CN109409287A - A kind of transfer learning method by macro sheet feelings to micro- expression - Google Patents
A kind of transfer learning method by macro sheet feelings to micro- expression Download PDFInfo
- Publication number
- CN109409287A CN109409287A CN201811250453.4A CN201811250453A CN109409287A CN 109409287 A CN109409287 A CN 109409287A CN 201811250453 A CN201811250453 A CN 201811250453A CN 109409287 A CN109409287 A CN 109409287A
- Authority
- CN
- China
- Prior art keywords
- micro
- matrix
- expression
- macro sheet
- projection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/174—Facial expression recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Cosmetics (AREA)
Abstract
The present invention relates to a kind of transfer learning methods by macro sheet feelings to micro- expression, the model is by the way of subspace projection in the training process, micro- expressive features matrix and macro sheet feelings eigenmatrix are projected in public subspace, and micro- expressive features matrix is reconstructed with macro sheet feelings eigenmatrix, minimize macro sheet feelings and micro- expressive features and structural difference in public subspace, so that macro sheet feelings with micro- expression are associated with maximization in subspace, iteration optimization target projection matrix, finally classify by test set data using the target projection matrix projection acquired to public subspace and using KNN.It realizes and the useful information of existing macro sheet feelings domain sample is moved into micro- expression domain, be equivalent to the number of samples for expanding marked micro- expression, overcome micro- expression sample and mark difficult drawback less.This method not only reduces the manpower waste of label, and greatly improves recognition performance, provides the another kind strategy of micro- Expression Recognition.
Description
Technical field
The invention belongs to computer visions and area of pattern recognition, are related to a kind of transfer learning by macro sheet feelings to micro- expression
Method more particularly to it is a kind of using Non-negative Matrix Factorization (NMF) carry out initialization and can keep projection front and back two domains it is public potential
Characteristic and promote two domain distances subspace projection by macro sheet feelings to micro- expression transfer learning method.
Technical background
Early in 1966, Haggard et al. found that the presence of micro- expression, referring to Haggard E A, Isaacs K
S.Micromomentary facial expressions as indicators of ego mechanisms in
psychotherapy[M]//Methods of research in psychotherapy.Springer,Boston,MA,
They think that micro- expression is one kind that human body self-protective mechanism embodies mode to 1966:154-165., can show people's heart
In constrain be reluctant by people find idea or emotion.But regrettably there is no draw in society at that time for their research
Play great repercussion.Ekman and Friesen et al. captured a kind of extremely quick in 1969 later under an accidental opportunity
The expression suppressed is attempted by people, they check and analyze frame by frame by the video recording to one section of depressive patients, in the process
It was found that and the micro- expression of formal definition.Referring to Ekman P, and Erika L R.What the face reveals:basic
and applied studies of spontaneous expression using the facial action coding
system(FACS)[M].Oxford University Press,1997.
In recent years, as society and scientific and technological are constantly progressive, the research temperature of micro- expression is even more constantly to rise.It is so far
Only, there is decades history for micro- expression research in foreign countries.Ekman et al. is even more one backbone of micro- expression research field
Strength, they have carried out a large amount of research, wherein famous has: short-term Expression Recognition tests (Brief Affect
Recognition Test, BART) (referring to: Ekman P, Friesen W V.Nonverbal behavior and
Psychopathology [J] .1974.), " Japanese and the of short duration Expression Recognition of Caucasian are test " (Japanese and
Caucasian Brief Affect Recognition Test, JACBART) (referring to: Matsumoto D, Leroux J,
Wilson Cohn C,et al.A new test to measure emotionrecognition ability:
Matsumoto and Ekman's Japanese and Caucasian Brief Affect Recognition Test
(JACBERT) [J] .Journal of Nonverbal Behavior, 2000,24 (3): 179-209.) etc..Under study for action he
Find that micro- expression can greatly help to detect a lie provide in terms of criminal investigation.While Ekman et al. (referring to: Ekman P, O'
Sullivan M.Who can catch a liar? [J] .American Psychologist, 1991,46 (46): 913-
920.) it is also proposed that, the people tested micro- Expression Recognition detection in performance and they lie test in score be breath breath
It is relevant.To 21 century, and Ekman et al. (referring to: WU Q, SHENG X B, FU X L.Micro-expression and
its applications[J].Advances in Psychological Science,2010,18(09):1359-1368.)
It developed micro- Expression Recognition training tool (Micro Expression Training Tool, METT), uses first in 2002
In the micro- Expression Recognition rate of raising.The identification of micro- expression sample can be significantly improved within a certain period of time using their training tool
Rate, corresponding training performance therefore can be improved simultaneously (referring to: Liang Jing, Yan Wenjing, Wu Qi wait the micro- expression progress of research of
With prospect [J] Chinese science fund, 2013,27 (2): 75-78.).The country also has many psychologists and scientist for micro-
Expression is studied.Professor Fu little Lan sets foot in this field at first, has carried out the micro- Expression Recognition for being directed to automatic lie identification
It is the first to be known as the micro- expression research of China for research work.2010, and Wu Qi et al. (referring to: Wu Qi, Shen Xunbing, Fu little Lan
Micro- expression research and its application [J] psychic science be in progress, 2010,18 (9): 1359-1368.) micro- expression that begins one's study and psychology
Relationship, and they summarize previous micro- Expression Recognition experiment and micro- expression training tool, to micro- expression in different necks
The application in domain is analyzed.2011 Nian Wuran (referring to: Wu Ran appoints the Set-out slide effect for spreading out and having the micro- expression of to study [J] application heart
Neo-Confucianism, 2011,17 (3): 241-248.) it is also to summarize from Psychological Angle and propose the Set-out slide effect of micro- expression.
Yan in 2013 et al. (referring to: Yan W J, Wu Q, Liu Y J, et al.CASME Database:a dataset of
spontaneous micro-expressions collected from neutralized faces[C].10th IEEE
Conference on Automatic Face and Gesture Recognition.Shanghai:IEEE, 2013.) for
Micro- expression duration is studied, the same year beam wait quietly people (referring to Liang Jing, Yan Wenjing, Wu Qi, wait micro- expression research into
Exhibition and prospect [J] Chinese science fund, 2013,27 (2): 75-78.) summarize existing micro- expression library and recognition methods.So far
Until the present, many researchers carry out related with science such as psychologic research method combination computer vision, machine learning
The work in every of micro- expression automatic identification, and have certain progress.
