CN108960123A - A kind of age estimation method - Google Patents
A kind of age estimation method Download PDFInfo
- Publication number
- CN108960123A CN108960123A CN201810691274.8A CN201810691274A CN108960123A CN 108960123 A CN108960123 A CN 108960123A CN 201810691274 A CN201810691274 A CN 201810691274A CN 108960123 A CN108960123 A CN 108960123A
- Authority
- CN
- China
- Prior art keywords
- age
- life
- feature
- face
- year
- 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
- G06V40/168—Feature extraction; Face representation
-
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
Abstract
The invention discloses a kind of age estimation methods, comprising the following steps: (1) extracts face feature by Face datection algorithm to the picture that data are concentrated；(2) face feature of extraction is divided into teenager, youth, the prime of life, the more robust years of one's life, reaches year, middle age and old age, wherein juvenile and old feature is directly extracted by linear SVM, and youth, the more robust years of one's life, up to five age level features of year and middle age uses the AGRC coding framework that joined time-constrain to be decoded at the prime of life.The accuracy rate of age estimation is effectively increased using the method for the present invention.
Description
Technical field
The invention belongs to technical field of image information processing, are related to a kind of age estimation method.
Background technique
Recognition of face and retrieval are always a popular topic in computer vision and MultiMedia Field.There are four types of main
Influence the principal element of recognition of face: posture, illumination, expression, age.Recently, with the accuracy of face key point location with
And the raising of Computing ability, Many researchers under no constraint environment (such as LFW face database) recognition of face
Achieve the effect identified close to the mankind.But LFW face database only changes in posture, illumination, expression, in face
Age on vary less.
Face changes highly significant with the increase at age, so the identification and matching in change of age human face are suitable
Difficult.Although age estimation and emulation aspect have research very much, the recognition of face and retrieval within annual age are still
Seldom.
Summary of the invention
The purpose of the present invention is to solve defects existing in the prior art, provide one kind and carry out people under change of age
Face identification and matched technology.
In order to achieve the above object, the present invention provides a kind of age estimation methods, comprising the following steps:
(1) face feature is extracted by Face datection algorithm to the picture that data are concentrated；
(2) by the face feature of extraction be divided into teenager, youth, the prime of life, the more robust years of one's life, up to year, middle age and old age, wherein it is juvenile and
Old feature is directly extracted by linear SVM, and youth, the more robust years of one's life, reaches year and middle aged five age levels at the prime of life
Feature is decoded using AGRC coding framework；
The AGRC coding framework is decoded, and steps are as follows:
(a) it calculates youth, the prime of life, the more robust years of one's life, reach year reference encoder matrix corresponding with middle aged five stages；
(b) granny rag Lars time matrix is utilized, increases time-constrain on the basis of matrix in granny rag Lars, so that young, strong
Year, the more robust years of one's life, be reduced to up to the change of age inside year and middle aged five stages it is minimum；
(c) final feature is obtained using maximum pond.
Wherein, in step (1) Face datection algorithm using JDA algorithm (joint cascade Face datection and registration Algorithm,
Joint Cascade Face Detection and Alignment)。
Specifically, face characteristic is extracted by following steps in step (1): extracting face 5 using JDA algorithm
Characteristic point: canthus, nose and sharp-tongued；And for five characteristic points of every face, using STCSR algorithm, (space-time cascades shape and returns
The face shape track algorithm returned, Spatio-Temporal Cascade Shape Regression), it will in conjunction with facial pose
It is extended to 16 characteristic points；Then to each feature point extraction local feature: to each characteristic point side of 5 different scales
Block goes to cover, and each square is divided into the grid of 4*4, and the uniform LBP feature of 59 dimensions is extracted from each grid, from same
The feature that one feature point extraction comes out is series connected as the description to each characteristic point, the characteristic dimension of each characteristic point
For 4720 dimensions, resulting 4720 dimension is then reduced to by 500 dimensions using Principal Component Analysis.
The present invention has the advantage that compared with prior art
The present invention is by carrying out divided stages for the age of people, wherein being not necessarily to since juvenile and old feature is obvious
By complicated feature extraction, directly by linear SVM (Linear Support Vector Machine,
LinSVM) can obtain it is preferable as a result, for other youth, the prime of life, the more robust years of one's life, up to passing through in year and middle aged five age levels
The AGRC coding framework (age bracket reference encoder frame algorithm, Age Groups Reference Coding) that text proposes, so that
16 face features that each stage extracts are in the same age stage (such as young stage) as close as in all ages and classes
It in the stage (such as young stage and middle aged stage), keeps off as much as possible, so that the accuracy rate of age estimation is improved, compared to
Without using AGRC frame, the accuracy rate of age estimation, which has directly obtained, to be significantly improved.
Detailed description of the invention
Fig. 1 is the basic flow chart of age estimation method of the present invention；
Fig. 2 is AGRC reference set of the present invention, feature relevant to age characteristics is extracted from the image of age characteristics, so
Age estimation is carried out afterwards；
Fig. 3 is the method for the present invention age data pretreatment operation JDA Face datection process；
Fig. 4 is STCSR face registration Algorithm of the present invention；
Fig. 5 is the master drawing for the CACD database that the present invention uses；
Fig. 6 is the location drawing of the present invention using 16 points of STCSR detection.
Specific embodiment
As shown in Figure 1, age estimation method detailed process of the present invention is as follows:
Using a large amount of famous person's information provided on the net, making reference set (is simply to be divided CACD data set
Class), as shown in Figure 2.
For each picture in data set, we first have to find out face from picture with people's face detection algorithm
The region at place.We use algorithm of the JDA algorithm as Face datection, and extract 5 characteristic points, five characteristic points point
It Wei not canthus, nose and sharp-tongued (as shown in Figure 3).As shown in figure 4, we are basic herein for five characteristic points of every face
On with STCSR algorithm combination facial pose 5 characteristic points are extended to 16 characteristic points, as shown in Figure 6.
After face's registration, we extract local feature from each characteristic point.In all local features, higher-dimension
Local binary patterns (High-Dimensional Local Binary Pattern, HD-LBP) face verification field
Achieved extraordinary effect.So there is employed herein a similar methods for extracting feature, in 16 characteristic points, I
Each characteristic point is gone to cover with the square of 5 different scales, each square is divided into the grid of 4*4, we are from each side
The uniform LBP feature of 59 dimensions is extracted in lattice.The feature come out from the same feature point extraction is series connected as right
The description of each characteristic point.The characteristic dimension of each characteristic point is 4720 dimensions.We use Principal Component Analysis (Principal
Component Analysis, PCA) by it is resulting 4720 dimension be reduced to 500 dimensions, feature is further processed behind aspect.
Then we using AGRC coding framework proposed in this paper by feature be decoded into youth, the prime of life, the more robust years of one's life, up to year and in
Year five age characteristics.AGRC mainly includes the steps that following three is main:
1, it calculates youth, the prime of life, the more robust years of one's life, reach year reference encoder matrix (word corresponding with middle aged five stages
Allusion quotation).
2, using granny rag Lars time matrix, it joined time change on the basis of original spatial variations expression formula, make
The young, prime of life, the more robust years of one's life, be reduced to up to the change of age inside year and middle aged five stages it is minimum.
3, final feature is obtained using maximum pond.
One, with reference to set representations:
By using the local feature extracted from the picture of reference crowd, we can be used following equation and calculate
Out with reference to expression collection:
Wherein,It is that the reference of i-th of people, k-th characteristic point in jth year indicates, d is characteristic dimension, n, m,
P is quantity, the range at age, the number of characteristic point with reference to personnel respectively.Its calculate the same person in same year some
All features of key point are divided by total number of persons.NijIt is the total number of persons of picture.Because CACD data set (as shown in Figure 5) is from interconnection
It is searched in net, may include noise.So taking mean value that can allow it to noise data with more Shandong for reference personnel
Stick.
Two, feature is decoded from reference space:
C(j,k)It is the d × n matrix indicated comprising n with reference to personnel, i.e.,
The feature that each characteristic point is extracted at k-th of characteristic point is x(k), we want using new to decode with reference to indicating
Feature.For this purpose, we define a new vector α(j,k)∈Rn×1To indicate the spy in the jth year extracted from k characteristic point
It levies (as shown in Figure 2).It is not difficult to find out that if if feature x(k)Relatively so with i-th of related personnelIt will be very
Greatly, on the contrary then can be small.We are solved with a least square problem comprising lucky big vast promise husband's regular terms:
But the relationship in this expression formula between the correlation table person of leting others have a look at there is no time-constrain, i.e., as a people and one
With reference to personnel when jth year is close, he there is a high likelihood that with this with reference to personnel jth -1 year and j+1 also close to, then
Age bracket is divided into 5 age brackets (youth, the more robust years of one's life, reaches year, middle age at the prime of life) by us, we add inside five stages respectively
Bound term is between added-time come the time-constrain that reflects in our coding frameworks.
It is as follows that we define a triple diagonal matrix L first:
L is one and makesWithAll very close to granny rag Lars matrix, its effect enables to
Difference afterwards between the two is smaller.It might as well enable
So we increase the constraint of time on the basis of original, and expression formula may be expressed as:
First item be in order to ensure error is enough small, Section 2 be in order to allow the same age stage people age characteristics
It is similar as far as possible.
Solve equation (6) very simply, because this is the least square problem of a L2 regularization, we can define two
A new matrixWithIt is as follows:
Wherein the birth age of personnel is respectively 1951-1990 in CACD data set, and the time for acquiring photo is 2004-
It 2013, then showing that the age photo section that the database is guaranteed replacement is 15-63 years old, is total up to 49 years.
Wherein 15-63 years old age was divided into 5 groups by us, as shown in table 4:
25 age levels of table
Note: being the time span of the age bracket in bracket behind age bracket
Wherein we can indicate formula (7) are as follows:
WithIt respectively indicates youth, the prime of life, the more robust years of one's life, reach year and middle age.For each
A, we can respectively indicate again:
It is young:
The prime of life:
The more robust years of one's life:
Up to year:
Middle age:
To matrixBe further processed, will wherein the 10th, 11,12,20,21,22,30,31,32,40,41,42 rows own
Non-zero element is all replaced into 0, obtains a new matrixThe purpose of so matrix is only in five age levels
Portion carries out time-constrain and does not change so that the difference inside feature is smaller to the feature other than age level.
Meanwhile our definition vectors
So we can expression (16) again:
The wherein analytic solutions of the equation are as follows:
If we defineSo we are, it can be seen that in factIt can be with
Regard as?Projection above, that is, a given picture, we can pass throughEfficiently to inquire corresponding sky
Between reference set.
Three, representative is never selected in the same year:
We expect the expression that space is referred between not the same year, and we used maximum ponds here to obtain this mesh
Mark:
By using maximum pond, last expression collection will be one with reference to maximum influence factor in personnel, be extracted
Age characteristics out reduces the influence with change of age in age level, to improve between age bracket and age bracket
Discrimination substantially increases the accuracy of age estimation to a certain extent.
Age estimation is carried out using AGRC frame to the present invention and existing method Non-AGRC carries out age estimation (this method
It is similar with the method for the present invention, only when using granny rag Lars matrix, it is added without time-constrain) compare, comparing result is as follows
Shown in table:
The comparison of 1 CACD database classification accuracy of table
| Method | Classification accuracy |
| Non-AGRC | 62.9 |
| AGRC | 78.8 |
As can be seen from the above table, compared to AGRC frame is not used, the accuracy rate that the present invention carries out age estimation is directly obtained
To significantly improving.
Claims (3)
1. a kind of age estimation method, it is characterised in that: the age estimation method the following steps are included:
(1) face feature is extracted by Face datection algorithm to the picture that data are concentrated；
(2) face feature of extraction is divided into teenager, youth, the prime of life, the more robust years of one's life, reaches year, middle age and old age, wherein juvenile and old
Feature is directly extracted by linear SVM, and youth, the more robust years of one's life, reaches year and middle aged five age level features at the prime of life
It is decoded using AGRC coding framework；
The AGRC coding framework is decoded, and steps are as follows:
(a) it calculates youth, the prime of life, the more robust years of one's life, reach year reference encoder matrix corresponding with middle aged five stages；
(b) granny rag Lars time matrix is utilized, increases time-constrain on the basis of matrix in granny rag Lars, so that youth, the prime of life, Sheng
Year, be reduced to up to the change of age inside year and middle aged five stages it is minimum；
(c) final feature is obtained using maximum pond.
2. age estimation method according to claim 1, it is characterised in that: Face datection algorithm is adopted in the step (1)
With JDA algorithm.
3. age estimation method according to claim 2, it is characterised in that: in the step (1) face characteristic by with
Lower step extracts: extracting 5 characteristic points of face using JDA algorithm: canthus, nose and sharp-tongued；And for every face
Five characteristic points are extended to 16 characteristic points using STCSR algorithm combination facial pose；Then to each feature point extraction
Local feature: going to cover to each characteristic point with the square of 5 different scales, and each square is divided into the grid of 4*4, from each
The uniform LBP feature that 59 dimensions are extracted in grid, the feature come out from the same feature point extraction are series connected conduct
Description to each characteristic point, the characteristic dimension of each characteristic point are 4720 dimensions, then will be resulting using Principal Component Analysis
4720 dimensions are reduced to 500 dimensions.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810691274.8A CN108960123A (en) | 2018-06-28 | 2018-06-28 | A kind of age estimation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810691274.8A CN108960123A (en) | 2018-06-28 | 2018-06-28 | A kind of age estimation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108960123A true CN108960123A (en) | 2018-12-07 |
Family
ID=64487938
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810691274.8A Pending CN108960123A (en) | 2018-06-28 | 2018-06-28 | A kind of age estimation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108960123A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113627520A (en) * | 2021-08-09 | 2021-11-09 | 中南大学 | Scale selection and noise robustness improved local binary pattern texture description method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106022287A (en) * | 2016-05-27 | 2016-10-12 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Over-age face verification method based on deep learning and dictionary representation |
| CN106650653A (en) * | 2016-12-14 | 2017-05-10 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Method for building deep learning based face recognition and age synthesis joint model |
-
2018
- 2018-06-28 CN CN201810691274.8A patent/CN108960123A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106022287A (en) * | 2016-05-27 | 2016-10-12 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Over-age face verification method based on deep learning and dictionary representation |
| CN106650653A (en) * | 2016-12-14 | 2017-05-10 | 广东顺德中山大学卡内基梅隆大学国际联合研究院 | Method for building deep learning based face recognition and age synthesis joint model |
Non-Patent Citations (2)
| Title |
|---|
| BOR-CHUN CHEN等: "Face Recognition and Retrieval Using Cross-Age Reference Coding With Cross-Age Celebrity Dataset", 《IEEE TRANSACTIONS ON MULTIMEDIA》 * |
| TIMO OJALA等: "Multiresolution Gray-Scale and Rotation Invariant Texture Classification with Local Binary Patterns", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113627520A (en) * | 2021-08-09 | 2021-11-09 | 中南大学 | Scale selection and noise robustness improved local binary pattern texture description method |
| CN113627520B (en) * | 2021-08-09 | 2023-09-05 | 中南大学 | Texture description method for improved local binary mode of scale selection and noise robustness |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yang et al. | Learning face age progression: A pyramid architecture of gans | |
| Zhang et al. | Joint low-rank and sparse principal feature coding for enhanced robust representation and visual classification | |
| CN105184298B (en) | A kind of image classification method of quick local restriction low-rank coding | |
| CN109522956B (en) | Low-rank discriminant feature subspace learning method | |
| CN105825183B (en) | Facial expression recognizing method based on partial occlusion image | |
| KR20120086728A (en) | Automatically mining person models of celebrities for visual search applications | |
| Sahbi et al. | A Hierarchy of Support Vector Machines for Pattern Detection. | |
| Huang et al. | Geometry-aware similarity learning on SPD manifolds for visual recognition | |
| CN112464730B (en) | Pedestrian re-identification method based on domain-independent foreground feature learning | |
| CN111339960B (en) | Face recognition method based on discrimination low-rank regression model | |
| CN104392246A (en) | Inter-class inner-class face change dictionary based single-sample face identification method | |
| Liao et al. | Face recognition based on dictionary learning and subspace learning | |
| CN112115881B (en) | Image feature extraction method based on robust identification feature learning | |
| CN105718898B (en) | Face age estimation method, system based on sparse undirected probability graph model | |
| Lai et al. | Visual speaker identification and authentication by joint spatiotemporal sparse coding and hierarchical pooling | |
| CN108090461A (en) | Three-dimensional face identification method based on sparse features | |
| CN108388918B (en) | Data feature selection method with structure retention characteristics | |
| CN108960123A (en) | A kind of age estimation method | |
| Wang et al. | Probabilistic nearest neighbor search for robust classification of face image sets | |
| CN110287973B (en) | Image feature extraction method based on low-rank robust linear discriminant analysis | |
| Zhao et al. | Multi-view dimensionality reduction via subspace structure agreement | |
| CN110287761A (en) | A kind of face age estimation method analyzed based on convolutional neural networks and hidden variable | |
| CN113887509B (en) | Rapid multi-modal video face recognition method based on image set | |
| Li et al. | Gabor Log-Euclidean Gaussian and its fusion with deep network based on self-attention for face recognition | |
| CN114943862A (en) | Two-stage image classification method based on structural analysis dictionary learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181207 |
|
| RJ01 | Rejection of invention patent application after publication |