CN110503072A - Face age estimation method based on multiple branch circuit CNN framework - Google Patents
Face age estimation method based on multiple branch circuit CNN framework Download PDFInfo
- Publication number
- CN110503072A CN110503072A CN201910806822.1A CN201910806822A CN110503072A CN 110503072 A CN110503072 A CN 110503072A CN 201910806822 A CN201910806822 A CN 201910806822A CN 110503072 A CN110503072 A CN 110503072A
- Authority
- CN
- China
- Prior art keywords
- age
- cnn
- face
- framework
- 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.)
- Granted
Links
Classifications
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
-
- 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
- G06V40/178—Human faces, e.g. facial parts, sketches or expressions estimating age from face image; using age information for improving recognition
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of face age estimation method based on multiple branch circuit CNN framework, this patent is taken into account face character (gender, colour of skin etc.) in age estimation task by designing multiple branch circuit CNN framework, to improve the accuracy rate of face age estimation.Be primarily based on a kind of black race male for identification of classical CNN architecture design, black race women, Caucasian male, Caucasian female, yellow race male, yellow race women these sixth types CNN model；Then the influence degree of model first half framework and its parameter as inclusion layer to reinforce face character to age estimation task is intercepted；Then the age of different faces attribute is estimated by separation by the design six CNN branches for six class face characters；The output result of last six branches is merged by the blending algorithm of public fused layer, to allow neural network to learn respectively to specific ethnic group, the face picture feature of gender, to improve the accuracy rate of age estimation.
Description
Technical field
The present invention relates to recognition of face, in particular to a kind of face age estimation method based on multiple branch circuit CNN framework.
Background technique
Currently, depth learning technology continues to develop, progress, become one of current most popular scientific trend.Convolution mind
Through network (Convolutional Neural Network, CNN) as the important algorithm in deep learning, it is good at very much processing
The problem of image correlation, is nowadays widely used in computer vision field, recognition of face, in terms of play focus on
It acts on.Age estimation (Age Estimation, AE) problem of face is always the popular problem in this field, and people are always
How to optimize AE algorithm in research, improves the accuracy rate of algorithm estimation.
Currently, in AE, for be usually two kinds of genders of more ethnic groups and men and women estimation, in training neural network
Generally use the public data collection that Morph, AgeDB and Webface etc. include more ethnic groups, gender and all ages and classes.So
And the face of different ethnic groups has notable difference, the colour of skin, face texture in terms of the facial characteristics such as eyes, nose type, cheekbone, lip
Etc. features it is also different.Equally, the face of masculinity femininity also has notable difference on many facial characteristics, for example, white people, black
The age information of kind people and yellow will be different in face embodiment, and for the same age, black race's face is in visual impression
Official embodies can be partially old, and white people are then partially young (for yellow)；Equally, for the same age, female
Property face compared with for male in visual perception's embodiment it is younger.
Overwhelming majority face age estimation method is all based on extensive human face data collection at present, i.e., directly will include more ethnic groups
And the image data of different sexes, as training set, by a single channel CNN come training pattern, although this mode ensure that
The diversity, rich of face picture needed for training, but ignore different ethnic groups, the nuance of gender facial image for
The influence of neural network, although there is scholar to have paid attention to influence of the age attribute to final AE performance, few researchs
Associated age attribute can be integrated into deep neural network framework by method appropriate, be trained to greatly limit
The AE model come is to specific ethnic group, the AE performance of the face picture of gender.
Actually, if it is possible to be trained respectively for different ethnic groups, the face picture of gender, neural network is allowed to distinguish
Learn to specific ethnic group, the face picture feature of gender, and is melted " knowledge " that learns using suitable policing algorithm
It closes, so that it may solve the problems, such as that above-mentioned existing face age estimation method exists, to a certain extent so as to improve the face age
The recognition performance for estimating model, enables it to have better estimation effect for different ethnic groups, the face picture of gender.
Summary of the invention
Goal of the invention: the object of the present invention is to provide a kind of face age estimation methods based on multiple branch circuit CNN framework, more
The problem that mending tradition single channel CNN, by constructing a novel multiple branch circuit CNN, by the facial image of different ethnic groups, gender
Different information take into account, improve final AE model for specific ethnic group, the AE performance of gender facial image.It is pair on the whole
The CNN model for being currently used for AE task is improved and is innovated, with improve final face age arbiter to different ethnic groups, no
The accuracy of identification of the other face picture of the same sex.