Existing micro- expression recognition method can be roughly divided into two types: and character representation method (referring to: Wu Q, Shen X, Fu
X.The machine knows what you are hiding:an automatic micro-expression
recognition system[M]//Affective Computing and Intelligent
Interaction.Springer, Berlin, Heidelberg, 2011:152-162. and Polikovsky S, Kameda Y,
Ohta Y.Facial micro-expressions recognition using high speed camera and 3D-
Gradient descriptor [J] .2009. and Pfister T, Li X, Zhao G, et al.Recognising
spontaneous facial micro-expressions[C]//Computer Vision(ICCV),2011IEEE
International Conference on.IEEE, 2011:1449-1456.) and linear subspaces learning method (referring to:
Wang S J,Chen H L,Yan W J,et al.Face recognition and micro-expression
recognition based on discriminant tensor subspace analysis plus extreme
Learning machine [J] .Neural processing letters, 2014,39 (1): 25-43. and Wang S J, Sun
M F,Chen Y H,et al.STPCA:sparse tensor principal component analysis for
feature extraction[C]//Pattern Recognition(ICPR),201221st International
Conference on.IEEE,2012:2278-2281.).The duration of micro- expression only has 0.04s-0.2s, and intensity is extremely micro-
It is weak, it is illuminated, the influence of the factors such as noise, the method recognition effect for causing micro- expressive features to indicate is still not ideal enough.Mesh
Preceding linear subspaces learning method makes micro- Expression Recognition already have better classification capacity, but this by sparse constraint
Kind restraining force is limited, still largely by human face's feature and light, noise effect, therefore to micro- table
The character representation of feelings and extraction cause difficulty.Meanwhile there is also micro- expression data library limited sample size, sample datas at present
The limitation of acquisition and label difficulty.
Summary of the invention
In view of the deficiencies of the prior art, the present invention provides a kind of transfer learning methods by macro sheet feelings to micro- expression;
Macro sheet feelings and micro- expression are projected to public subspace by the present invention, so that macro sheet feelings and micro- expression in public subspace
Feature distribution approach is consistent, keeps the public potential structure characteristic in both projection front and backs, so that macro sheet feelings and micro- in public subspace
Expression distance minimization, the drawback that micro- expression data library number of samples can be overcome limited, to effectively improve discrimination.
Summary of the invention:
Micro- expressive features matrix and macro sheet feelings eigenmatrix are projected to first in the subspace of P, because of macro sheet feelings and
Micro- expression contains certain potential common portions, therefore micro- expressive features matrix can pass through macro expressive features square in this sub-spaces
Battle array carries out linear reconstruction.Meanwhile linear reconstruction matrix also corresponds to reprojection's matrix for macro sheet feelings eigenmatrix, in public affairs
It is further by reprojection's macro sheet feelings eigenmatrix and micro- expressive features matrix distance by once projecting in subspace altogether
It furthers.By applying constraint to linear coefficient matrix and projection matrix, selective reconstruct is carried out, and keep macro in projection process
Expression and the public potential characteristic of micro- expression and after making projection micro- expressive features matrix and macro sheet feelings eigenmatrix in feature distribution
It reaches unanimity, so that the two distance minimization in public subspace.
Term is explained:
1, LBP feature: LBP refers to local binary patterns, full name in English: Local Binary Pattern is that one kind is used to
The operator of image local feature is described, LBP feature has the remarkable advantages such as gray scale invariance and rotational invariance.It is by
T.Ojala,M.It was proposed with D.Harwood [1] [2] in 1994.
2, LBP-TOP feature is expansion of the LBP from two-dimensional space to three-dimensional space, the full name of LBP-TOP are as follows: local
Binary patterns from three orthogonal planes, three orthogonal planes here refer to
Be exactly three orthogonal planes.
3, it opens into subspace, if x1, x2 ..., xr (r > 0) they are the r vectors of V, their all possible linear combination institutes
At collection be V a sub-spaces, referred to as x1, x2 ..., xr subspaces.
4, matrix ANorm regularization constraint:N is line number, t in formula
For columns.As the sum of 2 norms of row vector.
5, matrix ANorm regularization constraint:In formula m be line number, for column and norm, i.e.,
The maximum value of the sum of all rectangular array absolute value of a vector.
6, NMF refers to Non-negative Matrix Factorization;
7, KNN refers to K- nearest neighbour classification algorithm.
The technical solution of the present invention is as follows:
A kind of transfer learning method by macro sheet feelings to micro- expression, comprising the following steps:
(1) extract macro expressive features and micro- expressive features respectively: macro expression extraction LBP feature forms macro sheet feelings character representation
Matrix, micro- expression extraction LBP-TOP feature form micro- expressive features representing matrix, and by three orthogonal planes of LBP-TOP feature
Mean value is uniformly taken, script LBP-TOP feature is the LBP feature extracted in XY, XT, tri- orthogonal planes of YT, forms 3*59 dimension
Feature, unified here for the two dimension, the LBP feature on three orthogonal planes, which add up, takes mean value, forms 59 dimensions
LBP-TOP feature, being formed indicates feature with the unified micro- expression of macro sheet feelings dimension;
(2) by way of constructing macro sheet feelings to micro- expression transfer learning method objective function, by micro- expressive features matrix
P subspaces are projected to macro sheet feelings eigenmatrix;
(3) it is extracted by way of NMF and obtains initial projection matrix P, and by solving objective function iteration optimization mesh
Projection matrix is marked, exports target projection matrix P when meeting the condition of convergence1;
(4) take steps (1) identical mode, extracts macro expressive features and micro- expressive features to test set data respectively,
It is respectively formed test set macro sheet feelings eigenmatrix Xa teWith the micro- expressive features matrix X of test setb te, thrown by the resulting target of training
Shadow matrix P1Projection Z is carried out to micro- expressive features matrixb te=P1TXb te, pass through the resulting target projection matrix P of training1To macro sheet
Feelings eigenmatrix carries out projection Za te=P1TXa te, and distance is compared, by K nearest macro sheet feelings data sample of distance after projection
Label is assigned to micro- expression data sample as label,I.e.
Classified using KNN;
(5) the micro- expression prediction label of test set and true tag are compared, the ratio that correct number accounts for total number is to know
Not rate.
It is preferred according to the present invention, in the step (2), shown in the objective function such as formula (I):
In formula (I), α, beta, gamma is coefficient of balance, α > 0, β > 0, γ > 0;
XaIt is macro sheet feelings eigenmatrix, naMacro sheet feelings number of samples is represented, d represents intrinsic dimensionality;
XbIt is micro- expressive features matrix, nbDimension table feelings number of samples is represented, d represents intrinsic dimensionality;
P∈Rd×np, P is projection matrix, and np refers to subspace dimension;
V is reconstruction coefficients matrix (being equivalent to reprojection's matrix);
E is error matrix, stores respective characteristic ingredient in macro sheet feelings and micro- expression, and characteristic ingredient includes
Respective unique feature and structure;
Ω (P) is joint regularization term, public for keeping macro sheet feelings and the public potential characteristic of micro- expression in projection process
Potential characteristic includes feature and structure characteristic, so that macro sheet feelings and the distribution of micro- expressive features after projecting is reached unanimity, and further push away
Into subspace macro sheet feelings and micro- expression distance.
It is preferred according to the present invention, the step (2), comprising the following steps:
A, by macro sheet feelings eigenmatrix XaWith micro- expressive features matrix XbIt is projected in a sub-spaces by projection matrix P,
Macro sheet feelings eigenmatrix X because macro sheet feelings and micro- expression contain potential publicly-owned factor, after projectionaPass through reconstruction coefficients matrix V
To micro- expressive features matrix XbIt is reconstructed, is expressed as PTXb≈PTXaV;V is also considered as carrying out two to the macro sheet feelings after projection
Secondary projection is pushed further into micro- expressive features matrix and macro sheet feelings eigenmatrix distance in subspace.Reconstruct coefficient matrix V is applied
AddNorm regularization constraint, keeps reconstruction coefficients matrix V sparse, the effect of this regularization term are as follows: to make the l2 norm of the every a line of V most
It is small, occur 0 most elements in row, macro sheet feelings matrix multiple removes redundancy after 0 element and projection, completes to macro after projection
The feature selecting of expression.
It is preferred according to the present invention, the acquisition methods of the initial projection matrix P in the step A are as follows: pass through the side of NMF
Formula is in macro sheet feelings eigenmatrix XaMiddle extraction obtains initial projection matrix P, i.e., to macro sheet feelings eigenmatrix XaCarry out NMF decomposition
Xa=WH, W are basic matrix, and H is coefficient matrix;Basic matrix W is extracted as initial projection matrix P.
B、PTXb≈PTXaV is an approximate relationship, and accordingly, there exist an error matrix E, is made macro in public subspace
Expression and micro- expression are essentially equal, it may be assumed that PTXb=PTXaV+E;It can be understood as being stored with micro- expression and macro sheet in error matrix E
The respective characteristic part of feelings, meanwhile, error matrix has the function of reducing noise and eliminates over-fitting herein.To error matrix E
ApplyNorm constraint, so that error matrix E is sparse.
C, projection matrix P is constrained, to carry out feature selecting, keeps the potential of macro sheet feelings and micro- expression after projection
Publicly-owned characteristic simultaneously promotes macro sheet feelings and micro- expression distance, F norm constraint in public subspace to wish that projection matrix is sparse, carries out special
Levy the effect of selection;It is indicated to constrain the joint regularization constraint formed, Ω (P)=Ω 1 (P) by manifold constraint and MMD with Ω (P)
+ Ω 2 (P), for keeping macro sheet feelings and the public potential characteristic of micro- expression to promote the two distance in projection process, manifold constrains Ω 1
(P) shown in definition such as formula (II):
In formula (II), η > 0 is manifold regularization penalty factor, PTXb,tIt is micro- expressive features matrix t after once projecting
A sample, PTXb,kIt is micro- k-th of sample of expressive features matrix after once projecting, PTXaVtIt is macro expressive features square after reprojection
T-th of sample of battle array, PTXaVkIt is macro sheet feelings k-th of sample of eigenmatrix after reprojection, WtkIt is that t row, kth arrange in matrix W
Element;W is the weight matrix of micro- expression data manifold structure figure after projection, is reflected after projecting between micro- expression data sample
Syntople;L=D-W is figure Laplacian Matrix, and D is diagonal matrix, and diagonal element isWork as PTXb,i∈knn
(PTXb,j) when,Otherwise, Wij=0, PTXb,iFor micro- expression i-th after once projecting
A data sample, PTXb,jFor j-th of data sample of micro- expression after once projecting, it is herein constant that σ, which is Gaussian kernel bandwidth, I
For unit matrix;Look back our target be using micro- expression data after the linear reconstructing projection of macro sheet feelings data after projection,
PTXaV is PTXbAnother character representation form, therefore macro sheet feelings structural information can with micro- expression replace carrying out it is further simple
Change, manifold regularization constraint is for keeping the potential public architectural characteristic of macro sheet feelings and micro- expression after projecting;
MMD is constrained shown in the definition such as formula (III) of Ω 2 (P):
In formula (III), λ > 0 is MMD regularization penalty factor,Xa,i
For i-th of data sample of macro sheet feelings, naFor macro sheet feelings number of samples, Xb,jFor j-th of data sample of micro- expression, nbFor micro- expression sample
This number, MMD constraint wish after projecting in public subspace micro- expressive features matrix and macro sheet feelings eigenmatrix characteristic away from
From minimum, i.e. feature space after the two projection is intended to unanimously, i.e. projection back edge probability distribution is apart from nearest P (PTXb)≈P
(PTXa).Therefore, the definition of joint regularization term is as shown in formula (IV):
Ω (P)=Ω 1 (P)+Ω 2 (P)
=η tr (PTXbLXb TP)+λtr(PTMP) (Ⅳ)
s.t.PTXbDXb TP=I
In formula (IV), I is unit matrix;
Therefore, complete bound term definition is as shown in formula (V):
It is preferred according to the present invention, in the step (3), the initial projection matrix P by way of NMF, and by asking
Objective function iteration optimization target projection matrix is solved, exports target projection matrix P when meeting the condition of convergence1, comprise the following steps that
E, approximate form is converted by objective function to solve:
Shown in the objective function such as formula (VI):
In view of each item constraint, objective function be it is non-convex, approximate form is converted by objective function, such as formula (VII)
It is shown:
In formula (VII), Y is the matrix being made of the feature vector of WY=Λ DY problem solving as row vector, Λ be by
Corresponding characteristic value constitutes diagonal matrix as diagonal element;
F, it is solved by Lagrangian (ALM) mode of augmentation:
Objective function is rewritten as to the form of Augmented Lagrangian Functions, as shown in formula (VIII):
In formula (VIII), U is Lagrange multiplier, and μ is penalty factor, μ > 0.
