CN104112147B - A kind of face feature extraction method based on nearest feature line - Google Patents
A kind of face feature extraction method based on nearest feature line Download PDFInfo
- Publication number
- CN104112147B CN104112147B CN201410360742.5A CN201410360742A CN104112147B CN 104112147 B CN104112147 B CN 104112147B CN 201410360742 A CN201410360742 A CN 201410360742A CN 104112147 B CN104112147 B CN 104112147B
- Authority
- CN
- China
- Prior art keywords
- matrix
- sample
- feature
- dimentional
- nearest
- 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.)
- Expired - Fee Related
Links
Landscapes
- Image Analysis (AREA)
Abstract
A kind of face feature extraction method based on nearest feature line, directly calculated using image matrix data, seek it is same as facial image sample be matrix the subpoint on similar facial image sample characteristic curve generated, and minimize Scatter Matrix in the class based on two-dimentional nearest feature line, and obtain a linear transformation, so that facial image sample is by after the linear transformation, divergence is minimum in the class based on two-dimentional nearest feature line.Compared with traditional nearest feature line space-wise, present invention reduces operands, while can also the correlation as much as possible that retain between image array adjacent pixels point itself.
Description
Technical field
The present invention relates to the authentication of people and identification more particularly to a kind of face characteristic extractions based on nearest feature line
Method and apparatus.
Background technique
Face recognition technology is an emerging biometrics identification technology, is the high-end skill of current international and domestic sciemtifec and technical sphere
Art.In the process of face recognition, feature extraction is a wherein important step.Before being identified to facial image,
It needs first from multiple facial images of each classification, the set of the feature of category facial image is extracted, during this
Used facial image sample is referred to as training sample.
When executing recognition of face, first takes with the same mode of training sample is extracted, extracted from target facial image
Then feature again matches extracted feature with the feature of the training sample of each classification, to find therewith most
Close training sample, and it is determined as classification belonging to face in target facial image.
Nearest feature line [1] is a kind of classifier, and core is characteristic curve measurement.If y1And y2For two in same category
A sample, yqIt is that sample to be sorted does not calculate y in nearest feature line classifierqAnd y1Distance, also do not calculate yqAnd y2's
Distance, but calculate yqTo super straight line y1y2Distance, this super straight line is just called characteristic curve, and characteristic curve can have class label,
Here the class label of characteristic curve is defined as y1And y2Classification.If ypFor yqProjection on straight line, then yqTo straight line y1y2
Distance be equal to yqTo ypDistance.Same reason can calculate yqTo the distance of all characteristic curves, yqWith that feature
The distance of line is nearest, then nearest feature line classifier is just by yqAssign to classification corresponding to this feature line.Here yp=y1+t
(y2-y1), t=< yq-y1,y2-y1>/<y2-y1,y2-y1>。
A kind of nearest feature line space [2] method based on nearest feature line exists in the prior art, this method can be used
To extract the feature of facial image.IfIt is that N pair is pre- by dimension normalization, gray scale stretching etc.
The M of processing1×M2Facial image, wherein M1With M2It is positive integer, then nearest feature line space-wise specifically includes following step
It is rapid:
Step a, N pair training sample image is converted into vector respectively, it is N number of right respectively with N pair training sample image to obtain
(the M answered1×M2The dimensional vector y of) × 1i, i=1,2 ..., N here.
Step b, subpoint of each training sample of calculating to generic characteristic curve.If yiFor i-th of training sample, here
I=1,2 ..., N.
Step c, Scatter Matrix in the class based on nearest feature line is calculated.
Here, ym, yn∈P(l(yi)) indicate ym, ynWith yiIn same category,Indicate yiIn ym, ynIt is generated super straight
Projection on line, |P(l(yi))| it indicatesyiThe number of samples of place classification.
Step d, the smallest d characteristic value of calculating matrix A and corresponding feature vector w1,w2,…,wd.Step e, W is enabled
=[w1,w2,…,wd],Here i=1,2 ..., N.fiAs the mentioned feature of nearest feature line space-wise.
For target facial image to be sorted, after also passing through the pretreatments such as dimension normalization, gray scale stretching,
And also pass through step a and be converted into vector r, the mentioned feature of nearest feature line space-wise is R=WTr。
[1].Li,S.Z.,Lu,J.:‘Face recognition using the nearest feature line
method’,IEEE Trans.Neural Networks,1999,10,pp.439-443。
[2].Pang,Y.,Yuan,Y.,Li,X.:‘Generalised nearest feature line for
subspacelearning’,Electronics Letters,2007,43,pp.1079-1080。
Existing method needs face image data to be converted into vector, increases operand, it is also possible to cause to face in image
The loss of correlation between nearly pixel.
Summary of the invention
In order to solve the problems, such as in the prior art, the present invention provides a kind of, and the face characteristic based on nearest feature line extracts dress
It sets, includes sample matrix module, Scatter Matrix constructing module and projection matrix constructing module；Sample matrix module: being used for will be defeated
The current face's image pattern matrixing entered, is stored in sample matrix module, becomes sample image matrix；Sample image matrix
Set constitute face database；Scatter Matrix constructing module: for calculating the sample image square based on two-dimentional nearest feature line
Divergence between battle array, and the Scatter Matrix of current face's database is constructed, sample image matrix herein is to be stored in sample
Sample image matrix in matrix module；Projection matrix constructing module: for the Scatter Matrix according to current face's database, structure
The optimal projection matrix of current face's database is made, and the facial image sample in face database is transformed into feature space,
As face sample characteristics matrix, it is then store in facial image feature database, it is special by facial image feature database output face sample
Levy matrix.
A kind of face feature extraction method based on nearest feature line, sample matrix module obtain N facial image samples
Data setUsing the definition of nearest feature line, matrix regards the one of high dimension linear space as
It is a, thenHere the definition of two-dimentional nearest feature line is provided: one given
Image pattern to be sortedThe image pattern is in same category of two sample Xi,XjOn characteristic curve generated
Subpoint be
Xp=Xi+t(Xj-Xi), (1)
Wherein
T=< Xq-Xi,Xj-Xi>M/<Xj-Xi,Xj-Xi>M；
Then XqDistance to this feature line is | | Xq-Xp||2；
X has been calculatedqTo after the distance of all characteristic curves, find apart from the smallest characteristic curve, then by XqAssign to respective class
Not；Here XqAnd XpIt is matrix；
The objective function of two-dimentional nearest feature line analysis are as follows:
WhereinIndicate XmBy XiAnd XjProjection on characteristic curve generated, l (Xi) indicate XiClass label, by
Above-mentioned formula is it is found that the target of two-dimentional nearest feature line analysis is to minimize divergence in the class based on two-dimentional nearest feature line, warp
Cross matrix calculating, it is known that:
Here
The matrix of the composition of feature vector corresponding to the smallest characteristic value of matrix S in this way is exactly two-dimentional nearest feature line
Analyze the optimal transform matrix to be looked for.
As a further improvement of the present invention, the step of two-dimentional nearest feature line analysis method is as follows:
IfTo pass through the pre- places such as dimension normalization, gray scale stretching in sample matrix module
Data set composed by N number of training sample after reason:
Step a, using the definition of two-dimentional nearest feature line, according to formula (1), each facial image sample is calculated to similar
The subpoint of characteristic curve；
Step b, Scatter Matrix in the class based on two-dimentional nearest feature line is calculated
Here, Xm,Xn∈P(l(Xi)) indicate Xm,XnWith XiIn same category,Indicate XiIn Xm,XnIt is generated super straight
Projection on line, | Pl (Xi)) | indicate XiThe number of samples of place classification；
Step c, the smallest d characteristic value of calculating matrix S and corresponding feature vector w1,w2,…,wd；
Step d, W=[w is enabled1,w2,…,wd],Fi=WTXi, i=1,2 ..., N here.FiAs two-dimentional nearest feature line
The mentioned feature of analysis method；
For target facial image R to be sortedeFor, it is pre-processed by same dimension normalization, gray scale stretching etc.
Later, two-dimentional the mentioned feature of nearest feature line analysis method is FR=WTRe。
The beneficial effects of the present invention are:
In the invention patent, image is not converted into vector to carry out operation, but directly utilizes image matrix data
It is calculated, seeking it equally is the subpoint of matrix, and minimizes Scatter Matrix in class, it is desirable to a linear transformation is found,
Divergence in the class based on two-dimentional nearest feature line can be made minimum, to reduce operand, while can also be as more as possible
Retain image array adjacent pixels point itself between correlation.
Analysis of complexity:
Here in order to facilitate analysis of complexity, it is assumed that the dimension of each sample is m × m, enables D=m × m, while it is assumed that having
C classification, each classification have n training sample, then N=nc is total number of samples.Since multiplication compares consumption resource, here
The quantity of the multiplication of two methods is compared.Generally speaking, calculation amount is primarily present in identical three steps to both methods
In rapid.Firstly, both methods is required to find the subpoint of characteristic curve in all samples to class, either one-dimensional characteristic line also
It is two-dimentional nearest feature line, multiplication computation amount is the same, and is sought each subpoint and is needed Ο (m2), it needs to ask altogetherA subpoint, hereIt is the number of combinations that two are extracted from n-1 element, computation complexity is in two ways for institute
Ο(m2n3c).When and then calculating Scatter Matrix in class, in the two-dimentional nearest feature line analysis method that this patent is mentioned, due to
Scatter Matrix be a series of matrix products and, and wherein the computation complexity of each matrix product is Ο (m3), it is such
Matrix one is sharedIt is a, so total complexity is Ο (cn3m3), for one-dimensional characteristic space of lines, Scatter Matrix
It is the sum of matrix product as much, the computation complexity of each matrix product is Ο (D3), therefore total complexity is Ο
(cn3D3).Last two methods are both needed to solution characteristic equation, ask the characteristic value and feature vector of corresponding Scatter Matrix, calculate multiple
Miscellaneous degree is related with the dimension of Scatter Matrix, and for two-dimentional nearest feature line analysis method, Scatter Matrix dimension is m × m,
The complexity for solving characteristic equation is Ο (m3), for one-dimensional characteristic curve space-wise, Scatter Matrix dimension is D × D, is asked
The complexity for solving characteristic equation is Ο (D3).So by analysis above it is known that one-dimensional characteristic curve space-wise is total
Complexity is
Ο(m2n3c+cn3D3+D3)=Ο (n3cm2+(1+cn3)D3), total complexity of two-dimentional nearest feature line analysis method
Degree is Ο (m2n3c+cn3m3+m3)=Ο (n3cm2+(1+cn3)m3), due to D=m × m, so available conclusion here, two
The complexity of dimension nearest feature line analysis method is less than the complexity of one-dimensional characteristic space of lines method.
Detailed description of the invention
Fig. 1 is face feature deriving means structural schematic diagram of the present invention.
Specific embodiment
The present invention will be further described with reference to the accompanying drawing.
Face characteristic extraction element based on nearest feature line: include sample matrix module, Scatter Matrix constructing module and
Projection matrix constructing module.
Sample matrix module: current face's image pattern matrixing for that will input is stored in sample matrix module,
As sample image matrix；The set of sample image matrix constitutes face database；
Scatter Matrix constructing module: for calculating the divergence between the sample image matrix based on two-dimentional nearest feature line,
Sample image matrix herein is the sample image matrix being stored in sample matrix module, and constructs current face's database
Scatter Matrix；
Projection matrix constructing module: for the Scatter Matrix according to current face's database, current face's database is constructed
Optimal projection matrix, and the facial image sample in face database is transformed into feature space, becomes face sample characteristics
Matrix is then store in facial image feature database.Face sample characteristics matrix is exported by facial image feature database.
Face feature extraction method based on nearest feature line:
Sample matrix module has obtained N training sample face image data collectionThis
In provide a kind of feature extraction algorithm based on image array.In calculating process, do not need image array being converted to vector.
Using the definition of nearest feature line, in the definition of nearest feature line, sample number therein is not specifically noted
According to that should be vector or matrix, matrix data equally can be regarded as a point of high dimension linear space, thenHere provide the definition of two-dimentional nearest feature line: given one to be sorted
Image patternThe image pattern is in same category of two sample Xi,XjProjection on characteristic curve generated
It puts and is
Xp=Xi+t(Xj-Xi), (1)
Wherein
T=< Xq-Xi,Xj-Xi>M/<Xj-Xi,Xj-Xi>M。
Then XqDistance to this feature line is | | Xq-Xp||2.X has been calculatedqTo after the distance of all characteristic curves, find away from
From the smallest characteristic curve, then by XqAssign to respective classes.
Note that X hereqAnd XpIt is matrix.
The objective function for the two-dimentional nearest feature line analysis that this method is previously mentioned are as follows:
WhereinIndicate XmBy XiAnd XjProjection on characteristic curve generated, l (Xi) indicate XiClass label.By
Above-mentioned formula is it is found that the target of two-dimentional nearest feature line analysis is to minimize divergence in the class based on two-dimentional nearest feature line.Through
Cross matrix calculating, it is known that:
Here
The matrix of the composition of feature vector corresponding to the smallest characteristic value of matrix S in this way is exactly two-dimentional nearest feature line
Analyze the optimal transform matrix to be looked for.
The step of two-dimentional nearest feature line analysis method that this method is mentioned, is as follows:
IfTo pass through the pre- places such as dimension normalization, gray scale stretching in sample matrix module
Data set composed by N number of training sample after reason.
Step a, using the definition of two-dimentional nearest feature line, according to formula (1), each facial image sample is calculated to similar
The subpoint of characteristic curve.
Step b, Scatter Matrix in the class based on two-dimentional nearest feature line is calculated
Here, Xm,Xn∈P(l(Xi)) indicate Xm,XnWith XiIn same category,Indicate XiIn Xm,XnIt is generated super
Projection on straight line, | Pl (Xi)) | indicate XiThe number of samples of place classification.
Step c, the smallest d characteristic value of calculating matrix S and corresponding feature vector w1,w2,…,wd。
Step d, W=[w is enabled1,w2,…,wd],Fi=WTXi, i=1,2 ..., N here.FiAs two-dimentional nearest feature line
The mentioned feature of analysis method.
For target facial image R to be sortedeFor, it is pre-processed by same dimension normalization, gray scale stretching etc.
Later, two-dimentional the mentioned feature of nearest feature line analysis method is FR=WTRe。
The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be said that
Specific implementation of the invention is only limited to these instructions.For those of ordinary skill in the art to which the present invention belongs, exist
Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to of the invention
Protection scope.
Claims (2)
1. a kind of face feature extraction method based on nearest feature line, it is characterised in that:
Current face's image pattern matrixing of input is stored in sample matrix module, becomes sample by sample matrix module
Image array；The set of sample image matrix constitutes face database；
Scatter Matrix constructing module calculates the divergence between the sample image matrix based on two-dimentional nearest feature line, and constructs and work as
The Scatter Matrix of preceding face database, sample image matrix herein are the sample image square being stored in sample matrix module
Battle array；
Projection matrix constructing module constructs the optimal projection of current face's database according to the Scatter Matrix of current face's database
Matrix, and the facial image sample in face database is transformed into feature space, become face sample characteristics matrix, then deposits
Storage exports face sample characteristics matrix in facial image feature database, by facial image feature database；
Sample matrix module obtains N facial image sample data setsUtilize nearest feature
The definition of line, matrix data regard a point of high dimension linear space as, then
Here it provides the definition of two-dimentional nearest feature line: giving an image pattern to be sortedThe image pattern exists
Same category of two sample Xi,XjSubpoint on characteristic curve generated is
Xp=Xi+t(Xj-Xi), (1)
Wherein
T=< Xq-Xi,Xj-Xi>M/ < Xj-Xi,Xj-Xi>M；
Then XqDistance to this feature line is | | Xq-Xp||2；
X has been calculatedqTo after the distance of all characteristic curves, find apart from the smallest characteristic curve, then by XqAssign to respective classes；
Here XqAnd XpIt is matrix；
The objective function of two-dimentional nearest feature line analysis are as follows:
WhereinIndicate XmBy XiAnd XjProjection on characteristic curve generated, l (Xi) indicate XiClass label, by above-mentioned
Formula is it is found that the target of two-dimentional nearest feature line analysis is to minimize divergence in the class based on two-dimentional nearest feature line, by square
Battle array calculates, it is known that:
Here
The matrix of the composition of feature vector corresponding to the smallest characteristic value of matrix S in this way is exactly two-dimentional nearest feature line analysis
The optimal transform matrix to be looked for.
2. the face feature extraction method according to claim 1 based on nearest feature line, it is characterised in that: two dimension is recently
The step of feature provision of on-line analysis methodology, is as follows:
IfTo pass through the pretreatments such as dimension normalization, gray scale stretching in sample matrix module
Data set composed by N number of training sample afterwards:
Step a, each facial image sample is calculated to homogenous characteristics according to formula (1) using the definition of two-dimentional nearest feature line
The subpoint of line；
Step b, Scatter Matrix in the class based on two-dimentional nearest feature line is calculated
Here, Xm,Xn∈P(l(Xi)) indicate Xm,XnWith XiIn same category,Indicate XiIn Xm,XnOn super straight line generated
Projection, | Pl (Xi)) | indicate XiThe number of samples of place classification；
Step c, the smallest d characteristic value of calculating matrix S and corresponding feature vector w1,w2,L,wd；
Step d, W=[w is enabled1,w2,L,wd],Fi=WTXi, i=1,2 ..., N here；FiAs two-dimentional nearest feature line analysis side
The mentioned feature of method；
For target facial image R to be sortedeFor, after the pretreatments such as same dimension normalization, gray scale stretching,
Its two-dimentional mentioned feature of nearest feature line analysis method is FR=WTRe。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410360742.5A CN104112147B (en) | 2014-07-25 | 2014-07-25 | A kind of face feature extraction method based on nearest feature line |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410360742.5A CN104112147B (en) | 2014-07-25 | 2014-07-25 | A kind of face feature extraction method based on nearest feature line |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104112147A CN104112147A (en) | 2014-10-22 |
| CN104112147B true CN104112147B (en) | 2018-12-21 |
Family
ID=51708931
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410360742.5A Expired - Fee Related CN104112147B (en) | 2014-07-25 | 2014-07-25 | A kind of face feature extraction method based on nearest feature line |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104112147B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107680037B (en) * | 2017-09-12 | 2020-09-29 | 河南大学 | Improved face super-resolution reconstruction method based on nearest characteristic line manifold learning |
| CN109064454A (en) * | 2018-07-12 | 2018-12-21 | 上海蝶鱼智能科技有限公司 | Product defects detection method and system |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101324923A (en) * | 2008-08-05 | 2008-12-17 | 北京中星微电子有限公司 | Method and apparatus for extracting human face recognition characteristic |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102163279B (en) * | 2011-04-08 | 2012-11-28 | 南京邮电大学 | Color human face identification method based on nearest feature classifier |
| CN102402784B (en) * | 2011-12-16 | 2013-07-24 | 武汉大学 | Human face image super-resolution method based on nearest feature line manifold learning |
-
2014
- 2014-07-25 CN CN201410360742.5A patent/CN104112147B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101324923A (en) * | 2008-08-05 | 2008-12-17 | 北京中星微电子有限公司 | Method and apparatus for extracting human face recognition characteristic |
Non-Patent Citations (1)
| Title |
|---|
| 基于最近特征线的图像特征提取算法研究;闫立军;《万方数据》;20121225;第13-37页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104112147A (en) | 2014-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104657718B (en) | A kind of face identification method based on facial image feature extreme learning machine | |
| CN110188641A (en) | Image recognition and the training method of neural network model, device and system | |
| CN103544486B (en) | Human age estimation method based on self-adaptation sign distribution | |
| WO2016138838A1 (en) | Method and device for recognizing lip-reading based on projection extreme learning machine | |
| CN106991355A (en) | The face identification method of the analytical type dictionary learning model kept based on topology | |
| Zhang et al. | Unified detection method of aluminium profile surface defects: Common and rare defect categories | |
| CN110929836B (en) | Neural network training and image processing method and device, electronic equipment and medium | |
| Qian et al. | Robust low-rank regularized regression for face recognition with occlusion | |
| CN110321451A (en) | Image retrieval algorithm based on Distribution Entropy gain loss function | |
| CN106250909A (en) | A kind of based on the image classification method improving visual word bag model | |
| CN114821237A (en) | Unsupervised ship re-identification method and system based on multi-stage comparison learning | |
| Jia et al. | Fabric defect detection based on transfer learning and improved Faster R-CNN | |
| CN104112147B (en) | A kind of face feature extraction method based on nearest feature line | |
| Setyono et al. | Betawi traditional food image detection using ResNet and DenseNet | |
| Tran et al. | Improving the face recognition accuracy under varying illumination conditions for local binary patterns and local ternary patterns based on weber-face and singular value decomposition | |
| CN104318224A (en) | Face recognition method and monitoring equipment | |
| Gong et al. | Spectrum-irrelevant fine-grained representation for visible–infrared person re-identification | |
| Wang et al. | MsRAN: A multi-scale residual attention network for multi-model image fusion | |
| CN105868713A (en) | Method for identifying parallel features integrating human face expression based on nucleus LDA | |
| CN107657223B (en) | Face authentication method based on rapid processing multi-distance metric learning | |
| CN106485286B (en) | Matrix classification model based on local sensitivity discrimination | |
| CN114677535A (en) | Training method of domain-adaptive image classification network, image classification method and device | |
| Chen et al. | A GPU-accelerated approximate algorithm for incremental learning of Gaussian mixture model | |
| Tang et al. | Learning Hough regression models via bridge partial least squares for object detection | |
| Wang et al. | Research on pedestrian detection based on multi-level fine-grained YOLOX algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20181221 Termination date: 20190725 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |