CN104408405B - Face representation and similarity calculating method - Google Patents
Face representation and similarity calculating method Download PDFInfo
- Publication number
- CN104408405B CN104408405B CN201410604043.0A CN201410604043A CN104408405B CN 104408405 B CN104408405 B CN 104408405B CN 201410604043 A CN201410604043 A CN 201410604043A CN 104408405 B CN104408405 B CN 104408405B
- Authority
- CN
- China
- Prior art keywords
- image
- face
- face block
- specially
- facial image
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of face representations and similarity calculation algorithm.The face training set with mark is collected in off-line training step, everyone includes 50 100 facial images, the face block of 2*2 same size will be divided into after image normalization in set.To each face block, the fritter of dense extraction k*k pixels carries out mean value zero, after the pretreatment of variance normalizing, based on K means clustering learning convolution kernels.Each face block obtains K convolved image, to K*2*2 convolved image of whole input face, stretches the feature of all response diagrams after carrying out mean value Pooling operations and ReLU nonlinear operations respectively.To the feature after stretching, learn principal component analysis（PCA）Projection, after PCA reduces dimension, based on linear discriminant analysis（LDA）Algorithm study differentiates projection, obtains compact, robust face representation, and carrying out inner product operation to the face representation of two images obtains similarity.
Description
Technical field
The invention belongs to computer vision and technical field of image processing, expression and similarity meter in particular to face
Calculation method.
Background technology
Computer face identification refers to based on known Face Sample Storehouse, and image and pattern-recognition skill are analyzed using computer
For art from either statically or dynamically scene, identification or the one or more faces of verification extract the multiple of face using Feature Extraction Technology
Possible character representation, the technology are widely used in the important events such as public safety, proof of identification, effectively carry out face
The key of identification is quickly, accurately carries out face characteristic expression.There is the calculation of some face character representations in the prior art
Method carries out input facial image using convolution kernel process of convolution and obtains the multidimensional characteristic vectors expression of face, what operation was related to
Convolution kernel selection does not make full use of abundant face training data, and the dimension of feature vector sets too small influence precision, setting
Excessive influence speed, so as to constrain the accuracy of face characteristic expression and similarity calculation on the whole.
To overcome the above-mentioned shortcoming of the prior art, the present invention proposes a kind of face representation learnt based on convolution kernel
Algorithm and similarity calculating method, the convolution kernel obtained using training set off-line learning participate in the convolution operation of facial image
In, the dimensionality reduction and analysis matrix that are learnt using training set carry out human face discriminating feature extraction, are obtaining compact, robust people
After face represents, the similarity that can be quickly obtained between image, it is accurate to realize, fast to read, the face representation of robust and similar
Degree calculates.
Invention content
The present invention proposes a kind of face representation and similarity calculating method, it is characterised in that：
(1) generation network face training set is collected；
(2) it trains to obtain convolution kernel using the facial image in training set；
(3) convolution kernel is utilized, character representation is carried out to the facial image in training set；
(4) convolution kernel is utilized, character representation is carried out to input facial image；
(5) extraction differentiates feature in the facial image character representation in training set；
(6) extraction differentiates feature in facial image character representation is inputted；
(7) using feature is differentiated, image similarity is calculated.
Preferably, wherein step (1) is specially：C famous person, wherein c=1,2 ... ..., C are selected, famous person is corresponding with a figure
Picture, wherein 50<<100, the corresponding all image collections of famous person are to be denoted as training set,
Image in training set, wherein i=1,2 ... ..., N embody posture, illumination and the variation of expression.
Preferably, wherein step (2) is specially：Facial image in training set is aligned, the facial image after alignment
Size is 128*80 pixels, each image is divided into the identical face block of 2*2 size, each face block diagram is as size
64*40 pixels are respectively processed all face blocks of all images, obtain convolution kernel.
Preferably, wherein all face blocks to all images are specific the step of being respectively processed, obtain convolution kernel
For：The face block of each image corresponding position j is denoted as, wherein i=1,2 ... ... N, j=1, and 2,3,4, it is extracted on each face block
The fritter of k*k sizes, step-length 1, then the face block be expressed as：
, wherein 1<k<40, m=64-k+1, n=40-k+1 calculate, will subtract the mean value, and by the variance of every one-dimensional characteristic
Normalizing, the face block for obtaining each image corresponding position j represent：, the face block of all N width image corresponding position j is located successively
Reason, obtained N number of expression is connected, the face block for obtaining the position j of N width image in training set is expressed as：, to this progress
K-means clustering learnings obtain K cluster centre, are denoted as：, wherein dj, t expression t-th of cluster centre therein, t=1,
2 ... ... K, each cluster centre convolution kernel independent as one, to the face block of all position j of all images all in accordance with
Above-mentioned steps carry out convolution kernel study respectively, obtain, j=1,2,3,4 }.
Preferably, wherein step (3) is specially：To carrying out Pooling operations and ReLU after facial image convolution operation
It operates to carry out character representation.
Preferably, convolution operation therein is specially：By each cluster centre dj, t is as a convolution kernel, in progress
Product operation, two dimensional image is reduced to by the convolution vector stretched, to all in the facial image in training set, j=1, and 2,3,
4, above-mentioned inner product operation is carried out, the face block of same position corresponds to the convolution kernel that same position training obtains, is obtained 2*2*K
Convolved image, wherein the size of each convolved image is m*n.
Preferably, Pooling therein, which is operated, is specially：It operates to extract invariant features using mean value Pooling, it will
The convolved image of each m*n is divided into nonoverlapping grid of 4*4 pixel sizes, calculates the mean value of each grid, obtains mean value
Response diagram after Pooling, after Pooling is operated, the size of each response diagram is（m/4）*（n/4）.
Preferably, ReLU therein, which is operated, is specially：Negative sound in the response diagram exported after mean value Pooling is operated
It should be worth and return 0, retain positive response, operated by ReLU, add the non-linear of feature, for input, after ReLU is operated
All response diagrams stretch after obtained face block be expressed as, characteristic dimension K*（m/4）*（n/4）, 2*2 position is obtained
Face block expression be connected in turn, obtain the character representation of training set facial image, characteristic dimension is A=2*2*K*
（m/4）*（n/4）Dimension.
Preferably, wherein step (4) is specially：To carrying out Pooling and ReLU operations after facial image convolution operation
To carry out character representation.
Preferably, convolution operation therein is specially：To inputting facial image, it is aligned to the face figure of 128*80 pixels
Picture is divided into the identical face block of 2*2 size, and each face block diagram is 64*40 pixels as size, corresponding position j's
Face block is denoted as, j=1, and 2,3,4, the fritter of extraction k*k sizes on each face block, step-length 1, then face block expression
For：, wherein 1<k<40, m=64-k+1, n=40-k+1 calculate, will subtract the mean value, and the variance of every one-dimensional characteristic is returned
One, obtain the expression of the jth position face block of the image, by each cluster centre
Dj, t carry out inner product operation, the convolution vector stretched are reduced to two dimensional image, to inputting facial image as a convolution kernelYIn all above-mentioned inner product operations of carry out, j=1,2,3,4, the face block of same position corresponds to the volume that same position training obtains
Product core, is obtained 2*2*K convolved image, wherein the size of each convolved image is m*n.
Preferably, Pooling therein, which is operated, is specially：It operates to extract invariant features using mean value Pooling, it will
The convolved image of each m*n is divided into nonoverlapping grid of 4*4 pixel sizes, calculates the mean value of each grid, obtains mean value
Response diagram after Pooling, after Pooling is operated, the size of each response diagram is（m/4）*（n/4）.
Preferably, ReLU therein, which is operated, is specially：By the Negative Acknowledgment value in the response diagram exported after mean value Pooling
Return 0, retain positive response, operated by ReLU, add the non-linear of feature, for input, the institute after ReLU is operated
There is the face block that response diagram obtains after stretching to be expressed as, characteristic dimension K*（m/4）*（n/4）, people that 2*2 position is obtained
The expression of face block is connected in turn, and obtains the expression of input facial image, characteristic dimension is A=2*2*K*（m/4）*（n/4）
Dimension.
Preferably, wherein step (5) is specially：To facial image character representation by principal component analysis dimensionality reduction calculate and
Linear discriminant analysis obtains differentiating feature.
Preferably, wherein principal component analysis dimensionality reduction calculate the step of be specially：Calculate the association side of training set character representation
Poor matrix wherein being the mean value of the character representation of all N width images in training set, enables, wherein corresponding preceding r maximum eigenvalue
Corresponding feature vector, for input feature vector, it is obtained after principal component analysis dimensionality reduction, all face figures in training set
The character representation of picture is by dimensionality reduction postscript.
Preferably, wherein linear discriminant analysis is specially：Respectively calculate dimensionality reduction after character representation class scatter and class
Interior divergenceSW, wherein, it represents the mean value after the feature vector principal component analysis dimensionality reduction degree of person ownership image, is training set
In all images character representationLMean value, PCA_Dim=r, wherein being q feature of following generalized eigenvalues problem maximum
It is worth corresponding feature vector：, to carrying out differentiation feature extraction, obtain differentiating feature.
Preferably, wherein step（6）Specially：To facial image character representation by principal component analysis dimensionality reduction calculate and
Linear discriminant analysis obtains differentiating feature.
Preferably, principal component analysis dimensionality reduction therein calculate the specific steps are：To the character representation of input picture, meter
It calculates, after obtaining, differentiation feature is calculated.
Preferably, wherein step（7）Specially：It is obtained to carrying out two norm normalizings：, to carrying out two models
Number normalizing obtains：, carry out inner product calculating) obtain input pictureYWith
The similarity of training set facial image.
Description of the drawings
Fig. 1 is face representation learning process figure
Fig. 2 is the facial image after alignment
Fig. 3 is the convolution kernel example based on K-means clustering learnings
Fig. 4 is fast convolution process schematic.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with the accompanying drawings and the specific embodiments
The present invention is made a further detailed description.
Fig. 1 is the face representation learning process figure disclosed in the present invention, is described with reference to Fig. 1.
Step 1：Training set is set and normalization
A network face data set is collected, selects C famous person, wherein c=1,2 ... ..., C are corresponding with an image, wherein
50<<100, the corresponding all image collections of famous person are to be denoted as training set,
Image in training set should embody the variation of posture, illumination and expression.
Based on eye position, face is aligned, the facial image See Figure 2 after alignment, it is preferred that the people after alignment
It is bold small for 128*80 pixels.
Step 2：Face representation learns
（1）Facial image after alignment is divided into the face block of 2*2 same size, the face table of corresponding 128*80
Show, the size of each face block is 64*40 pixels.
The N images included in gathering training, the face block of every image corresponding position j are denoted as, wherein i=1,
2 ... ... N, j=1,2,3,4.
（2）Each face block is handled as follows：
On each face block extract k*k sizes fritter, step-length 1, then the face block can be expressed as：
Formula 1
Wherein m=64-k+1, n=40-k+1 are preferably provided with k=9, then m=56, n=32.
It calculates, the mean value will be subtracted, and by the variance normalizing of every one-dimensional characteristic, obtain the i-th width image
The expression of jth position face block：
Formula 2
（3）To being handled successively per piece image in training set, by step（2）The feature of the corresponding position of obtained image
Matrix expression connects, and obtains the expression of the jth position face block of all N width images：
Formula 3
（4）To what is obtained, based on K cluster centre of K-means clustering learnings, it is denoted as：
Formula 4
Wherein dj, t represent t-th of cluster centre therein, t=1,2 ... ... K, and each cluster centre can regard one as
Independent convolution kernel, is preferable to provide K=32, that is, corresponds to each face block and obtain K convolution kernel.Fig. 3 is the example of convolution kernel.
（5）Convolution kernel study is carried out respectively all in accordance with above-mentioned steps to the face block of all positions of all images, is obtained
D1 to D4.
Step 3：Face characteristic extracts
The feature extraction of face is divided into three steps：Convolution, Pooling and ReLU.
（1）Convolution operation
For inputting facial imageY, by step 2（1）With（2）Two-step pretreatment obtains the imageYJth position
The expression of face block：
Formula 5
By each cluster centre dj, t is considered as a convolution kernel, carries out convolution operation, which can regard as grasps as inner product
Make, obtain
Formula 6
It is two dimensional image by inner product and then by the vector stretched " reduction ", Fig. 4 is the signal of shown convolution operation
Figure.The facial image of input is divided into 2*2 face block, and all face blocks are carried out with above-mentioned convolution respectively, j=1, and 2,3,4, phase
Inner product is carried out using the convolution kernel of same position with the face block that position is trained, that is, corresponds to D1, corresponding D4, then for input
ImageY, 2*2*K=2*2*32=128 convolved image is always obtained, wherein the size of each convolved image is m*n=56*32.
（2）Pooling is operated
It operates to extract invariant features by mean value Pooling.First, the convolved image of each 56*32 is divided into 4*4
Nonoverlapping grid of pixel size calculates the mean value of each grid, obtains the response diagram after mean value Pooling, passes through
After Pooling operations, the size of each response diagram is 14*8.
（3）ReLU is operated
Negative Acknowledgment value in the response diagram exported after mean value Pooling is returned 0 by ReLU operations, is retained positive response, is passed through
ReLU is operated, and adds the non-linear of feature.
For input picture face block, the face representation that all response diagrams after ReLU is operated obtain after stretching is, special
Sign dimension is K*（m/4）*（n/4）=32*14*8 will input facial imageYWhat all positions obtained connects, and is inputted
Facial imageYExpression, characteristic dimension is A=2*2* K*（m/4）*（n/4）=14336 dimensions.
（4）All images in gathering training are by step 2（1）With（2）Two-step pretreatment obtains the of the image
The expression of j positions face block：
By each cluster centredj,tIt is considered as a convolution kernel, carries out convolution operation, which can regard as grasps as inner product
Make, obtain
Formula 7
It is two dimensional image by inner product and then by the vector stretched " reduction ", Fig. 4 is the signal of shown convolution operation
Figure.To the facial image, 2*2 face block is divided into, to all face blocksJ=1,2,3,4, above-mentioned convolution is carried out respectively,
The face block of same position carries out inner product using the convolution kernel of same position, i.e., correspond to D1, correspondence D4, then for the image, always
2*2*K=2*2*32=128 convolved image is obtained, wherein the size of each convolved image is m*n=56*32.
（5）Pooling is operated
It operates to extract invariant features by mean value Pooling.First, the convolved image of each 56*32 is divided into 4*4
Nonoverlapping grid of pixel size calculates the mean value of each grid, obtains the response diagram after mean value Pooling, passes through
After Pooling operations, the size of each response diagram is 14*8.
（6）ReLU is operated
Negative Acknowledgment value in the response diagram exported after mean value Pooling is returned 0 by ReLU operations, is retained positive response, is passed through
ReLU is operated, and adds the non-linear of feature.
For training set image face block, the face block that all response diagrams after ReLU is operated obtain after stretching represents
For characteristic dimension K*（m/4）*（n/4）=32*14*8 is connected what all positions of training set facial image obtained, is obtained
To the expression of training set facial image, characteristic dimension is A=2*2* K*（m/4）*（n/4）=14336 dimensions.
Step 4：Differentiate feature extraction
Differentiation feature extraction is carried out to the face representation f obtained in step 3, is divided into principal component analysis dimensionality reduction and linear discriminant
Analyze two steps.
Principal component analysis dimensionality reduction
Calculate the covariance matrix of training set character representation
Formula 8,
Wherein it is the mean value of the feature vector of all N width images in training set.
It enablesFormula 9
The corresponding feature vector of wherein corresponding preceding r maximum eigenvalue, it is preferred that PCA_Dim=r=2000.For defeated
Enter feature, obtained after principal component analysis dimensionality reduction
Formula 10
The feature of face images in training set is by dimensionality reduction postscript
（2）Linear discriminant analysis
Respectively calculate dimensionality reduction after feature class scatter and class in divergence.
Wherein formula 11
Formula 12
It wherein represents the mean value after the feature vector principal component analysis dimensionality reduction degree of person ownership image, is institute in training set
There is the mean value of the feature vector L of image N.
Formula 13
Wherein it is the corresponding feature vector of q characteristic value of following generalized eigenvalues problem maximum：
Formula 14
Preferably, q=800 are set, to carrying out differentiation feature extraction
Formula 15
To carrying out two norm normalizings：Formula 16
Step 5：Similarity calculation
For input pictureY, its character representation is obtained by step 3, is updated in formula 10 and obtained, will be updated to
It in formula 15, obtains, formula 16 will be substituted into and carry out two norm normalizings：Formula 17
Similarity is calculated based on cosine distances, since two norm normalizings have been carried out in feature, so only characteristics of needs
Inner product
) formula 18
Obviously, those skilled in the art can carry out the present invention essence of various changes and deformation without departing from the present invention
God and range, if in this way, these modification and variation of the present invention belong to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to comprising including these changes and deformation.
Claims (15)
1. a kind of face representation and similarity calculating method, it is characterised in that：
(1) generation network face training set is collected；
(2) it trains to obtain convolution kernel using the facial image in training set；
(3) convolution kernel is utilized, character representation is carried out to the facial image in training set；
(4) convolution kernel is utilized, character representation is carried out to input facial image；
(5) extraction differentiates feature in the facial image character representation in training set；
(6) extraction differentiates feature in facial image character representation is inputted；
(7) using feature is differentiated, image similarity is calculated;
Wherein step (1) is specially：Select C famous person, wherein c=1,2 ... ..., C, famous personIt is corresponding withImage is opened,
In 50<<100, famous personCorresponding all image collections are, training set is denoted as,, train the image in set, wherein i=1,2 ... ..., N embody posture, illumination and the variation of expression;
Wherein step (2) is specially：Facial image in training set is aligned, the facial image after alignmentSize is 128*80
Pixel, by each imageThe identical face block of 2*2 size is divided into, each face block diagram is 64*40 pixels as size, right
All face blocks of all images are respectively processed, and obtain convolution kernel;
The step of being wherein respectively processed to all face blocks of all images, obtain convolution kernel be specially：Each image
The face block of corresponding position j is denoted as, wherein i=1,2 ... ... N, j=1,2,3,4, in each face blockUpper extraction k*k is big
Small fritter, step-length 1, then the face blockIt is expressed as：
=[, ... ... ,], wherein 1<k<40, m=64-k+1, n=40-k+1 are calculated,
Mean value will be subtracted, and by the variance normalizing of every one-dimensional characteristic, the face block for obtaining each image corresponding position j represents：, the face block of all N width image corresponding position j is handled successively, will be obtained
N number of expression connect, the face block for obtaining the position j of N width image in training set is expressed as：, to the carry out K-means clustering learnings, K cluster centre is obtained,
It is denoted as：, wherein dj, t expression t-th of cluster centre therein, t=1,2 ... ...
K, each cluster centre convolution kernel independent as one, to the face block of all position j of all images all in accordance with above-mentioned step
Suddenly convolution kernel study is carried out respectively, obtain, j=1,2,3,4 }.
2. the method as described in claim 1, wherein step (3) are specially：To carrying out Pooling after facial image convolution operation
Operation and ReLU operate to carry out character representation.
3. method as claimed in claim 2, convolution operation therein are specially：By each cluster centre dj, t is as a volume
Product core, carries out inner product operation, by the convolution stretched vectorTwo dimensional image is reduced to, to training
The facial image of concentrationIn it is all, j=1,2,3,4, above-mentioned inner product operation is carried out, the face block of same position corresponds to
The convolution kernel that same position training obtains, is obtained 2*2*K convolved image, wherein the size of each convolved image is m*n.
4. method as claimed in claim 3, Pooling operations therein are specially：It operates to extract using mean value Pooling
The convolved image of each m*n is divided into nonoverlapping grid of 4*4 pixel sizes, calculates the equal of each grid by invariant features
Value, obtains the response diagram after mean value Pooling, after Pooling is operated, the size of each response diagram is（m/4）*（n/4）.
5. method as claimed in claim 4, ReLU operations therein are specially：The sound exported after mean value Pooling is operated
Negative Acknowledgment value that should be in figure returns 0, retains positive response, is operated by ReLU, the non-linear of feature is added, for input, face block that all response diagrams after ReLU is operated obtain after stretching represents
For, characteristic dimension K*（m/4）*（n/4）, the face block that 2*2 position is obtained representsIt is connected in turn, is instructed
Practice collection facial imageCharacter representation, characteristic dimension is A=2*2*K*（m/4）*（n/4）Dimension.
6. the method as described in claim 1, wherein step (4) are specially：To carrying out Pooling after facial image convolution operation
It operates to carry out character representation with ReLU.
7. method as claimed in claim 6, convolution operation therein is specially：To inputting facial image, it is aligned to 128*80
The facial image of pixel is divided into the identical face block of 2*2 size, and each face block diagram is 64*40 pixels as size,
The face block of corresponding position j is denoted as, j=1,2,3,4, in each face blockThe fritter of upper extraction k*k sizes, step-length 1,
The then face blockIt is expressed as：, wherein 1<k<40, m=64-k+1,n=
40-k+1 is calculated, willSubtract mean value, and by the variance normalizing of every one-dimensional characteristic, obtain the image's
The expression of jth position face block, by each cluster centre dj, t is as a volume
Product core, carries out inner product operation, by the convolution stretched vectorTwo dimensional image is reduced to, to input
Facial imageIn it is allCarry out above-mentioned inner product operation, j=1,2,3,4, the face block of same position corresponds to same position instruction
2*2*K convolved image is obtained in the convolution kernel got out, wherein the size of each convolved image is m*n.
8. the method for claim 7, Pooling operations therein are specially：It operates to extract using mean value Pooling
The convolved image of each m*n is divided into nonoverlapping grid of 4*4 pixel sizes, calculates the equal of each grid by invariant features
Value, obtains the response diagram after mean value Pooling, after Pooling is operated, the size of each response diagram is（m/4）*（n/4）.
9. method as claimed in claim 8, ReLU operations therein are specially：The response diagram that will be exported after mean value Pooling
In Negative Acknowledgment value return 0, retain positive response, operated by ReLU, the non-linear of feature is added, for input, face block that all response diagrams after ReLU is operated obtain after stretching represents
For, characteristic dimension K*（m/4）*（n/4）, the face block that 2*2 position is obtained representsIt is connected in turn, obtains
Input facial imageExpression, characteristic dimension is A=2*2*K*（m/4）*（n/4）Dimension.
10. method as claimed in claim 9, wherein step (5) are specially：To facial image character representation by principal component point
Analysis dimensionality reduction calculates and linear discriminant analysis obtains differentiating feature.
11. the step of method as claimed in claim 10, wherein principal component analysis dimensionality reduction calculate is specially：It is special to calculate training set
Levy the covariance matrix represented, whereinIt is the mark sheet of all N width images in training set
ShowMean value, enable, wherein
It is correspondingThe corresponding feature vector of preceding r maximum eigenvalue, for input feature vector, after principal component analysis dimensionality reduction
It arrives,, the character representation of the face images in training set is by dimensionality reduction postscript。
12. method as claimed in claim 10, wherein linear discriminant analysis are specially：It is dissipated between the class of character representation after calculating dimensionality reduction respectively
DegreeWith divergence Sw in class, wherein,
,Represent thePerson ownership imageFeature vector principal component analysis dimensionality reduction degree after mean value,It is institute in training set
There is the character representation of imageMean value,, PCA_
Dim=r, whereinIt is q characteristic value of following generalized eigenvalues problem maximum
Corresponding feature vector：,, it is rightDifferentiation feature extraction is carried out, obtains differentiating feature。
13. method as claimed in claim 12, wherein step（6）Specially：Principal component is passed through to facial image character representation
Analysis dimensionality reduction calculates and linear discriminant analysis obtains differentiating feature.
14. method as claimed in claim 13, principal component analysis dimensionality reduction therein calculate the specific steps are：To input picture
Character representation, calculate, obtainAfterwards, differentiation feature is calculated。
15. method as claimed in claim 13, wherein step（7）Specially：It is rightTwo norm normalizings are carried out to obtain：,
It is rightTwo norm normalizings are carried out to obtain：, carry out inner product calculatingIt obtains
Input pictureWith training set facial imageSimilarity.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410604043.0A CN104408405B (en) | 2014-11-03 | 2014-11-03 | Face representation and similarity calculating method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410604043.0A CN104408405B (en) | 2014-11-03 | 2014-11-03 | Face representation and similarity calculating method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104408405A CN104408405A (en) | 2015-03-11 |
| CN104408405B true CN104408405B (en) | 2018-06-15 |
Family
ID=52646036
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410604043.0A Expired - Fee Related CN104408405B (en) | 2014-11-03 | 2014-11-03 | Face representation and similarity calculating method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104408405B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104881798A (en) * | 2015-06-05 | 2015-09-02 | 北京京东尚科信息技术有限公司 | Device and method for personalized search based on commodity image features |
| CN105046272B (en) * | 2015-06-29 | 2018-06-19 | 电子科技大学 | A kind of image classification method based on succinct non-supervisory formula convolutional network |
| CN106372073A (en) * | 2015-07-21 | 2017-02-01 | 北京大学 | Mathematical formula retrieval method and apparatus |
| CN106611177A (en) * | 2015-10-27 | 2017-05-03 | 北京航天长峰科技工业集团有限公司 | Big data-based image classification method |
| CN105426914B (en) * | 2015-11-19 | 2019-03-15 | 中国人民解放军信息工程大学 | A Location Recognition-Oriented Image Similarity Detection Method |
| CN106803054B (en) * | 2015-11-26 | 2019-04-23 | 腾讯科技（深圳）有限公司 | Faceform's matrix training method and device |
| CN105590115A (en) * | 2015-12-18 | 2016-05-18 | 北京畅景立达软件技术有限公司 | Visual-angle-invariant human feature expression method |
| CN106355202A (en) * | 2016-08-31 | 2017-01-25 | 广州精点计算机科技有限公司 | Image feature extraction method based on K-means clustering |
| CN108805151B (en) * | 2017-05-05 | 2021-05-25 | 杭州微禾迅科技有限公司 | Image classification method based on depth similarity network |
| CN107609481B (en) * | 2017-08-14 | 2020-11-20 | 百度在线网络技术（北京）有限公司 | Method, apparatus and computer storage medium for generating training data for face recognition |
| CN109299746A (en) * | 2018-10-22 | 2019-02-01 | 广州星唯信息科技有限公司 | A kind of segment chord similar degree calculation method |
| CN112597862B (en) * | 2020-12-16 | 2024-07-19 | 上海芯翌智能科技有限公司 | Method and equipment for cleaning face data |
| CN116246311B (en) * | 2022-10-25 | 2025-11-21 | 南京审计大学 | Face recognition method based on multi-layer cascade boundary median discrimination projection |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008081152A3 (en) * | 2006-12-22 | 2008-09-18 | France Telecom | System and method for recognizing an object in an image |
| CN101470896A (en) * | 2007-12-24 | 2009-07-01 | 南京理工大学 | Automotive target flight mode prediction technique based on video analysis |
| CN103020261A (en) * | 2012-12-24 | 2013-04-03 | 南京邮电大学 | Image automatic marking method |
| CN103530657A (en) * | 2013-09-26 | 2014-01-22 | 华南理工大学 | Deep learning human face identification method based on weighting L2 extraction |
-
2014
- 2014-11-03 CN CN201410604043.0A patent/CN104408405B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2008081152A3 (en) * | 2006-12-22 | 2008-09-18 | France Telecom | System and method for recognizing an object in an image |
| CN101470896A (en) * | 2007-12-24 | 2009-07-01 | 南京理工大学 | Automotive target flight mode prediction technique based on video analysis |
| CN103020261A (en) * | 2012-12-24 | 2013-04-03 | 南京邮电大学 | Image automatic marking method |
| CN103530657A (en) * | 2013-09-26 | 2014-01-22 | 华南理工大学 | Deep learning human face identification method based on weighting L2 extraction |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104408405A (en) | 2015-03-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104408405B (en) | Face representation and similarity calculating method | |
| Zheng et al. | Recent advances of deep learning for sign language recognition | |
| Rekha et al. | Shape, texture and local movement hand gesture features for indian sign language recognition | |
| Sun et al. | DL-SFA: Deeply-learned slow feature analysis for action recognition | |
| Sivan et al. | Computer vision based assistive technology for blind and visually impaired people | |
| Dominio et al. | Combining multiple depth-based descriptors for hand gesture recognition | |
| Hasanuzzaman et al. | Robust and effective component-based banknote recognition for the blind | |
| Pigou et al. | Gesture and sign language recognition with temporal residual networks | |
| Ansari et al. | Nearest neighbour classification of Indian sign language gestures using kinect camera | |
| CN102938065B (en) | Face feature extraction method and face identification method based on large-scale image data | |
| CN104915673B (en) | A kind of objective classification method and system of view-based access control model bag of words | |
| CN112818850B (en) | Cross-pose face recognition method and system based on progressive neural network and attention mechanism | |
| CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
| Haque et al. | Two-handed bangla sign language recognition using principal component analysis (PCA) and KNN algorithm | |
| Zamani et al. | Saliency based alphabet and numbers of American sign language recognition using linear feature extraction | |
| Hu et al. | Surveillance video face recognition with single sample per person based on 3D modeling and blurring | |
| Mohane et al. | Object recognition for blind people using portable camera | |
| CN103336835A (en) | Image retrieval method based on weight color-sift characteristic dictionary | |
| Pradeep Kumar et al. | Design and development of asl recognition by kinect using bag of features | |
| CN108520215A (en) | Single-sample face recognition method based on multi-scale joint feature encoder | |
| Sadeghzadeh et al. | Triplet loss-based convolutional neural network for static sign language recognition | |
| Engoor et al. | Occlusion-aware dynamic human emotion recognition using landmark detection | |
| Umakanthan et al. | Spatio temporal feature evaluation for action recognition | |
| Dlshad et al. | Kurdish sign language recognition system | |
| CN108256405A (en) | A kind of face identification method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: 100080, room 106, unit 5, Yi Sheng garden, No. 331 East Zhongguancun Road, Beijing, Haidian District Applicant after: BEIJING CKLEADER SOFTWARE TECHNOLOGY Co.,Ltd. Address before: 100080, room 1509, block F, Zhongguancun mansion, No. 16 Haidian street, Haidian District, Beijing Applicant before: BEIJING CKLEADER SOFTWARE TECHNOLOGY Co.,Ltd. |
|
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180615 Termination date: 20211103 |