We cannot simultaneously all variables of direct solution, so we are solved by the way of ADMM.
Its dependent variable is fixed, is successively solved in the form of iteration;Include:
A, its dependent variable is fixed, to formula (VIII) derivation zero setting, solves V, as shown in formula (Ⅸ):
In formula (Ⅸ), G is diagonal matrix, GiiFor diagonal element, work as viWhen=0, Gii=0;Otherwise,
B, its dependent variable is fixed, E is solved, as shown in formula (Ⅹ):
C, its dependent variable is fixed, by solving P to formula (VII) derivation zero setting, as shown in formula (Ⅺ):
In formula (Ⅺ),
G, judge whether to meet stop condition | | PTXb-PTXaV-E | | < ε, 0 < ε < < 0.1, ε is outage threshold, yes
Words stop iteration and export target projection matrix P1, otherwise, return step a.
The invention has the benefit that
During macro sheet feelings and micro- expression are migrated, micro- expressive features matrix and macro sheet feelings eigenmatrix are created
One public projector space carries out selective reconstruct to micro- expression data with macro expression data in public projection subspace, and
Storing macro sheet feelings and micro- expression characteristic part by error term reduces noise, and addition joint regularization constraint item keeps macro sheet feelings
With the potential common features of micro- expression, make projection after feature space be all intended to geometry it is consistent so that it is public son sky
Both interior distance is minimum, so that sufficiently extracting macro sheet feelings useful information is conducive to micro- Expression Recognition.
Detailed description of the invention
Fig. 1 is a kind of transfer learning method general flow chart by macro sheet feelings to micro- expression of the present invention;
Fig. 2 is that objective function of the present invention solves schematic diagram;
Fig. 3 is that balance parameters of the present invention influence schematic diagram;
Fig. 4 is that subspace dimension of the present invention influences schematic diagram;
Specific embodiment
The present invention is further qualified with embodiment with reference to the accompanying drawings of the specification, but not limited to this.
Embodiment 1
A kind of transfer learning method by macro sheet feelings to micro- expression, as shown in Figure 1, comprising the following steps:
(1) extract macro expressive features and micro- expressive features respectively: macro expression extraction LBP feature forms macro sheet feelings character representation
Matrix Xa, micro- expression extraction LBP-TOP feature forms micro- expressive features representing matrix Xb, and it is orthogonal flat by LBP-TOP feature three
Face uniformly takes mean value, and script LBP-TOP feature is the LBP feature extracted in XY, XT, tri- orthogonal planes of YT, forms 3*59 dimension
Feature, it is unified here for the two dimension, by the LBP feature on three orthogonal planes carry out it is cumulative take mean value, form 59 dimensions
LBP-TOP feature, being formed indicates feature with the unified micro- expression of macro sheet feelings dimension;
(2) by way of constructing macro sheet feelings to micro- expression transfer learning method objective function, by micro- expressive features matrix
P subspaces are projected to macro sheet feelings eigenmatrix;
(3) it is extracted by way of NMF and obtains initial projection matrix P, and by solving objective function iteration optimization mesh
Projection matrix is marked, exports target projection matrix P when meeting the condition of convergence1;
(4) take steps (1) identical mode, extracts macro expressive features and micro- expressive features to test set data respectively,
It is respectively formed test set macro sheet feelings eigenmatrix Xa teWith the micro- expressive features matrix X of test setb te, thrown by the resulting target of training
Shadow matrix P1Projection Z is carried out to micro- expressive features matrixb te=P1TXb te, pass through the resulting target projection matrix P of training1To macro sheet
Feelings eigenmatrix carries out projection Za te=P1TXa te, and distance is compared, by K nearest macro sheet feelings data sample of distance after projection
Label is assigned to micro- expression data sample as label,I.e.
Classified using KNN;
(5) the micro- expression prediction label of test set and true tag are compared, the ratio that correct number accounts for total number is to know
Not rate.
Embodiment 2
According to a kind of transfer learning method by macro sheet feelings to micro- expression described in embodiment 1, difference is,
In step (2), shown in the objective function such as formula (I):
In formula (I), α, beta, gamma is coefficient of balance, α > 0, β > 0, γ > 0;
XaIt is macro sheet feelings eigenmatrix, naMacro sheet feelings number of samples is represented, d represents intrinsic dimensionality;
XbIt is micro- expressive features matrix, nbDimension table feelings number of samples is represented, d represents intrinsic dimensionality;
P∈Rd×np, P is projection matrix, and np refers to subspace dimension;
V is reconstruction coefficients matrix (being equivalent to reprojection's matrix);
E is error matrix, stores respective characteristic ingredient in macro sheet feelings and micro- expression, and characteristic ingredient includes
Respective unique feature and structure;
Ω (P) is joint regularization term, public for keeping macro sheet feelings and the public potential characteristic of micro- expression in projection process
Potential characteristic includes feature and structure characteristic, so that macro sheet feelings and the distribution of micro- expressive features after projecting is reached unanimity, and further push away
Into subspace macro sheet feelings and micro- expression distance.
Step (2), comprising the following steps:
A, by macro sheet feelings eigenmatrix XaWith micro- expressive features matrix XbIt is projected in a sub-spaces by projection matrix P,
Macro sheet feelings eigenmatrix X because macro sheet feelings and micro- expression contain potential publicly-owned factor, after projectionaPass through reconstruction coefficients matrix V
To micro- expressive features matrix XbIt is reconstructed, is expressed as PTXb≈PTXaV;V is also considered as carrying out two to the macro sheet feelings after projection
Secondary projection is pushed further into micro- expressive features matrix and macro sheet feelings eigenmatrix distance in subspace.By to reconstruction coefficients square
Battle array V appliesNorm regularization constraint, makes reconstruction coefficients matrix V structural sparse, the effect of this regularization term are as follows: keep V each
Capable l2 Norm minimum, row is interior to there is 0 most elements, and macro sheet feelings matrix multiple removal redundancy, complete after 0 element and projection
The feature selecting of macro sheet feelings after projecting in pairs;
The acquisition methods of projection matrix P in step A are as follows: in macro sheet feelings eigenmatrix X by way of NMFaMiddle extraction
Initial projection matrix P is obtained, i.e., to macro sheet feelings eigenmatrix XaIt carries out NMF and decomposes Xa=WH, W are basic matrix, and H is coefficient square
Battle array;Basic matrix W is extracted as initial projection matrix P.
B、PTXb≈PTXaV is an approximate relationship, and accordingly, there exist an error matrix E, is made macro in public subspace
Expression and micro- expression are essentially equal, it may be assumed that PTXb=PTXaV+E;It can be understood as being stored with micro- expression and macro sheet in error matrix E
The respective characteristic part of feelings, meanwhile, error matrix has the function of reducing noise and eliminates over-fitting herein.To error matrix E
ApplyNorm constraint, so that error matrix E is sparse.
C, projection matrix P is constrained, to carry out feature selecting, keeps the potential of macro sheet feelings and micro- expression after projection
Publicly-owned characteristic simultaneously promotes macro sheet feelings and micro- expression distance, F norm constraint in public subspace to wish that projection matrix is sparse, carries out special
Levy the effect of selection;It is indicated to constrain the joint regularization constraint formed, Ω (P)=Ω 1 (P) by manifold constraint and MMD with Ω (P)
+ Ω 2 (P), for keeping macro sheet feelings and the public potential characteristic of micro- expression to promote the two distance in projection process, for projection matrix
P applies F norm and is constrained, and passes through minimum | | P | |2 FKeep its sparse, to extract two useful letters being projected in matrix
Breath;Joint regularization term Ω (P)=Ω 1 (P)+Ω 2 (P) is applied to projection matrix P, includes manifold regularization Ω 1 (P) and MMD
Regularization Ω 2 (P) keeps macro sheet feelings and the public potential characteristic of micro- expression in projection process.
Manifold constrains shown in the definition such as formula (II) of Ω 1 (P):
In formula (II), η > 0 is manifold regularization penalty factor, PTXb,tIt is micro- expressive features matrix t after once projecting
A sample, PTXb,kIt is micro- k-th of sample of expressive features matrix after once projecting, PTXaVtIt is macro expressive features square after reprojection
T-th of sample of battle array, PTXaVkIt is macro sheet feelings k-th of sample of eigenmatrix after reprojection, WtkIt is that t row, kth arrange in matrix W
Element;W is the weight matrix of micro- expression data manifold structure figure after projection, is reflected after projecting between micro- expression data sample
Syntople;L=D-W is figure Laplacian Matrix, and D is diagonal matrix, and diagonal element isWork as PTXb,i∈knn
(PTXb,j) when,Otherwise, Wij=0, PTXb,iFor micro- expression i-th after once projecting
A data sample, PTXb,jFor j-th of data sample of micro- expression after once projecting, it is herein constant that σ, which is Gaussian kernel bandwidth, I
For unit matrix;
The target for looking back us is using micro- expression data after the linear reconstructing projection of macro sheet feelings data after projection, PTXaV
For PTXbAnother character representation form, therefore macro sheet feelings structural information can with micro- expression replacement be further simplified, flow
The potential public architectural characteristic of macro sheet feelings and micro- expression after shape regularization constraint is used to keep projecting;PTXDXTP=I is in order to anti-
Only there is the trivial solution of P.
MMD is constrained shown in the definition such as formula (III) of Ω 2 (P):
In formula (III), λ > 0 is MMD regularization penalty factor,
Xa,iFor i-th of data sample of macro sheet feelings, naFor macro sheet feelings number of samples, Xb,jFor j-th of data sample of micro- expression, nbFor micro- table
Micro- expressive features matrix and macro sheet feelings eigenmatrix characteristic in public subspace are wished after projecting in feelings number of samples, MMD constraint
Minimum according to distance, i.e., the feature space after the two projection is intended to unanimously, i.e. projection back edge probability distribution is apart from nearest P
(PTXb)≈P(PTXa).Therefore, the definition of joint regularization term is as shown in formula (IV):
In formula (IV), I is unit matrix;
Complete bound term definition is as shown in formula (V):
Embodiment 3
A kind of transfer learning method by macro sheet feelings to micro- expression according to embodiment 1 or 2 is distinguished and is,
In step (3), the initial projection matrix P by way of NMF, and by solving objective function iteration optimization target
Projection matrix exports target projection matrix P when meeting the condition of convergence1, comprise the following steps that
E, it converts approximate form for objective function to solve: as described in Figure 2;
Shown in the objective function such as formula (VI):
In view of each item constraint, objective function be it is non-convex, approximate form is converted by objective function, such as formula (VII)
It is shown:
In formula (VII), Y is the matrix being made of the feature vector of WY=Λ DY problem solving as row vector, Λ be by
Corresponding characteristic value constitutes diagonal matrix as diagonal element;
F, it is solved by Lagrangian (ALM) mode of augmentation:
Objective function is rewritten as to the form of Augmented Lagrangian Functions, as shown in formula (VIII):
In formula (VIII), U is Lagrange multiplier, and μ is penalty factor, μ > 0,
We cannot simultaneously all variables of direct solution, so we are solved by the way of ADMM.
Its dependent variable is fixed, is successively solved in the form of iteration;Include:
A, its dependent variable is fixed, to formula (VIII) derivation zero setting, solves V, as shown in formula (Ⅸ):
In formula (Ⅸ), G is diagonal matrix, GiiFor diagonal element, work as viWhen=0, Gii=0;Otherwise,
B, its dependent variable is fixed, E is solved, as shown in formula (Ⅹ):
C, its dependent variable is fixed, by solving P to formula (VII) derivation zero setting, as shown in formula (Ⅺ):
In formula (Ⅺ),
G, judge whether to meet stop condition | | PTXb-PTXaV-E | | < ε, 0 < ε < < 0.1, ε is outage threshold, yes
Words stop iteration and export target projection matrix P1, otherwise, return step a.
Technical effect in order to better illustrate the present invention, inventor have also carried out following experiment:
It in invention, applies two popular databases and is migrated, is i.e. extension Cohn-Kanade (CK+) table
Feelings database and the micro- expression data library CASME2.
Firstly, having probed into influence of the number of iterations for experiment, the results are shown in Table 1.
Table 1
The number of iterations | 10 | 20 | 30 | 40 | 50 | 60 | 70 |
Discrimination | 0.579 | 0.592 | 0.603 | 0.612 | 0.667 | 0.667 | 0.667 |
After 50 iteration, discrimination no longer changes, and therefore, subsequent experimental maximum number of iterations is set as 50.
It is 300 that subspace dimension is chosen in this experiment, has probed into influence of the parameter for discrimination, as a result as shown in Figure 3.It says
Bright each balance parameters can play a role.
Secondly, preset parameter is α=0.01, β=0.01, γ=0.1, when η=0.01, λ=0.01, has probed into selection
Influence of the Spatial Dimension (i.e. the order of NMF) for discrimination is as a result, as shown in Figure 4.
In order to assess the macro sheet feelings of proposition to the transfer learning method performance of micro- expression, the present invention and in the micro- expression of CASME2
Other state-of-the-art micro- expression recognition methods are (as differentiated detection sub-space analysis method (DTSA), facial dynamics figure on database
(FDM) and the methods of LBP-TOP) be compared.Table 3 lists four kinds of method discriminations.From table 2 it can be seen that the present invention mentions
Model performance out is better than other several state-of-the-art methods.
Table 2
The method of the present invention | FDM | LBP-top | DTSA |
0.667 | 0.426 | 0.418 | 0.361 |
As shown in table 2, the micro- Expression Recognition rate highest of method proposed by the invention, reach 0.667, DTSA then show it is minimum
0.361.The limited of micro- expression sample greatly reduces micro- expression identification and nicety of grading.However, the model proposed utilizes table
The sample in feelings database training stage, and using the thought of transfer learning, it can solve the limited problem of micro- expression sample.
Claims (5)
1. a kind of transfer learning method by macro sheet feelings to micro- expression, which comprises the following steps:
(1) extract macro expressive features and micro- expressive features respectively: macro expression extraction LBP feature forms macro sheet feelings character representation square
Battle array, micro- expression extraction LBP-TOP feature forms micro- expressive features representing matrix, and three orthogonal planes of LBP-TOP feature are united
One takes mean value, and being formed indicates feature with the unified micro- expression of macro sheet feelings dimension;
(2) by way of constructing macro sheet feelings to micro- expression transfer learning method objective function, by micro- expressive features matrix and macro
The subspace that expressive features matrix projection is opened to P;
(3) it is extracted by way of NMF and obtains initial projection matrix P, and thrown by solving objective function iteration optimization target
Shadow matrix exports target projection matrix P when meeting the condition of convergence1;
(4) take steps (1) identical mode, extracts macro expressive features and micro- expressive features to test set data respectively, respectively
Form test set macro sheet feelings eigenmatrix Xa teWith the micro- expressive features matrix X of test setb te, pass through the resulting target projection square of training
Battle array P1Projection Z is carried out to micro- expressive features matrixb te=P1TXb te, pass through the resulting target projection matrix P of training1It is special to macro sheet feelings
Sign matrix carries out projection Za te=P1TXa te, and distance is compared, by the label of K nearest macro sheet feelings data sample of distance after projection
Micro- expression data sample is assigned to as label, is classified using KNN;
(5) the micro- expression prediction label of test set and true tag are compared, the ratio that correct number accounts for total number is discrimination.
2. a kind of transfer learning method by macro sheet feelings to micro- expression according to claim 1, which is characterized in that the step
Suddenly in (2), shown in the objective function such as formula (I):
In formula (I), α, beta, gamma is coefficient of balance, α > 0, β > 0, γ > 0;
XaIt is macro sheet feelings eigenmatrix, naMacro sheet feelings number of samples is represented, d represents intrinsic dimensionality;
XbIt is micro- expressive features matrix, nbDimension table feelings number of samples is represented, d represents intrinsic dimensionality;
P∈Rd×np, P is projection matrix, and np refers to subspace dimension;
V is reconstruction coefficients matrix;
E is error matrix, stores respective characteristic ingredient in macro sheet feelings and micro- expression, and characteristic ingredient includes respectively only
Special feature and structure;
Ω (P) is joint regularization term, public potential for keeping macro sheet feelings and the public potential characteristic of micro- expression in projection process
Characteristic includes feature and structure characteristic, so that macro sheet feelings and the distribution of micro- expression structure and features after projecting is reached unanimity, and further
Promote subspace macro sheet feelings and micro- expression distance.
3. a kind of transfer learning method by macro sheet feelings to micro- expression according to claim 2, which is characterized in that the step
Suddenly (2), comprising the following steps:
A, by macro sheet feelings eigenmatrix XaWith micro- expressive features matrix XbIt is projected in a sub-spaces, is projected by projection matrix P
Macro sheet feelings eigenmatrix X afterwardsaBy reconstruction coefficients matrix V to micro- expressive features matrix XbIt is reconstructed, is expressed as PTXb≈
PTXaV applies l to reconstruct coefficient matrix V21Norm constraint makes reconstruction coefficients matrix V structural sparse, makes the l2 model of the every a line of V
Number is minimum, occurs 0 most elements in row, and macro sheet feelings matrix multiple removes redundancy after 0 element and projection, completes to projection
The feature selecting of macro sheet feelings afterwards;
B, there are an error matrix E, keep macro sheet feelings and micro- expression in public subspace essentially equal, it may be assumed that PTXb=PTXaV+E,
L is applied to error matrix E1Norm constraint, so that error matrix E is sparse;
C, projection matrix P is constrained, to carry out feature selecting, keeps the potential publicly-owned of macro sheet feelings and micro- expression after projection
Characteristic simultaneously promotes macro sheet feelings and micro- expression distance, F norm constraint in public subspace to wish that projection matrix is sparse, carries out feature choosing
The effect selected;It is indicated to constrain the joint regularization constraint formed, Ω (P)=Ω 1 (P)+Ω 2 by manifold constraint and MMD with Ω (P)
(P), for keeping macro sheet feelings and the public potential characteristic of micro- expression to promote the two distance in projection process, manifold constrains Ω 1 (P)
Definition is as shown in formula (II):
In formula (II), η > 0 is manifold regularization penalty factor, PTXb,tIt is micro- t-th of sample of expressive features matrix after once projecting
This, PTXb,kIt is micro- k-th of sample of expressive features matrix after once projecting, PTXaVtIt is macro sheet feelings eigenmatrix after reprojection
T sample, PTXaVkIt is macro sheet feelings k-th of sample of eigenmatrix after reprojection, WtkIt is the member of t row in matrix W, kth column
Element;W is the weight matrix of micro- expression data manifold structure figure after projection, reflects the neighbour after projecting between micro- expression data sample
Connect relationship;L=D-W is figure Laplacian Matrix, and D is diagonal matrix, and diagonal element isWork as PTXb,i∈knn
(PTXb,j) when,Otherwise, Wij=0, PTXb,iFor micro- expression i-th after once projecting
A data sample, PTXb,jFor j-th of data sample of micro- expression after once projecting, it is herein constant that σ, which is Gaussian kernel bandwidth, I
For unit matrix;
MMD is constrained shown in the definition such as formula (III) of Ω 2 (P):
In formula (III), λ > 0 is MMD regularization penalty factor,Xa,iFor
I-th of data sample of macro sheet feelings, naFor macro sheet feelings number of samples, Xb,jFor j-th of data sample of micro- expression, nbFor micro- expression sample
Micro- expressive features matrix and macro sheet feelings eigenmatrix characteristic distance in public subspace are wished after projecting in number, MMD constraint
Minimum, i.e. feature space after the two projection are intended to unanimously, i.e. projection back edge probability distribution is apart from nearest P (PTXb)≈P
(PTXa), joint regularization term definition is as shown in formula (IV):
In formula (IV), I is unit matrix;
Therefore, complete bound term definition is as shown in formula (V):
4. a kind of transfer learning method by macro sheet feelings to micro- expression according to claim 3, which is characterized in that the step
The acquisition methods of initial projection matrix P in rapid A are as follows: in macro sheet feelings eigenmatrix X by way of NMFaMiddle extraction obtains
Initial projection matrix P, i.e., to macro sheet feelings eigenmatrix XaIt carries out NMF and decomposes Xa=WH, W are basic matrix, and H is coefficient matrix;It mentions
Take basic matrix W as initial projection matrix P.
5. a kind of transfer learning method by macro sheet feelings to micro- expression according to claim 3 or 4, which is characterized in that institute
It states in step (3), the initial projection matrix P by way of NMF, and by solving objective function iteration optimization target projection square
Battle array exports target projection matrix P when meeting the condition of convergence1, comprise the following steps that
E, approximate form is converted by objective function to solve:
Shown in the objective function such as formula (VI):
Approximate form is converted by objective function, as shown in formula (VII):
In formula (VII), Y is the matrix being made of the feature vector of WY=Λ DY problem solving as row vector, and Λ is by corresponding to
Characteristic value as diagonal element constitute diagonal matrix;
F, it is solved by augmentation Lagrange mode:
Objective function is rewritten as to the form of Augmented Lagrangian Functions, as shown in formula (VIII):
In formula (VIII), U is Lagrange multiplier, and μ is penalty factor, μ > 0,
Its dependent variable is fixed, is successively solved in the form of iteration;Include:
A, its dependent variable is fixed, to formula (VIII) derivation zero setting, solves V, as shown in formula (Ⅸ):
In formula (Ⅸ), G is diagonal matrix, GiiFor diagonal element, work as viWhen=0, Gii=0;Otherwise,
B, its dependent variable is fixed, E is solved, as shown in formula (Ⅹ):
C, its dependent variable is fixed, by solving P to formula (VII) derivation zero setting, as shown in formula (Ⅺ):
In formula (Ⅺ),
G, judge whether to meet stop condition | | PTXb-PTXaV-E | | < ε, 0 < ε < < 0.1, ε is outage threshold, if being, is stopped
Only iteration and export target projection matrix P1, otherwise, return step a.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811250453.4A CN109409287B (en) | 2018-10-25 | 2018-10-25 | Transfer learning method from macro expression to micro expression |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811250453.4A CN109409287B (en) | 2018-10-25 | 2018-10-25 | Transfer learning method from macro expression to micro expression |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109409287A true CN109409287A (en) | 2019-03-01 |
CN109409287B CN109409287B (en) | 2021-05-14 |
Family
ID=65469892
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811250453.4A Active CN109409287B (en) | 2018-10-25 | 2018-10-25 | Transfer learning method from macro expression to micro expression |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109409287B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110097020A (en) * | 2019-05-10 | 2019-08-06 | 山东大学 | A kind of micro- expression recognition method based on joint sparse dictionary learning |
CN111048117A (en) * | 2019-12-05 | 2020-04-21 | 南京信息工程大学 | Cross-library speech emotion recognition method based on target adaptation subspace learning |
CN111460981A (en) * | 2020-03-30 | 2020-07-28 | 山东大学 | Micro-expression recognition method for generating confrontation network model based on reconstructed cross-domain video |
CN112837212A (en) * | 2021-01-28 | 2021-05-25 | 南京大学 | Image arbitrary style migration method based on manifold alignment |
CN112991765A (en) * | 2021-02-03 | 2021-06-18 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Method, terminal and storage medium for updating road high-emission source recognition model |
CN113496147A (en) * | 2020-03-19 | 2021-10-12 | 山东大学 | Micro-expression recognition method based on joint non-negative matrix secondary decomposition of dual graph regularization |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080260212A1 (en) * | 2007-01-12 | 2008-10-23 | Moskal Michael D | System for indicating deceit and verity |
CN103186774A (en) * | 2013-03-21 | 2013-07-03 | 北京工业大学 | Semi-supervised learning-based multi-gesture facial expression recognition method |
CN106971180A (en) * | 2017-05-16 | 2017-07-21 | 山东大学 | A kind of micro- expression recognition method based on the sparse transfer learning of voice dictionary |
CN107273876A (en) * | 2017-07-18 | 2017-10-20 | 山东大学 | A kind of micro- expression automatic identifying method of ' the grand micro- transformation models of to ' based on deep learning |
WO2017190645A1 (en) * | 2016-05-03 | 2017-11-09 | 中兴通讯股份有限公司 | Method, device and storage medium for micro-expression recognition |
CN108629314A (en) * | 2018-05-07 | 2018-10-09 | 山东大学 | A kind of micro- expression recognition method based on active migration study |
CN108647628A (en) * | 2018-05-07 | 2018-10-12 | 山东大学 | A kind of micro- expression recognition method based on the sparse transfer learning of multiple features multitask dictionary |
-
2018
- 2018-10-25 CN CN201811250453.4A patent/CN109409287B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080260212A1 (en) * | 2007-01-12 | 2008-10-23 | Moskal Michael D | System for indicating deceit and verity |
CN103186774A (en) * | 2013-03-21 | 2013-07-03 | 北京工业大学 | Semi-supervised learning-based multi-gesture facial expression recognition method |
WO2017190645A1 (en) * | 2016-05-03 | 2017-11-09 | 中兴通讯股份有限公司 | Method, device and storage medium for micro-expression recognition |
CN106971180A (en) * | 2017-05-16 | 2017-07-21 | 山东大学 | A kind of micro- expression recognition method based on the sparse transfer learning of voice dictionary |
CN107273876A (en) * | 2017-07-18 | 2017-10-20 | 山东大学 | A kind of micro- expression automatic identifying method of ' the grand micro- transformation models of to ' based on deep learning |
CN108629314A (en) * | 2018-05-07 | 2018-10-09 | 山东大学 | A kind of micro- expression recognition method based on active migration study |
CN108647628A (en) * | 2018-05-07 | 2018-10-12 | 山东大学 | A kind of micro- expression recognition method based on the sparse transfer learning of multiple features multitask dictionary |
Non-Patent Citations (2)
Title |
---|
SU-JING WANG ET AL.: "Micro-Expression Recognition Using Color Spaces", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
李文书 等: "微表情识别方法综述", 《人类工效学》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110097020A (en) * | 2019-05-10 | 2019-08-06 | 山东大学 | A kind of micro- expression recognition method based on joint sparse dictionary learning |
CN110097020B (en) * | 2019-05-10 | 2023-04-07 | 山东大学 | Micro-expression recognition method based on joint sparse dictionary learning |
CN111048117A (en) * | 2019-12-05 | 2020-04-21 | 南京信息工程大学 | Cross-library speech emotion recognition method based on target adaptation subspace learning |
CN111048117B (en) * | 2019-12-05 | 2022-06-17 | 南京信息工程大学 | Cross-library speech emotion recognition method based on target adaptation subspace learning |
CN113496147A (en) * | 2020-03-19 | 2021-10-12 | 山东大学 | Micro-expression recognition method based on joint non-negative matrix secondary decomposition of dual graph regularization |
CN113496147B (en) * | 2020-03-19 | 2023-08-29 | 山东大学 | Micro-expression recognition method based on dual graph regularization and combining non-negative matrix factorization |
CN111460981A (en) * | 2020-03-30 | 2020-07-28 | 山东大学 | Micro-expression recognition method for generating confrontation network model based on reconstructed cross-domain video |
CN111460981B (en) * | 2020-03-30 | 2022-04-01 | 山东大学 | Micro-expression recognition method for generating confrontation network model based on reconstructed cross-domain video |
CN112837212A (en) * | 2021-01-28 | 2021-05-25 | 南京大学 | Image arbitrary style migration method based on manifold alignment |
CN112837212B (en) * | 2021-01-28 | 2023-09-26 | 南京大学 | Image arbitrary style migration method based on manifold alignment |
CN112991765A (en) * | 2021-02-03 | 2021-06-18 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Method, terminal and storage medium for updating road high-emission source recognition model |
Also Published As
Publication number | Publication date |
---|---|
CN109409287B (en) | 2021-05-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109409287A (en) | A kind of transfer learning method by macro sheet feelings to micro- expression | |
Baydilli et al. | Classification of white blood cells using capsule networks | |
CN104657718B (en) | A kind of face identification method based on facial image feature extreme learning machine | |
CN104866810B (en) | A kind of face identification method of depth convolutional neural networks | |
CN103632138B (en) | A kind of face identification method of low-rank piecemeal rarefaction representation | |
CN107480623B (en) | Neighbor preserving face recognition method based on collaborative representation | |
Rahman et al. | Convolutional neural networks performance comparison for handwritten Bengali numerals recognition | |
CN107832786B (en) | A kind of recognition of face classification method dictionary-based learning | |
CN108960201A (en) | A kind of expression recognition method extracted based on face key point and sparse expression is classified | |
Cao et al. | Pose and illumination variable face recognition via sparse representation and illumination dictionary | |
CN110097020A (en) | A kind of micro- expression recognition method based on joint sparse dictionary learning | |
CN110956113B (en) | Robust face recognition method based on secondary cooperation representation identification projection | |
Kolouri et al. | Explaining distributed neural activations via unsupervised learning | |
Wu et al. | Outside box and contactless palm vein recognition based on a wavelet denoising ResNet | |
Ouyang et al. | Robust automatic facial expression detection method based on sparse representation plus LBP map | |
Li et al. | Cervical histopathology image classification using ensembled transfer learning | |
Yousefnezhad et al. | Local discriminant hyperalignment for multi-subject fmri data alignment | |
Mao et al. | Cell mitosis event analysis in phase contrast microscopy images using deep learning | |
Górriz et al. | Automatic selection of ROIs in functional imaging using Gaussian mixture models | |
Schiffer et al. | Contrastive representation learning for whole brain cytoarchitectonic mapping in histological human brain sections | |
Niu et al. | Discriminative video representation with temporal order for micro-expression recognition | |
Fadel et al. | Chessboard EEG images classification for BCI systems using deep neural network | |
CN114511901A (en) | Age classification assisted cross-age face recognition algorithm | |
CN111611963B (en) | Face recognition method based on neighbor preservation canonical correlation analysis | |
Castro et al. | Generation of synthetic structural magnetic resonance images for deep learning pre-training |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |