true, otherwise it is 0 that the result of function, which is 1,
It is described to assume that function is the popularization to logistic regression, therefore cost function is readily modified as：
Its partial derivative is asked for SoftMax cost function J (θ), obtains gradient formula：
For a vector, its first of elementIt is J (θ) to θjFirst of component partial derivative,
After obtaining the above solution partial derivative formula, cost function J (θ) is minimized using stochastic gradient descent algorithm,
It requires to be updated parameter in each iterative process：Finally realize SoftMax
Return disaggregated model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810780758.XA CN108921123A (en) | 2018-07-17 | 2018-07-17 | A kind of face identification method based on double data enhancing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810780758.XA CN108921123A (en) | 2018-07-17 | 2018-07-17 | A kind of face identification method based on double data enhancing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108921123A true CN108921123A (en) | 2018-11-30 |
Family
ID=64415352
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810780758.XA Pending CN108921123A (en) | 2018-07-17 | 2018-07-17 | A kind of face identification method based on double data enhancing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921123A (en) |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109543827A (en) * | 2018-12-02 | 2019-03-29 | 清华大学 | Production fights network equipment and training method |
| CN109740608A (en) * | 2018-12-28 | 2019-05-10 | 北京工业大学 | A kind of image partition method based on deep learning |
| CN109840467A (en) * | 2018-12-13 | 2019-06-04 | 北京飞搜科技有限公司 | A kind of in-vivo detection method and system |
| CN109858488A (en) * | 2018-12-28 | 2019-06-07 | 众安信息技术服务有限公司 | A kind of handwriting samples recognition methods and system based on sample enhancing |
| CN109918523A (en) * | 2019-02-14 | 2019-06-21 | 广东工业大学 | A kind of circuit board element detection method based on YOLO9000 algorithm |
| CN109919890A (en) * | 2019-03-06 | 2019-06-21 | 颐保医疗科技(上海)有限公司 | A kind of data enhancement methods applied to medicine identification |
| CN110311870A (en) * | 2019-06-10 | 2019-10-08 | 哈尔滨工业大学（威海） | A kind of SSL VPN method for recognizing flux based on density data description |
| CN110334698A (en) * | 2019-08-30 | 2019-10-15 | 上海聚虹光电科技有限公司 | Glasses detection system and method |
| CN110705402A (en) * | 2019-09-19 | 2020-01-17 | 天津天地伟业机器人技术有限公司 | Face recognition confidence value mapping algorithm |
| CN111047584A (en) * | 2019-12-23 | 2020-04-21 | 福州大学 | Data enhancement method for improving defect classification accuracy of solar cell module |
| CN111079624A (en) * | 2019-12-11 | 2020-04-28 | 北京金山云网络技术有限公司 | Method, device, electronic equipment and medium for collecting sample information |
| CN111414928A (en) * | 2019-01-07 | 2020-07-14 | 中国移动通信有限公司研究院 | Method, device and equipment for generating face image data |
| WO2020168796A1 (en) * | 2019-02-19 | 2020-08-27 | 深圳先进技术研究院 | Data augmentation method based on high-dimensional spatial sampling |
| CN111950619A (en) * | 2020-08-05 | 2020-11-17 | 东北林业大学 | Active learning method based on dual-generation countermeasure network |
| CN112183500A (en) * | 2020-11-27 | 2021-01-05 | 北京易真学思教育科技有限公司 | Data enhancement method, training method, computer device and storage medium |
| CN112257787A (en) * | 2020-10-23 | 2021-01-22 | 天津大学 | Image semi-supervised classification method based on generation type dual-condition confrontation network structure |
| CN112396005A (en) * | 2020-11-23 | 2021-02-23 | 平安科技（深圳）有限公司 | Biological characteristic image recognition method and device, electronic equipment and readable storage medium |
| CN112990225A (en) * | 2021-05-17 | 2021-06-18 | 深圳市维度数据科技股份有限公司 | Image target identification method and device in complex environment |
| CN113034411A (en) * | 2020-12-19 | 2021-06-25 | 北京工业大学 | Road disease picture enhancement method for resisting generation network by coupling traditional method and deep convolution |
| CN113043292A (en) * | 2021-03-22 | 2021-06-29 | 哈尔滨理工大学 | Polar museum welcome robot eye device with eye spirit simulation function and control |
| CN113127806A (en) * | 2021-04-19 | 2021-07-16 | 上海工程技术大学 | Regression analysis model selection method based on machine learning |
| CN114549283A (en) * | 2022-01-14 | 2022-05-27 | 同济大学 | Training method of image generation model and image generation method |
| CN117743719A (en) * | 2023-12-22 | 2024-03-22 | 北京京航计算通讯研究所 | Page element identification method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106951867A (en) * | 2017-03-22 | 2017-07-14 | 成都擎天树科技有限公司 | Face identification method, device, system and equipment based on convolutional neural networks |
| CN107220929A (en) * | 2017-06-23 | 2017-09-29 | 深圳市唯特视科技有限公司 | A kind of non-paired image method for transformation using the consistent confrontation network of circulation |
| CN107423701A (en) * | 2017-07-17 | 2017-12-01 | 北京智慧眼科技股份有限公司 | The non-supervisory feature learning method and device of face based on production confrontation network |
| US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
| CN107945118A (en) * | 2017-10-30 | 2018-04-20 | 南京邮电大学 | A kind of facial image restorative procedure based on production confrontation network |
| CN108171770A (en) * | 2018-01-18 | 2018-06-15 | 中科视拓（北京）科技有限公司 | A kind of human face expression edit methods based on production confrontation network |
| CN108268860A (en) * | 2018-02-09 | 2018-07-10 | 重庆科技学院 | A kind of gas gathering and transportation station equipment image classification method based on convolutional neural networks |
-
2018
- 2018-07-17 CN CN201810780758.XA patent/CN108921123A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180075581A1 (en) * | 2016-09-15 | 2018-03-15 | Twitter, Inc. | Super resolution using a generative adversarial network |
| CN106951867A (en) * | 2017-03-22 | 2017-07-14 | 成都擎天树科技有限公司 | Face identification method, device, system and equipment based on convolutional neural networks |
| CN107220929A (en) * | 2017-06-23 | 2017-09-29 | 深圳市唯特视科技有限公司 | A kind of non-paired image method for transformation using the consistent confrontation network of circulation |
| CN107423701A (en) * | 2017-07-17 | 2017-12-01 | 北京智慧眼科技股份有限公司 | The non-supervisory feature learning method and device of face based on production confrontation network |
| CN107945118A (en) * | 2017-10-30 | 2018-04-20 | 南京邮电大学 | A kind of facial image restorative procedure based on production confrontation network |
| CN108171770A (en) * | 2018-01-18 | 2018-06-15 | 中科视拓（北京）科技有限公司 | A kind of human face expression edit methods based on production confrontation network |
| CN108268860A (en) * | 2018-02-09 | 2018-07-10 | 重庆科技学院 | A kind of gas gathering and transportation station equipment image classification method based on convolutional neural networks |
Non-Patent Citations (3)
| Title |
|---|
| TENG Z.等: "TV-GAN: Generative Adversarial Network Based Thermal to Visible Face Recognition", 《2018 INTERNATIONAL CONFERENCE ON BIOMETRICS (ICB)》 * |
| XI C.等: "InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets", 《百度学术，HTTPS://ARXIV.ORG/PDF/1606.03657.PDF》 * |
| 姚乃明 等: "基于生成式对抗网络的鲁棒人脸表情识别", 《自动化学报》 * |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11574199B2 (en) | 2018-12-02 | 2023-02-07 | Tsinghua University | Generative adversarial network device and training method thereof |
| CN109543827B (en) * | 2018-12-02 | 2020-12-29 | 清华大学 | Generating type confrontation network device and training method |
| CN109543827A (en) * | 2018-12-02 | 2019-03-29 | 清华大学 | Production fights network equipment and training method |
| CN109840467A (en) * | 2018-12-13 | 2019-06-04 | 北京飞搜科技有限公司 | A kind of in-vivo detection method and system |
| CN109740608A (en) * | 2018-12-28 | 2019-05-10 | 北京工业大学 | A kind of image partition method based on deep learning |
| CN109858488A (en) * | 2018-12-28 | 2019-06-07 | 众安信息技术服务有限公司 | A kind of handwriting samples recognition methods and system based on sample enhancing |
| CN111414928A (en) * | 2019-01-07 | 2020-07-14 | 中国移动通信有限公司研究院 | Method, device and equipment for generating face image data |
| CN109918523A (en) * | 2019-02-14 | 2019-06-21 | 广东工业大学 | A kind of circuit board element detection method based on YOLO9000 algorithm |
| CN109918523B (en) * | 2019-02-14 | 2021-03-30 | 广东工业大学 | Circuit board component detection method based on YOLO9000 algorithm |
| WO2020168796A1 (en) * | 2019-02-19 | 2020-08-27 | 深圳先进技术研究院 | Data augmentation method based on high-dimensional spatial sampling |
| CN109919890B (en) * | 2019-03-06 | 2023-01-20 | 颐保医疗科技(上海)有限公司 | Data enhancement method applied to medicine identification |
| CN109919890A (en) * | 2019-03-06 | 2019-06-21 | 颐保医疗科技(上海)有限公司 | A kind of data enhancement methods applied to medicine identification |
| CN110311870B (en) * | 2019-06-10 | 2022-08-02 | 哈尔滨工业大学（威海） | SSL VPN flow identification method based on density data description |
| CN110311870A (en) * | 2019-06-10 | 2019-10-08 | 哈尔滨工业大学（威海） | A kind of SSL VPN method for recognizing flux based on density data description |
| CN110334698A (en) * | 2019-08-30 | 2019-10-15 | 上海聚虹光电科技有限公司 | Glasses detection system and method |
| CN110705402A (en) * | 2019-09-19 | 2020-01-17 | 天津天地伟业机器人技术有限公司 | Face recognition confidence value mapping algorithm |
| CN111079624A (en) * | 2019-12-11 | 2020-04-28 | 北京金山云网络技术有限公司 | Method, device, electronic equipment and medium for collecting sample information |
| CN111079624B (en) * | 2019-12-11 | 2023-09-01 | 北京金山云网络技术有限公司 | Sample information acquisition method and device, electronic equipment and medium |
| CN111047584A (en) * | 2019-12-23 | 2020-04-21 | 福州大学 | Data enhancement method for improving defect classification accuracy of solar cell module |
| CN111950619A (en) * | 2020-08-05 | 2020-11-17 | 东北林业大学 | Active learning method based on dual-generation countermeasure network |
| CN111950619B (en) * | 2020-08-05 | 2022-09-09 | 东北林业大学 | Active learning method based on dual-generation countermeasure network |
| CN112257787A (en) * | 2020-10-23 | 2021-01-22 | 天津大学 | Image semi-supervised classification method based on generation type dual-condition confrontation network structure |
| WO2022105179A1 (en) * | 2020-11-23 | 2022-05-27 | 平安科技（深圳）有限公司 | Biological feature image recognition method and apparatus, and electronic device and readable storage medium |
| CN112396005A (en) * | 2020-11-23 | 2021-02-23 | 平安科技（深圳）有限公司 | Biological characteristic image recognition method and device, electronic equipment and readable storage medium |
| CN112183500A (en) * | 2020-11-27 | 2021-01-05 | 北京易真学思教育科技有限公司 | Data enhancement method, training method, computer device and storage medium |
| CN113034411A (en) * | 2020-12-19 | 2021-06-25 | 北京工业大学 | Road disease picture enhancement method for resisting generation network by coupling traditional method and deep convolution |
| CN113034411B (en) * | 2020-12-19 | 2024-02-27 | 北京工业大学 | Road disease picture enhancement method coupling traditional method and depth convolution countermeasure generation network |
| CN113043292A (en) * | 2021-03-22 | 2021-06-29 | 哈尔滨理工大学 | Polar museum welcome robot eye device with eye spirit simulation function and control |
| CN113127806A (en) * | 2021-04-19 | 2021-07-16 | 上海工程技术大学 | Regression analysis model selection method based on machine learning |
| CN112990225A (en) * | 2021-05-17 | 2021-06-18 | 深圳市维度数据科技股份有限公司 | Image target identification method and device in complex environment |
| CN114549283A (en) * | 2022-01-14 | 2022-05-27 | 同济大学 | Training method of image generation model and image generation method |
| CN117743719A (en) * | 2023-12-22 | 2024-03-22 | 北京京航计算通讯研究所 | Page element identification method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921123A (en) | A kind of face identification method based on double data enhancing | |
| CN109815826B (en) | Method and device for generating face attribute model | |
| Niu et al. | Ordinal regression with multiple output cnn for age estimation | |
| CN113221639B (en) | Micro-expression recognition method for representative AU (AU) region extraction based on multi-task learning | |
| CN112950661B (en) | Attention-based generation method for generating network face cartoon | |
| CN110021051A (en) | One kind passing through text Conrad object image generation method based on confrontation network is generated | |
| CN110222668A (en) | Based on the multi-pose human facial expression recognition method for generating confrontation network | |
| CN104361316B (en) | Dimension emotion recognition method based on multi-scale time sequence modeling | |
| Zhai et al. | BeautyNet: Joint multiscale CNN and transfer learning method for unconstrained facial beauty prediction | |
| CN109034034A (en) | A kind of vein identification method based on nitrification enhancement optimization convolutional neural networks | |
| CN108932517A (en) | A kind of multi-tag clothes analytic method based on fining network model | |
| KR102326902B1 (en) | Image-based Posture Preservation Virtual Fitting System Supporting Multi-Poses | |
| Sun et al. | A discriminatively deep fusion approach with improved conditional GAN (im-cGAN) for facial expression recognition | |
| CN108985464A (en) | The continuous feature generation method of face for generating confrontation network is maximized based on information | |
| CN113393550B (en) | Fashion garment design synthesis method guided by postures and textures | |
| CN104915658B (en) | A kind of emotion component analyzing method and its system based on emotion Distributed learning | |
| Olague et al. | Evolving head tracking routines with brain programming | |
| Zhu et al. | Convolutional ordinal regression forest for image ordinal estimation | |
| Zhu et al. | Facial aging and rejuvenation by conditional multi-adversarial autoencoder with ordinal regression | |
| CN112883756A (en) | Generation method of age-transformed face image and generation countermeasure network model | |
| Agarwal et al. | Synthesis of realistic facial expressions using expression map | |
| CN111400525A (en) | Intelligent fashionable garment matching and recommending method based on visual combination relation learning | |
| Kim et al. | Embedded face recognition based on fast genetic algorithm for intelligent digital photography | |
| Wang et al. | SURVS: A Swin-Unet and game theory-based unsupervised segmentation method for retinal vessel | |
| Surya et al. | ReStGAN: A step towards visually guided shopper experience via text-to-image synthesis |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181130 |