Technical solution: the face age estimation method of the present invention based on multiple branch circuit CNN framework includes the following steps
Step 1, data set is introduced, CNN framework is constructed, data set imports in CNN framework and trains one for identification
Black race male, black race women, Caucasian male, Caucasian female, yellow race male and yellow race women CNN model；Image is introduced into
The CNN model carries out ethnic group and recognition result is screened and exported to gender, and as a result output is one 6 × 1 column vector, uses Fk group,
K ∈ { BM, BF, WM, WF, YM, YF } indicates that value indicates the matching probability of input picture and specific face character k；
Step 2, the first half framework for the model that training obtains in step 1 and parameter are saved, as inclusion layer, setting
The purpose of inclusion layer is to reinforce age attribute for the influence degree of next age estimation task.
Step 3, it is made of in inclusion layer and then secondary building CNN framework, the CNN framework 6 branches, 6 branch difference
Act on the age attribute of identification black race male, black race women, Caucasian male, Caucasian female, yellow race male and yellow race women
The age of face picture estimates that the output result of the last full articulamentum of 6 branches is the column vector of m × 1, and m is
The quantity at age, uses Fk, k ∈ { BM, BF, WM, WF, YM, YF } expression, matching of the value expression input picture in face character k
On the basis of matching degree with the specific m age；
Step 4, public fused layer is established after the CNN model being made of 6 branches, is used in public fused layer part
6 branch Fusion Features, the formula of public fused layer be by formula 1
Fa=∑ Fk group·Fk, k ∈ { BM, BF, WM, WF, YM, YF } (1)
FaIt is fused parameter vector, is the column vector of m × 1, m is the quantity at age, indicates the picture most
The whole age estimates weight, obtains result at this time and takes into account the different information of the facial image of different ethnic groups, gender；
The objective function that public fused layer is followed by uses cross entropy loss function, and the objective function is true for the computation model identification age
Error between value and predicted value is lost and is updated with weight parameter of the stochastic gradient descent algorithm to 6 branches.
The utility model has the advantages that this patent devises the branch of the different sexes of 6 different ethnic groups for age attribute, every picture is first
It first passes through inclusion layer and a series of convolution, Chi Hua, full context layer forms a kind of CNN model, for identification black race male, black
Equal six branches of kind women.This 6 branch structures are identical, and parameter is different, in public fused layer below, this 6 branch
The multiplied by weight that the output on road is calculated with every branch that pre-training obtains before, assigns this six CNN branch by blending algorithm
The other meaning of the corresponding race and sex in road, and when last backpropagation Optimal Parameters, it can also play different-effect.From
And neural network is pointedly trained respectively to different ethnic groups, the face picture of gender, and then improve face
Age estimates the recognition performance of model, it is enable to have better estimation effect for different ethnic groups, the face picture of gender.
Detailed description of the invention
Fig. 1 is the schematic diagram of the face age estimation method based on multiple branch circuit CNN framework.
Specific embodiment
As shown in Figure 1, a kind of face age estimation method based on multiple branch circuit CNN framework, comprising the following steps:
Step 1: it is primarily based on the existing large data collection such as Morph II, Webface and passes through the warp such as VGG or Resnet
Allusion quotation CNN framework establishes black race male, black race women, Caucasian male, Caucasian female, yellow race male and a yellow race for identification
The model of this six classes of women, loss function use cross entropy loss function, utilize stochastic gradient descent algorithm (stochastic
Gradient descent, SGD) model training is extremely restrained.Using stochastic gradient descent algorithm as model optimization algorithm, by
It is avoided by way of calculating a sample every time in the algorithm compared to traditional batch gradient descent algorithm and is being joined every time
Number carries out the redundancy that computes repeatedly of gradient to similar sample before updating, therefore has faster optimal speed, under stochastic gradient
Algorithm is dropped with a training sample xiWith predicted value yiParameter updates, and parameter more new formula isWherein θ is
Weight, η are learning rate, and J is objective function to be optimized.After training, the output result of the final full articulamentum of the model
For 6 × 1 column vector, respectively represent from top to down black race male, black race women, Caucasian male, Caucasian female, yellow race male with
And yellow race women these sixth types age attribute, numerical value can characterize identification knot, numerical value is bigger, belong to it is higher a possibility that such,
Use Fk group, k ∈ { BM, BF, WM, WF, YM, YF } expression is for specific face character k.
Step 2: the first half framework for the model that training obtains in step 1 and parameter are saved, as inclusion layer.If
Setting the layer is to incorporate this correlation in entire model for face character sexual enlightenment related to age objective reality.It should
Inclusion layer is CNN framework trained in advance in step 1 (being what age attribute for differentiating) front section, wherein remaining
Part CNN framework and the weight parameter of training.
Step 3: 6 independent CNN are arranged based on classics CNN frameworks such as VGG, Resnet again after inclusion layer
Framework (6 branches), this 6 branches are respectively acting on Black-Male, Black-Female, White-Male, White-
The age of six kinds of age attribute face pictures of Female, Yellow-Male and Yellow-Female estimates that initial parameter is arranged
It is random initializtion.For an input face picture, the output result of the last full articulamentum of 6 branches is one
The column vector of m × 1, m are the quantity at age, use Fk, k ∈ { BM, BF, WM, WF, YM, YF } is indicated, indicates that this inputs picture most
It may be which (full articulamentum export age representated by the corresponding position of maximum numerical value be to estimate the age) at age
Step 4: use formula 1 by 6 branch Fusion Features in public fused layer part.
Fa=∑ Fk group·Fk, k ∈ { BM, BF, WF, YM, YF } (1)
FaIt is fused parameter vector, is the column vector of m × 1, indicates the final age estimation power of the picture
Weight, obtains result at this time and takes into account the different information of the facial image of different ethnic groups, gender.Last objective function
Using cross entropy loss function, cross entropy loss function formula isWherein N is training picture
Total amount, yiPass through the age prediction result that CNN framework obtains for i-th trained picture,For the true year of i-th trained picture
Age label, the loss function calculate the true value of trained picture and pass through the accumulative mistake between the predicted value that CNN framework obtains
Difference, and with stochastic gradient descent algorithm by entire CNN model training to restraining, so far age for fully considering face character
Estimate that model training is completed.
So far, whole flow process terminates.
Claims (5)
1. a kind of face age estimation method based on multiple branch circuit CNN framework, it is characterised in that include the following steps
Step 1, data set is introduced, CNN framework is constructed, data set imports in CNN framework and trains a black race for identification
Male, black race women, Caucasian male, Caucasian female, yellow race male and yellow race women CNN model；Image is introduced into the CNN
Model carries out ethnic group and recognition result is screened and exported to gender, and as a result output is one 6 × 1 column vector, uses Fk group, k ∈
{ BM, BF, WM, WF, YM, YF } is indicated, value indicates the matching probability of input picture and specific face character k；
Step 2, the first half framework for the model that training obtains in step 1 and parameter are saved, as inclusion layer, setting is shared
The purpose of layer is to reinforce age attribute for the influence degree of next age estimation task；
Step 3, it is made of in inclusion layer and then secondary building CNN framework, the CNN framework 6 branches, 6 branches act on respectively
In the age attribute face for identifying black race male, black race women, Caucasian male, Caucasian female, yellow race male and yellow race women
The age of picture estimates that the output result of the last full articulamentum of 6 branches is the column vector of m × 1, and m is the age
Quantity, use Fk, k ∈ { BM, BF, WM, WF, YM, YF } expression, value expression input picture is on the matching basis of face character k
The upper matching degree with the specific m age；
Step 4, public fused layer is established after the CNN model being made of 6 branches, uses formula in public fused layer part
6 branch Fusion Features, the formula of public fused layer are by 1
Fa=Σ Fk group·Fk, k ∈ { BM, BF, WM, WF, YM, YF } (1)
FaIt is fused parameter vector, is the column vector of m × 1, m is the quantity at age, indicates the picture final year
Age estimates weight, obtains result at this time and takes into account the different information of the facial image of different ethnic groups, gender；It is public to melt
The objective function that conjunction layer is followed by is using cross entropy loss function, and the objective function is for computation model identification age true value and in advance
Error between measured value is lost and is updated with weight parameter of the stochastic gradient descent algorithm to 6 branches.
2. the face age estimation method according to claim 1 based on multiple branch circuit CNN framework, it is characterised in that in step
In 1, the model for constructing CNN model includes VGG and Resnet model, and specific model selection is depending on the circumstances.
3. the face age estimation method according to claim 2 based on multiple branch circuit CNN framework, it is characterised in that use phase
Same 6 CNN models of model construction, and use identical parameter initialization method.
4. the face age estimation method according to claim 1 based on multiple branch circuit CNN framework, it is characterised in that in step
In 4, cross entropy loss function formula isWherein N is training picture total amount, yiIt is instructed for i-th
Practice picture and pass through the age prediction result that CNN framework obtains,For the real age label of i-th trained picture, the loss letter
Number calculates the true value of trained picture and passes through the cumulative errors between the predicted value that CNN framework obtains.
5. the face age estimation method according to claim 2 based on multiple branch circuit CNN framework, it is characterised in that in step
In 4, using stochastic gradient descent algorithm as model optimization algorithm, calculated since the algorithm declines compared to traditional batch gradient
Method avoids the repetition for carrying out gradient to similar sample before every subparameter updates by way of calculating a sample every time
Computing redundancy, therefore there is faster optimal speed, stochastic gradient descent algorithm is with a training sample xiWith predicted value yiParameter
It updates, parameter more new formula isWherein θ is weight, and η is learning rate, and J is objective function to be optimized.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910806822.1A CN110503072B (en) | 2019-08-29 | 2019-08-29 | Face age estimation method based on multi-branch CNN architecture |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910806822.1A CN110503072B (en) | 2019-08-29 | 2019-08-29 | Face age estimation method based on multi-branch CNN architecture |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110503072A true CN110503072A (en) | 2019-11-26 |
| CN110503072B CN110503072B (en) | 2023-05-26 |
Family
ID=68590306
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910806822.1A Active CN110503072B (en) | 2019-08-29 | 2019-08-29 | Face age estimation method based on multi-branch CNN architecture |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503072B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112733717A (en) * | 2021-01-11 | 2021-04-30 | 广州市金其利信息科技有限公司 | Method for optimizing face recognition based on face attributes |
| CN113095300A (en) * | 2021-05-13 | 2021-07-09 | 华南理工大学 | Age prediction method and system fusing race information |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106447625A (en) * | 2016-09-05 | 2017-02-22 | 北京中科奥森数据科技有限公司 | Facial image series-based attribute identification method and device |
| CN106529402A (en) * | 2016-09-27 | 2017-03-22 | 中国科学院自动化研究所 | Multi-task learning convolutional neural network-based face attribute analysis method |
-
2019
- 2019-08-29 CN CN201910806822.1A patent/CN110503072B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106447625A (en) * | 2016-09-05 | 2017-02-22 | 北京中科奥森数据科技有限公司 | Facial image series-based attribute identification method and device |
| CN106529402A (en) * | 2016-09-27 | 2017-03-22 | 中国科学院自动化研究所 | Multi-task learning convolutional neural network-based face attribute analysis method |
Non-Patent Citations (1)
| Title |
|---|
| MINGXING DUAN等: "An Ensemble CNN2ELM for Age Estimation", 《HTTPS://IEEEXPLORE.IEEE.ORG/ABSTRACT/DOCUMENT/80825610》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112733717A (en) * | 2021-01-11 | 2021-04-30 | 广州市金其利信息科技有限公司 | Method for optimizing face recognition based on face attributes |
| CN112733717B (en) * | 2021-01-11 | 2021-11-23 | 广州市金其利信息科技有限公司 | Method for optimizing face recognition based on face attributes |
| CN113095300A (en) * | 2021-05-13 | 2021-07-09 | 华南理工大学 | Age prediction method and system fusing race information |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110503072B (en) | 2023-05-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106503623B (en) | Facial image age estimation method based on convolutional neural networks | |
| CN109815826B (en) | Method and device for generating face attribute model | |
| CN106529402B (en) | The face character analysis method of convolutional neural networks based on multi-task learning | |
| CN109299262A (en) | A kind of text implication relation recognition methods for merging more granular informations | |
| CN108984530A (en) | A kind of detection method and detection system of network sensitive content | |
| CN108921123A (en) | A kind of face identification method based on double data enhancing | |
| CN107368831A (en) | English words and digit recognition method in a kind of natural scene image | |
| CN106650827A (en) | Human body posture estimation method and system based on structure guidance deep learning | |
| CN109325547A (en) | Non-motor vehicle image multi-tag classification method, system, equipment and storage medium | |
| CN107133211A (en) | A kind of composition methods of marking based on notice mechanism | |
| CN113642431B (en) | Training method and device of target detection model, electronic equipment and storage medium | |
| CN112800292B (en) | Cross-modal retrieval method based on modal specific and shared feature learning | |
| CN111582342B (en) | Image identification method, device, equipment and readable storage medium | |
| CN109817276A (en) | A kind of secondary protein structure prediction method based on deep neural network | |
| CN111784121B (en) | Action quality evaluation method based on uncertainty score distribution learning | |
| CN109783641A (en) | A kind of entity relationship classification method based on two-way-GRU and improved attention mechanism | |
| CN108932517A (en) | A kind of multi-tag clothes analytic method based on fining network model | |
| CN112860945B (en) | Method for multi-mode video question answering by using frame-subtitle self-supervision | |
| CN110110663A (en) | A kind of age recognition methods and system based on face character | |
| CN113569732B (en) | Face attribute identification method and system based on parallel sharing multitasking network | |
| CN110503072A (en) | Face age estimation method based on multiple branch circuit CNN framework | |
| CN113220891B (en) | Method for generating confrontation network image description based on unsupervised concept-to-sentence | |
| CN104952067A (en) | Method for segmenting color images on basis of NSGA-II (non-dominated sorting genetic algorithm-II) evolution algorithms | |
| CN111652798A (en) | Human face pose migration method and computer storage medium | |
| CN115374189B (en) | Block chain-based food safety tracing method, device and equipment |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |