CN106919897A - A kind of facial image age estimation method based on three-level residual error network - Google Patents
A kind of facial image age estimation method based on three-level residual error network Download PDFInfo
- Publication number
- CN106919897A CN106919897A CN201611264499.2A CN201611264499A CN106919897A CN 106919897 A CN106919897 A CN 106919897A CN 201611264499 A CN201611264499 A CN 201611264499A CN 106919897 A CN106919897 A CN 106919897A
- Authority
- CN
- China
- Prior art keywords
- residual error
- error network
- level
- training
- age
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- 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 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
A kind of facial image age estimation method based on three-level residual error network, belong to technical field of data processing, purpose is to improve the facial image age estimation level under the conditions of untethered, and its technical scheme is that methods described sets up three-level residual error network on the basis of basic residual error network frame first；Then pre-training is carried out to ImageNet data sets using three-level residual error network, obtains ImageNet residual error network models；Training is finely adjusted to acquired ImageNet residual errors network model on the face age data collection under the conditions of untethered again；The three-level residual error network for finally being trained using fine setting carries out the estimation of facial image age.The present invention is estimated using the three-level residual error real-time performance facial image age, not only substantially increase the learning ability of DCNN network models, and solve the problems, such as that the over-fitting in training process and gradient disappear well, so as to improve the accuracy that the facial image age under the conditions of untethered estimates.
Description
Technical field
The age true method estimated can be realized according to facial image under the conditions of untethered the present invention relates to a kind of, belong to several
According to processing technology field.
Background technology
Face is extremely abundant information source, and people can obtain a large amount of useful informations, such as identity, property from facial image
Not, age and expression etc..One of key message as face, the age serves basic work in the social interaction of people
With, therefore realize that the automatic estimation at age is one of important process of artificial intelligence field by facial image.At present, face year
Age is estimated in the multiple intelligence such as the man-machine interaction based on age level, access control, vision monitoring, the marketing and Law enforcement
There is good application prospect in energy field.
The main thought that the facial image age is estimated is that principal character is extracted from facial image, then uses classification or returns
Method is returned to realize that the age is estimated, wherein character classification by age is used to estimate the age packet of people, and it is then to estimate that people's is accurate that the age returns
Age.Overwhelming majority face age estimation method carried out age estimation, initially, Kwon et al. using the feature of engineer in the past
The geometric properties that face is obtained by calculating the distance between human face characteristic point ratio are estimated for the age；Cootes et al. is in people
Global textural characteristics are added on the basis of face geometric properties, it is proposed that AAM models；Then, LBP, SFP and BIF feature respectively by
It is used as the face characteristic of age estimation.Feature based on these engineers, returns and sorting technique be used to estimate people
Face, the method based on SVM is used for age group classification, is returned for the age, then mainly include the methods such as SVR, PLS, CCA.This
A little artificial methods for extracting feature obtain good result, such as face age number on the human face data collection under confined condition
According to collection FG-NET and MORPH.But these methods seem barely satisfactory in the face age estimation task under the conditions of untethered,
Such as face age data collection Adience under the conditions of untethered, the facial image in these data sets is not by artificial mistake
Filter and adjustment, include a variety of changes, such as noise, illumination, posture, expression etc. in image, these changes turn to age estimation
Bring new challenge.
At present, depth convolutional neural networks (DCNN) have turned into the focus of computer vision field research.From 5-conv+3-
The VGG networks and the GoogleNet of 21-conv+1-fc of AlexNet to the 16-conv+3fc of fc, then to thousands of layers
ResNets, either the learning ability or depth of network be all significantly improved.Therefore, increasing scholar in recent years
Begin attempt to using DCNN solve age estimation problem, and demonstrate its be obtained in that under the conditions of untethered be substantially better than in the past
The result of manual extraction characterization method.Yi etc. proposes multiple dimensioned DCNN age estimation methods at first, and is carried out on MORPH
Checking；Wang etc. extracts face feature using DCNN, then carries out age estimation using SVR, is obtained on FG-NET and MORPH
Obtained preferable effect；Levi etc. carries out age and sex point using DCNN on the Adience data sets of untethered condition
Class, the age is estimated more to challenge under the conditions of finding untethered；Ekmekji proposes sex and the character classification by age side of a kind of chain type
Method, the method is respectively trained DCNN on Adience data sets for different sexes.The DCNN that above method is used only is wrapped
Containing a small amount of convolutional layer and full articulamentum, this greatly limits learning abilities of the DCNN in terms of age estimation.In order to enter one
Step improves the network knot using the similar VGG-16 of deep layer such as the learning ability of model, and then the accuracy rate that the raising age is estimated, Hou
Structure and the smooth activation primitive (SAAF) that adapts to carry out age estimation on Adience data sets, achieve preferable effect；
Rothe etc. proposes depth expectational model (DEX), and it uses model based on the VGG-16 network architectures, using ImageNet
Data set carries out pre-training, is then finely adjusted using IMDB-WIKI large-scale data set pair networks, finally in different faces
Training is finely adjusted on data set, current best result has been obtained, but on Adience data sets under the conditions of untethered
Still it is difficult to reach the age estimation level of people.
Existing DCNN age estimation methods are primarily present problems with：
1. age estimation effect of the existing method under the conditions of untethered is poor, and this is the DCNN that use is estimated due to the age
E-learning scarce capacity is caused.So the stronger DCNN networks of design learning ability are needed, to improve age estimated capacity.
2. because human face data collection image scale is generally smaller, when being trained using larger DCNN networks, it is easy to
Over-fitting problem is caused, so being needed using the technological means for suppressing over-fitting in model training.
3. deeper DCNN networks can cause gradient disappearance problem, and the serious study for hindering disaggregated model of gradient disappearance is entered
Journey, limits the learning ability of DCNN networks, the accuracy that the influence facial image age is estimated.
The content of the invention
A kind of drawback it is an object of the invention to be directed to prior art, there is provided facial image based on three-level residual error network
Age estimation method, to improve the accuracy that the facial image age is estimated under the conditions of untethered.
Problem of the present invention is solved with following technical proposals：
A kind of facial image age estimation method based on three-level residual error depth convolutional neural networks, methods described exists first
Three-level residual error network is set up on the basis of basic residual error network frame；Then ImageNet data sets are entered using three-level residual error network
Row pre-training, obtains ImageNet residual error network models；Again to having obtained on the face age data collection under the conditions of untethered
ImageNet residual error network models be finely adjusted training；The three-level residual error network for finally being trained using fine setting carries out face figure
As the age is estimated.Three-level residual error network is a kind of new depth convolutional neural networks structure of foundation on residual error network foundation,
It adds extra two-stage shortcut connections on the basis of raw residual network structure.
The above-mentioned facial image age estimation method based on three-level residual error network, the described method comprises the following steps：
A. three-level residual error network is set up
The model based on basic residual error network model (ResNets), adds first outside all residual blocks
Shortcut branch roads, referred to as first order shortcut branch roads；Then shortcut branch roads are added on the basis of each residual block group,
Referred to as second level shortcut branch roads；Shortcut branch roads in raw residual block are referred to as third level shortcut branch roads, the
One-level and second level shortcut branch roads include a convolutional layer；
B. pre-training is carried out to ImageNet data sets using three-level residual error network, obtains ImageNet three-level residual error networks
Pre-training model, during model pre-training, is carried out using the basic residual error network model trained to three-level residual error network
Initialization, i.e., convolutional layer in raw residual block, BN layers initialized by basic residual error network model, first and second grades
Convolutional layer on shortcut branch roads is initialized using MSRA methods；
C. acquired ImageNet residual errors network model is carried out on the face age data collection under the conditions of untethered
Fine setting training, random depth algorithm is used in model fine setting training process, i.e., disconnect some residual errors at random when training is finely tuned
Mapping branch road, to change path of information flow, reaches the purpose for suppressing that over-fitting and gradient disappear；
D. the estimation of facial image age is carried out with the three-level residual error network for training.
The present invention not only substantially increases DCNN network models using the estimation of three-level residual error real-time performance facial image age
Learning ability, and solve the problems, such as that the over-fitting in training process and gradient disappear well, it is non-so as to improve
The accuracy that confined condition servant's face image age is estimated.
Brief description of the drawings
Fig. 1 is ResNets-34 and RoR-3-34 schematic diagrames；
Fig. 2 is random depth residual error network diagram.
Each symbol is in text：LBP (local binary patterns), SFP (spatial elastic feature), BIF (biological characteristic), SVR (branch
Hold vector regression), PLS (partial least squares), CCA (typical association analysis), shortcut (shortcut branch road), epoch (
Individual training circulation).
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.
The present invention estimates presently, there are to solve the problems, such as the face age, establishes three-level residual error network, and use three
Level residual error network carries out ImageNet data set pre-training, and then right on face age estimated data collection under the conditions of untethered
The model that pre-training is obtained is finely adjusted, and over-fitting problem is suppressed using random depth algorithm in trim process.
The present invention includes following 3 steps：
1. three-level residual error network is set up on the basis of basic residual error network frame first, to improve of DCNN network models
Habit ability.
2. it is limited to the scale of face age estimated data collection, ImageNet data sets is carried out using three-level residual error network pre-
Training, obtains ImageNet residual error network models.
3. acquired ImageNet residual errors network model is carried out on the face age data collection under the conditions of untethered
Fine setting training；Over-fitting problem is suppressed using random depth algorithm in training process is finely tuned.
1st step sets up three-level residual error network, improves DCNN network model learning abilities.This patent chooses basic residual error network
Model based on model (ResNets), by taking 34 layers of basic residual error network model (ResNets-34) as an example, it includes four groups altogether
Residual block, each residual block group is respectively comprising the residual block that quantity is 3,4,6 and 3, and the residual block output dimension in every group
It is identical, respectively 64,128,256 and 512, each residual block maps branch road by a residual error and a shortcut branch road is superimposed
Composition, residual error mapping branch road is included in order：Convolutional layer-BN layers of-BN layers of-ReLU layers-convolutional layer, first in each residual block group
The shortcut branch roads of individual residual block include a convolutional layer, and the shortcut branch roads of remaining residual block (are not wrapped for identical mapping
Containing any intermediate layer), residual error mapping branch road and shortcut branch roads are connected one ReLU layers after being superimposed, ResNets-34 and its
Comprising residual error block structure such as Fig. 1 (a) shown in.
Fig. 1 (b) shows the detailed process for building three-level residual error network, first, is added outside all residual blocks
Shortcut branch roads, referred to as first order shortcut branch roads；Then shortcut branch roads are added on the basis of each residual block group,
Referred to as second level shortcut branch roads；Shortcut branch roads in raw residual block are referred to as third level shortcut branch roads, the
One-level and second level shortcut branch roads include a convolutional layer.This patent represents 34 layers of three-level of structure using RoR-3-34
Residual error network.As shown in Fig. 1 (b), the first order or second level shortcut branch roads are superimposed with the output of residual block group, form new
Residual block, and the residual error block structure of nested form has been collectively constituted with raw residual block.These additive processes can use following formula (1)
Represent.Wherein xlWith xl+1L-th input and output of residual block are represented, F represents residual error mapping function, h (xl)=xlRepresent permanent
Deng mapping, g (xl) represent first and second grade of convolution mapping of shortcut branch roads.
x3+1=g (x1)+h(x3)+F(x3,W3)
x3+4+1=g (x3+1)+h(x3+4)+F(x3+4,W3+4)
2nd step is that pre-training is carried out to ImageNet data sets using three-level residual error network, obtains ImageNet three-levels residual
Difference network pre-training model.Because face age estimated data collection image scale is relatively small, easily produced in the training process
The problem of fitting, so this patent uses three-level residual error network training ImageNet data sets, to obtain the essential characteristic of image
Expression model, and then model using pre-training is finely adjusted on face age estimated data collection, so as to alleviate directly training
The over-fitting problem that face age data collection brings.Due to directly using three-level residual error network RoR- on ImageNet data sets
The more basic residual error network of convergence rate is many slowly when 3-34 is trained, so in order to save the training time, using having trained
Basic residual error network model three-level residual error network RoR-3-34 is initialized, i.e. convolution in model in original residual block
Layer is initialized with BN layers by basic residual error network model, and the convolutional layer on first and second grades of shortcut branch roads is used
MSRA methods are initialized.Once being input into a collection of picture during model training during each iteration carries out batch processing, and using random
Gradient descent method updates the weight of network, and every batch of size is 128.The maximum epoch numbers of training are 10, and initial learning rate is
Learning rate is reduced to 0.0001 after 0.001,5 epoch.
3rd step is finely adjusted instruction on face age data collection to acquired ImageNet residual errors network pre-training model
Practice；Over-fitting problem is suppressed using random depth algorithm in training process is finely tuned.Detailed process is as follows：
(1) network is broadened while three-level residual error network adds extra shortcut branch roads, and adds more parameters,
Over-fitting problem may be caused more serious.This patent disconnects some network paths so as to change at random when model is finely tuned and trained
Path of information flow, the purpose of over-fitting and gradient disappearance problem is suppressed to reach.This patent disconnects residual error mapping branch at random
Rather than shortcut branch roads, this is the random disconnection because shortcut branch roads are the main transmission paths of information on road
Shortcut branch roads can cause the not convergent result of training.Make plRepresent unimpeded general of residual error mapping branch road of l-th residual block
Rate, L is the number of original residual block in three-level residual error network, formula (2) display plSuccessively decrease with residual block position linearity, pLRepresent most
The unimpeded probability of latter residual block residual error mapping branch road.Fig. 2 shows and works as L=5, pLRandom depth algorithm is used when=0.5
Residual error schematic network structure, flRepresent residual error mapping function, xlRepresent l-th input of residual block, plControl residual error mapping
flConnection and closure.
(2) the three-level residual error network pre-training model that random depth algorithm and step 2 are obtained more than is in face year
It is finely adjusted on age estimated data collection.Different training parameters is selected for different face age estimated data collection.Received with non-
, it is necessary to the full articulamentum fc-1000 in the model that pre-training is obtained is replaced with as a example by Adience data sets under the conditions of limit
Fc-8, output classification number is changed to 8 by 1000, and being once input into a collection of picture during each iteration during model training carries out batch processing, adopts
The weight of network is updated with stochastic gradient descent method, every batch of size is 128, the maximum epoch numbers of training are 120, initial study
Rate is that learning rate is reduced to 0.001 after 0.01,80 epoch.
The present invention has advantages below：
1st, the present invention adds shortcut branch roads step by step on basic residual error network foundation, establishes three-level residual error network.
Original residual block is converted into residual error mapping by three-level residual error network, then the residual error mapping in raw residual block is converted into residual error and reflects
The residual error mapping hit, and then learning difficulty is reduced, improve e-learning on the premise of substantially network parameter is not increased
Ability.In addition, three-level residual error network makes high-rise residual block be transmitted to bottom residual block and believes by adding three-level shortcut branch roads
Breath, serves and suppresses the effect that gradient disappears.
2nd, because face age estimated data collection image scale is relatively small, the problem of over-fitting is produced easily in training,
This patent carries out pre-training using three-level residual error network to ImageNet data sets, and mould is expressed with the essential characteristic for obtaining image
Type, and then be finely adjusted on face age estimated data collection using pre-training model, so as to alleviate the directly training face age
The over-fitting problem that estimated data collection brings, improves the degree of accuracy of face age estimation.
3rd, because three-level residual error network broadens network while extra shortcut branch roads are added, and add more
Training parameter, may cause over-fitting problem more serious.This patent uses random depth algorithm when model is finely tuned and trained, with
Machine disconnects some residual errors and maps branch road so as to change path of information flow, and over-fitting and gradient disappearance problem are suppressed to reach,
Improve the purpose of face age accuracy of estimation.
Experimental analysis：
In order to illustrate the advantage of three-level residual error network, the Adience face age estimated data collection under the conditions of untethered
On carried out the face age estimate experiment.
Pre-training is carried out on ImageNet data sets using three-level residual error network first, 18 layers and 34 layers have been respectively adopted
Basic residual error network and 18 layers and 34 layers of three-level residual error network are tested, the ImageNet authentication error rates result such as institute of table 1
Show.From experimental result, the three-level residual error network that this patent is proposed can drop compared with the basic residual error network of the identical number of plies
The classification error rate of low ImageNet data sets, it was demonstrated that three-level residual error network more basic residual error network has more preferable network science
Habit ability.
The ImageNet classification results of table 1
Secondly, tested on the Adience face age estimated data collection under the conditions of untethered, be respectively adopted 34 layers
Basic residual error network and 34 layers of three-level residual error network, every kind of network are respectively adopted three kinds of strategies and are tested, and three kinds of strategies are respectively
For：1. directly train, directly model is finely adjusted on Adience after 2.ImageNet pre-training, 3.ImageNet is instructed in advance
Model is finely adjusted on Adience using random depth algorithm (SD) after white silk.Face age estimated result is as shown in table 2,
From experimental result, three-level residual error network is in the instruction combined with the fine setting training of random depth algorithm using ImageNet pre-training
The highest classification degree of accuracy is obtained when practicing method, it was demonstrated that the validity of this patent institute extracting method.
The Adience character classification by age results of table 2
Claims (2)
1. a kind of facial image age estimation method based on three-level residual error network, it is characterized in that, methods described is first basic
Three-level residual error network is set up on the basis of residual error network frame；Then ImageNet data sets are carried out using three-level residual error network pre-
Training, obtains ImageNet residual error network models；Again to acquired on the face age data collection under the conditions of untethered
ImageNet residual error network models are finely adjusted training；The three-level residual error network for finally being trained using fine setting carries out facial image
Age is estimated.
2. a kind of facial image age estimation method based on three-level residual error network according to claim 1, it is characterized in that,
The described method comprises the following steps：
A. three-level residual error network is set up
The model based on basic residual error network model (ResNets), adds shortcut outside all residual blocks first
Branch road, referred to as first order shortcut branch roads；Then shortcut branch roads, referred to as second are added on the basis of each residual block group
Level shortcut branch roads；Shortcut branch roads in raw residual block are referred to as third level shortcut branch roads, the first order and
Two grades of shortcut branch roads include a convolutional layer；
B. pre-training is carried out to ImageNet data sets using three-level residual error network, obtains ImageNet three-level residual error networks and instruct in advance
Practice model, during model pre-training, three-level residual error network is carried out initially using the basic residual error network model trained
Change, i.e., convolutional layer in raw residual block, BN layers initialized by basic residual error network model, first and second grades
Convolutional layer on shortcut branch roads is initialized using MSRA methods；
C. acquired ImageNet residual errors network model is finely adjusted on the face age data collection under the conditions of untethered
Training, random depth algorithm is used in model fine setting training process, i.e., disconnect the mapping of some residual errors at random when training is finely tuned
Branch road, to change path of information flow, reaches the purpose for suppressing that over-fitting and gradient disappear；
D. the estimation of facial image age is carried out with the three-level residual error network for training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611264499.2A CN106919897B (en) | 2016-12-30 | 2016-12-30 | Human face image age estimation method based on three-level residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611264499.2A CN106919897B (en) | 2016-12-30 | 2016-12-30 | Human face image age estimation method based on three-level residual error network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106919897A true CN106919897A (en) | 2017-07-04 |
| CN106919897B CN106919897B (en) | 2020-05-22 |
Family
ID=59453371
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611264499.2A Active CN106919897B (en) | 2016-12-30 | 2016-12-30 | Human face image age estimation method based on three-level residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106919897B (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107247947A (en) * | 2017-07-07 | 2017-10-13 | 北京智慧眼科技股份有限公司 | Face character recognition methods and device |
| CN107545245A (en) * | 2017-08-14 | 2018-01-05 | 中国科学院半导体研究所 | A kind of age estimation method and equipment |
| CN107633520A (en) * | 2017-09-28 | 2018-01-26 | 福建帝视信息科技有限公司 | A kind of super-resolution image method for evaluating quality based on depth residual error network |
| CN108108499A (en) * | 2018-02-07 | 2018-06-01 | 腾讯科技（深圳）有限公司 | Face retrieval method, apparatus, storage medium and equipment |
| CN108108677A (en) * | 2017-12-12 | 2018-06-01 | 重庆邮电大学 | One kind is based on improved CNN facial expression recognizing methods |
| CN108280436A (en) * | 2018-01-29 | 2018-07-13 | 深圳市唯特视科技有限公司 | A kind of action identification method based on the multi-grade remnant network for stacking recursive unit |
| CN109065033A (en) * | 2018-09-19 | 2018-12-21 | 华南理工大学 | A kind of automatic speech recognition method based on random depth time-delay neural network model |
| CN109271884A (en) * | 2018-08-29 | 2019-01-25 | 厦门理工学院 | Face character recognition methods, device, terminal device and storage medium |
| CN109345476A (en) * | 2018-09-19 | 2019-02-15 | 南昌工程学院 | High spectrum image super resolution ratio reconstruction method and device based on depth residual error network |
| CN109636780A (en) * | 2018-11-26 | 2019-04-16 | 深圳先进技术研究院 | Breast density automatic grading method and device |
| CN109934761A (en) * | 2019-01-31 | 2019-06-25 | 中山大学 | Jpeg image steganalysis method based on convolutional neural networks |
| CN110458084A (en) * | 2019-08-06 | 2019-11-15 | 南京邮电大学 | A kind of face age estimation method based on inversion residual error network |
| CN110717401A (en) * | 2019-09-12 | 2020-01-21 | Oppo广东移动通信有限公司 | Age estimation method and device, equipment and storage medium |
| CN111814611A (en) * | 2020-06-24 | 2020-10-23 | 重庆邮电大学 | Multi-scale face age estimation method and system embedded with high-order information |
| CN111860039A (en) * | 2019-04-26 | 2020-10-30 | 四川大学 | Cross-connection CNN + SVR-based street space quality quantification method |
| CN113506620A (en) * | 2021-01-20 | 2021-10-15 | 夏凤兰 | Medical management system and method based on block chain |
| CN113537382A (en) * | 2021-07-29 | 2021-10-22 | 中国科学技术大学 | Equipment diagnosis method and system based on residual shrinkage network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824054A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded depth neural network-based face attribute recognition method |
| CN104463172A (en) * | 2014-12-09 | 2015-03-25 | 中国科学院重庆绿色智能技术研究院 | Face feature extraction method based on face feature point shape drive depth model |
| CN105975916A (en) * | 2016-04-28 | 2016-09-28 | 西安电子科技大学 | Age estimation method based on multi-output convolution neural network and ordered regression |
| CN106228162A (en) * | 2016-07-22 | 2016-12-14 | 王威 | A kind of quick object identification method of mobile robot based on degree of depth study |
-
2016
- 2016-12-30 CN CN201611264499.2A patent/CN106919897B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824054A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded depth neural network-based face attribute recognition method |
| CN104463172A (en) * | 2014-12-09 | 2015-03-25 | 中国科学院重庆绿色智能技术研究院 | Face feature extraction method based on face feature point shape drive depth model |
| CN105975916A (en) * | 2016-04-28 | 2016-09-28 | 西安电子科技大学 | Age estimation method based on multi-output convolution neural network and ordered regression |
| CN106228162A (en) * | 2016-07-22 | 2016-12-14 | 王威 | A kind of quick object identification method of mobile robot based on degree of depth study |
Non-Patent Citations (1)
| Title |
|---|
| MASOUD ABDI AND SAEID NAHAVANDI: ""Multi-Residual Networks"", 《ARXIV》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107247947A (en) * | 2017-07-07 | 2017-10-13 | 北京智慧眼科技股份有限公司 | Face character recognition methods and device |
| CN107247947B (en) * | 2017-07-07 | 2021-02-09 | 智慧眼科技股份有限公司 | Face attribute identification method and device |
| CN107545245A (en) * | 2017-08-14 | 2018-01-05 | 中国科学院半导体研究所 | A kind of age estimation method and equipment |
| CN107633520A (en) * | 2017-09-28 | 2018-01-26 | 福建帝视信息科技有限公司 | A kind of super-resolution image method for evaluating quality based on depth residual error network |
| CN108108677A (en) * | 2017-12-12 | 2018-06-01 | 重庆邮电大学 | One kind is based on improved CNN facial expression recognizing methods |
| CN108280436A (en) * | 2018-01-29 | 2018-07-13 | 深圳市唯特视科技有限公司 | A kind of action identification method based on the multi-grade remnant network for stacking recursive unit |
| CN108108499A (en) * | 2018-02-07 | 2018-06-01 | 腾讯科技（深圳）有限公司 | Face retrieval method, apparatus, storage medium and equipment |
| CN108108499B (en) * | 2018-02-07 | 2023-05-26 | 腾讯科技（深圳）有限公司 | Face retrieval method, device, storage medium and equipment |
| CN109271884A (en) * | 2018-08-29 | 2019-01-25 | 厦门理工学院 | Face character recognition methods, device, terminal device and storage medium |
| CN109345476A (en) * | 2018-09-19 | 2019-02-15 | 南昌工程学院 | High spectrum image super resolution ratio reconstruction method and device based on depth residual error network |
| CN109065033B (en) * | 2018-09-19 | 2021-03-30 | 华南理工大学 | Automatic speech recognition method based on random deep time delay neural network model |
| CN109065033A (en) * | 2018-09-19 | 2018-12-21 | 华南理工大学 | A kind of automatic speech recognition method based on random depth time-delay neural network model |
| CN109636780A (en) * | 2018-11-26 | 2019-04-16 | 深圳先进技术研究院 | Breast density automatic grading method and device |
| CN109934761A (en) * | 2019-01-31 | 2019-06-25 | 中山大学 | Jpeg image steganalysis method based on convolutional neural networks |
| CN109934761B (en) * | 2019-01-31 | 2022-11-29 | 中山大学 | JPEG image steganalysis method based on convolutional neural network |
| CN111860039A (en) * | 2019-04-26 | 2020-10-30 | 四川大学 | Cross-connection CNN + SVR-based street space quality quantification method |
| CN111860039B (en) * | 2019-04-26 | 2022-08-02 | 四川大学 | Cross-connection CNN + SVR-based street space quality quantification method |
| CN110458084B (en) * | 2019-08-06 | 2022-08-26 | 南京邮电大学 | Face age estimation method based on inverted residual error network |
| CN110458084A (en) * | 2019-08-06 | 2019-11-15 | 南京邮电大学 | A kind of face age estimation method based on inversion residual error network |
| CN110717401A (en) * | 2019-09-12 | 2020-01-21 | Oppo广东移动通信有限公司 | Age estimation method and device, equipment and storage medium |
| CN111814611A (en) * | 2020-06-24 | 2020-10-23 | 重庆邮电大学 | Multi-scale face age estimation method and system embedded with high-order information |
| CN113506620A (en) * | 2021-01-20 | 2021-10-15 | 夏凤兰 | Medical management system and method based on block chain |
| CN113537382A (en) * | 2021-07-29 | 2021-10-22 | 中国科学技术大学 | Equipment diagnosis method and system based on residual shrinkage network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106919897B (en) | 2020-05-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106919897A (en) | A kind of facial image age estimation method based on three-level residual error network | |
| CN109145939B (en) | Semantic segmentation method for small-target sensitive dual-channel convolutional neural network | |
| CN109523463A (en) | A kind of face aging method generating confrontation network based on condition | |
| CN102201236B (en) | Speaker recognition method combining Gaussian mixture model and quantum neural network | |
| CN105678248B (en) | Face key point alignment algorithm based on deep learning | |
| CN103544483B (en) | A kind of joint objective method for tracing based on local rarefaction representation and system thereof | |
| CN109584251A (en) | A kind of tongue body image partition method based on single goal region segmentation | |
| CN107506702A (en) | Human face recognition model training and test system and method based on multi-angle | |
| CN105956560A (en) | Vehicle model identification method based on pooling multi-scale depth convolution characteristics | |
| CN107368845A (en) | A kind of Faster R CNN object detection methods based on optimization candidate region | |
| CN108876774A (en) | A kind of people counting method based on convolutional neural networks | |
| CN106096535A (en) | A kind of face verification method based on bilinearity associating CNN | |
| CN105787482A (en) | Specific target outline image segmentation method based on depth convolution neural network | |
| CN106778584A (en) | A kind of face age estimation method based on further feature Yu shallow-layer Fusion Features | |
| CN107123123A (en) | Image segmentation quality evaluating method based on convolutional neural networks | |
| CN109583379A (en) | A kind of pedestrian's recognition methods again being aligned network based on selective erasing pedestrian | |
| CN106250829A (en) | Digit recognition method based on lip texture structure | |
| CN106780482A (en) | A kind of classification method of medical image | |
| CN106127108A (en) | A kind of staff image region detection method based on convolutional neural networks | |
| CN108052884A (en) | A kind of gesture identification method based on improvement residual error neutral net | |
| CN108665485A (en) | A kind of method for tracking target merged with twin convolutional network based on correlation filtering | |
| CN106022363A (en) | Method for recognizing Chinese characters in natural scene | |
| CN106980812A (en) | Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net | |
| CN111259950B (en) | Method for training YOLO neural network based on 3D model | |
| CN107301376A (en) | A kind of pedestrian detection method stimulated based on deep learning multilayer |
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 |