CN107451537B - Face recognition method based on deep learning multi-layer non-negative matrix decomposition - Google Patents

Face recognition method based on deep learning multi-layer non-negative matrix decomposition Download PDF

Info

Publication number
CN107451537B
CN107451537B CN201710568578.0A CN201710568578A CN107451537B CN 107451537 B CN107451537 B CN 107451537B CN 201710568578 A CN201710568578 A CN 201710568578A CN 107451537 B CN107451537 B CN 107451537B
Authority
CN
China
Prior art keywords
matrix
layer
projection
test sample
coefficient
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.)
Active
Application number
CN201710568578.0A
Other languages
Chinese (zh)
Other versions
CN107451537A (en
Inventor
同鸣
李明阳
陈逸然
席圣男
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN201710568578.0A priority Critical patent/CN107451537B/en
Publication of CN107451537A publication Critical patent/CN107451537A/en
Application granted granted Critical
Publication of CN107451537B publication Critical patent/CN107451537B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • G06V40/172—Classification, e.g. identification
    • 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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • Biomedical Technology (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a Face recognition method based on deep learning multilayer nonnegative matrix decomposition, which mainly solves the problem of low recognition rate of the existing Face recognition technology under complex appearance change.

Description

Face recognition method based on deep learning multi-layer non-negative matrix decomposition
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a face image identification method which can be applied to the fields of identity authentication and information security.
Background
With the continuous development of human society, the face recognition has wide application in a plurality of fields such as security, finance, electronic government affairs and the like, and the improvement of the face recognition performance is beneficial to the expansion of the application of the face recognition. The main current research on face recognition is to extract efficient, robust and more discriminative features and to design classifiers with better classification capabilities. The key to improve the robustness of face recognition is to select more robust and discriminative features and design a classifier with good classification capability.
The non-negative matrix decomposition is a characteristic extraction method for matrix decomposition under non-negative constraint, has good data representation capability, can greatly reduce the dimensionality of data characteristics, has decomposition characteristics conforming to the visual experience of human visual perception, and has interpretable and clear physical significance for decomposition results. Basic non-negative matrix factorization NMF directly decomposes the original coefficient matrices into basis and coefficient matrices and requires that both the basis and coefficient matrices are non-negative, which indicates that non-negative matrix factorization NMF only has additive combinations. Therefore, the non-negative matrix factorization NMF can be regarded as a model based on partial representation, which can provide a local structure of observed data, but in some cases, the NMF algorithm also gives a global feature, which results in limited classification performance.
Deep learning is a new research direction of feature representation in the field of machine learning, and in recent years, breakthrough progress is made in various applications such as speech recognition and computer vision, and the deep learning forms more abstract high-level representation or features by combining bottom-level features. The deep learning model has more nonlinear transformation layers and stronger generalization capability. However, in practical applications, the performance of deep learning is degraded due to changes in appearance caused by factors such as head posture, lighting, and shading, and no good solution has been found so far.
Disclosure of Invention
In view of the above-mentioned shortcomings in the prior art, an object of the present invention is to provide a face recognition method based on deep learning multi-layer non-negative matrix factorization, so as to obtain low-rank robust features with more discriminative deep layers and improve the face recognition rate under complex appearance changes.
The technical key point for realizing the method is to introduce a new multi-layer non-negative matrix decomposition on the basis of deep learning so as to improve the existing deep learning method. Specifically, the invention obtains a low-rank characteristic representation with better discrimination by performing multiple times of nonnegative matrix decomposition on sample characteristics obtained by deep learning, thereby improving the face recognition rate, and the steps comprise:
(1) inputting each channel data of a training sample into a VGG-Face deep convolution neural network to obtain characteristic data X (K) of each channel data of the training sample, wherein K is 1, 2.
(2) Respectively carrying out the characteristic extraction processes of normalization, nonlinear transformation and matrix decomposition on the characteristic data X (k) obtained in the step (1) to obtain a coefficient matrix H (k);
(3) repeating the feature extraction process in the step (2) for L times to obtain the low-rank robust feature hj(k) Wherein j is 1,2, and n is the total number of training samples;
(4) obtaining a low-rank robust feature h according to the step (3)j(k) Constructing K nearest neighbor classifiers;
(5) inputting each channel data of the test sample into a VGG-Face deep convolution neural network to obtain characteristic data Y (k) of each channel data of the test sample;
(6) performing a projection process according to the characteristic data Y (k) obtained in the step (5) to obtain a projection coefficient vector
Figure BDA0001349019230000021
(7) The projection coefficient vector obtained in the step (6) is processed
Figure BDA0001349019230000022
Inputting the test samples into K nearest neighbor classifiers to obtain a classification result of each channel of the test samples, wherein i is 1, 2.
(8) And (5) integrating the classification result of each channel of the test sample obtained in the step (7) to obtain the classification result of the test sample.
Compared with the prior art, the invention has the following advantages:
1) the method combines multi-layer non-negative matrix factorization on the basis of deep learning, and can obtain characteristic representation with higher discriminative power;
2) the invention further improves the face recognition rate under the complex appearance change by integrating the classification results of different channels.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention.
Detailed Description
Referring to fig. 1, the face recognition based on deep learning multi-layer non-negative matrix factorization of the present invention comprises the following steps:
step 1, obtaining characteristic data X (k) of each channel data of a training sample.
(1a) Obtaining a face data set VtrainThe method comprises the following steps of (1) taking the training samples as a training data set, wherein the total number of the training samples in the training data set is n, the number of the classes of the training data set is c, each training sample in the training data set is equally divided into K regions, each region is taken as 1 channel data of the training sample, and the training samples contain K channel data;
(1b) according to the training data set, under an L inux operating system, a Caffe deep learning framework is utilized to finely adjust the VGG-Face deep convolution neural network parameters;
(1c) inputting each channel data of each training sample in a training data set into a VGG-Face deep convolution neural network to obtain characteristic data X (K) of each training channel data, wherein K is 1, 2. K is the number of channels of the training sample.
And 2, acquiring a coefficient matrix H (k) according to the characteristic data X (k).
Respectively carrying out the characteristic extraction processes of normalization, nonlinear transformation and matrix decomposition on the characteristic data X (k) to obtain a coefficient matrix H (k);
(2a) normalizing the characteristic data X (k) by using L2 norm;
(2b) carrying out nonlinear transformation on the result subjected to normalization processing in the step (2a) by using a sigmoid function to obtain a transformed result B (k);
(2c) performing matrix decomposition on the result B (k) after the nonlinear transformation in the step (2b) by using soft constraint non-negative matrix decomposition to obtain B (k) approximately equal to Z (k) A (k) F (k), wherein B (k) is a matrix of m × n order, Z (k) is a base matrix of m × phi order, A (k) is an auxiliary matrix of phi × c order, F (k) is a prediction label matrix of c × n order, m is an original feature dimension, phi is a decomposition dimension, c is a category number, and n is the total number of training samples;
(2c1) random initialization base matrix Z(1)(k) Auxiliary matrix A(1)(k) And a predictive label matrix F(1)(k) As a result of iteration 1, where the basis matrix Z(1)(k) Any element in (1) satisfies
Figure BDA0001349019230000031
Figure BDA0001349019230000038
Is a basis matrix Z(1)(k) Row p and column q elements; auxiliary matrix A(1)(k) Any element in (1) satisfies
Figure BDA0001349019230000032
Figure BDA0001349019230000039
As an auxiliary matrix A(1)(k) α row β column element, predictive tag matrix F(1)(k) Any element in (1) satisfies
Figure BDA0001349019230000033
Figure BDA00013490192300000310
For predicting label matrix F(1)(k) Row y of (2)
Figure BDA0001349019230000034
A column element, p 1,2, a, m, q 1,2, a, phi, α 1,2, a, phi, β 1,2, a, c, gamma 1,2, a, c,
Figure BDA0001349019230000037
(2c2) for the element Z in the base matrix Z, the following formula is usedp,qUpdating:
Figure BDA0001349019230000035
where T is the number of iterations, T2., iter, iter being the maximum number of iterations, T being the matrix transpose,
Figure BDA0001349019230000036
for non-normalized basis matrix Z obtained after t iterations(t)′(k) Row p and column q elements;
(2c3) for the base matrix Z obtained in step (2c2)(t)' (k) carrying out normalization processing to obtain a base matrix Z iterated for t times(t)(k)；
(2c4) For the element A in the auxiliary matrix A (k), the following formula is usedα,β(k) Updating:
Figure BDA0001349019230000041
wherein the content of the first and second substances,
Figure BDA0001349019230000042
for the auxiliary matrix A obtained after t iterations(t)(k) α row β column element A(t)(k) The auxiliary matrix is obtained after t iterations;
(2c5) according to the following formula, the elements in the prediction label matrix F (k)
Figure BDA0001349019230000043
Updating:
Figure BDA0001349019230000044
wherein the content of the first and second substances,
Figure BDA0001349019230000045
predicting label matrix F after t iterations(t)(k) Row y of (2)
Figure BDA0001349019230000046
A column element; f(t)(k) Predicting a label matrix after iterating for t times; λ is a regular term coefficient;
Figure BDA0001349019230000047
is the gamma row of the predefined local label matrix C (k)
Figure BDA0001349019230000048
A column element;
(2c6) judging whether the iteration time t reaches the maximum iteration time iter: if so, stopping iteration, and performing the basic matrix Z obtained by the iter iteration(iter)(k) Auxiliary matrix A(iter)(k) And a predictive label matrix F(iter)(k) As final basis matrix z (k), auxiliary matrix a (k) and predictive label matrix f (k); otherwise, returning to the step (2c 2);
(2d) obtaining a coefficient matrix according to the auxiliary matrix A (k) and the prediction label matrix F (k) obtained after the soft constraint non-negative matrix decomposition in the step (2 c): h (k) ═ a (k) f (k).
And 3, acquiring the low-rank robust features h (k) of the training samples.
Repeating the feature extraction process in the step 2 to obtain low-rank robust features h (k) of feature data X (k) of each channel of the training sample;
(3a) processing the characteristic data X (k) of each channel of the training sample according to the step 2 to obtain a layer 1 basis matrix Z1(k) And a layer 1 coefficient matrix H1(k)；
(3b) According to the step 2, the 1 st layer coefficient matrix H obtained in the step (3a)1(k) Processing to obtain a layer 2 base matrix Z2(k) And a layer 2 coefficient matrix H2(k)；
(3c) Continuing to repeat the same steps according to steps (3a) and (3b), according to the l-1 th layer coefficient matrix Hl-1(k) To obtain the first layer basis matrix Zl(k) And the first layer coefficient matrix Hl(k) Until the repetition number l is L, a L th layer basis matrix Z is obtainedL(k) And L th layer coefficient matrix HL(k) Wherein l 2.. L is the number of layers of the multilayer non-negative matrix decomposition;
(3d) l th layer coefficient matrix H obtained according to the step (3c)L(k) Obtaining the low-rank robust features h of each channel of the training samplej(k) Wherein j is 1, 2.
Step 4, obtaining low-rank robust features h according to the step 3j(k) And constructing K nearest neighbor classifiers.
(4a) Selecting the low-rank robust feature h of the kth channel of each training sample from the result obtained in the step 3j(k) Forming a feature set;
(4b) forming a nearest neighbor classifier according to the feature set obtained in the step (4 a);
(4c) repeating the steps (4a) and (4b) for different channels to obtain K nearest neighbor classifiers.
And 5, acquiring characteristic data Y (k) of each channel data of the test sample.
(5a) Acquiring a face data set V with the same attribute as the training data settestAs a test data set, the total number of test samples in the test data set is e, the number of categories of the test data set is c, and each test sample in the test data set is divided into K channel data according to the step (1 a);
(5b) setting parameters of the VGG-Face deep convolution neural network according to the step (1 b);
(5c) and inputting each channel data of the test sample into the VGG-Face deep convolution neural network to obtain the characteristic data Y (k) of each channel data of the test sample.
Step 6, respectively projecting the characteristic data Y (k) of each channel data of the test sample obtained in the step 5, and outputting a projection coefficient vector
Figure BDA0001349019230000051
(6a) Carrying out normalization processing, nonlinear transformation and projective transformation on the characteristic data Y (k) of the test sampleTo obtain a layer 1 projection matrix
Figure BDA0001349019230000052
(6a1) Normalizing the characteristic data Y (k) of the test sample by using L2 norm;
(6a2) performing nonlinear transformation on the result obtained after the normalization processing in the step (6a1) by using a Sigmoid function to obtain a transformation result f (Y (k)) after the nonlinear transformation, wherein f (·) represents that the nonlinear transformation is performed by using the Sigmoid function;
(6a3) respectively carrying out the nonlinear transformation on the result f (Y (k)) obtained in the step (6a2) in the step (3a) to obtain the layer 1 base matrix Z1(k) Carrying out projection transformation to obtain a1 st layer projection matrix:
Figure BDA0001349019230000053
wherein the content of the first and second substances,
Figure BDA0001349019230000054
representing a generalized inverse operation;
(6b) the layer 1 projection matrix obtained according to the step (6a)
Figure BDA0001349019230000055
And a layer 2 basis matrix Z2(k) Performing the same processing procedure to obtain the 2 nd layer projection matrix
Figure BDA0001349019230000056
(6c) Continuing to repeat the same steps according to steps (6a) and (6b), projecting the matrix according to layer l-1
Figure BDA0001349019230000057
And the l-th layer basis matrix Zl(k) Obtaining the first layer projection matrix
Figure BDA0001349019230000058
Until the repetition time l is L, a L layer projection matrix is obtained
Figure BDA0001349019230000059
Wherein, l ═ 2.., L;
(6d) l th layer projection matrix obtained according to the step (6c)
Figure BDA0001349019230000061
Obtaining a projection coefficient vector of each test sample
Figure BDA0001349019230000062
Wherein, i is 1, 2.
Step 7, the projection coefficient vector obtained in the step 6 is used
Figure BDA0001349019230000063
And inputting the data into K nearest neighbor classifiers to obtain a classification result of each channel of the test sample.
(7a) Computing low-rank robust features h of training samplesj(k) Projection coefficient vector with test sample
Figure BDA0001349019230000064
Low dimension euclidean distance therebetween
Figure BDA0001349019230000065
Obtaining a set of distances
Figure BDA0001349019230000066
Where j ═ 1, 2., n, i ∈ {1, 2., e }, | · | | computationally |, y2Represents a2 norm;
(7b) the distance set obtained according to step (7a)
Figure BDA0001349019230000067
Minimum value in distance set
Figure BDA0001349019230000068
A corresponding ξ th training sample class is used as a classification result of the ith test sample on the kth nearest classifier, wherein ξ∈ {1, 2.. multidot.n };
(7c) and (5) classifying the K channels of each test sample according to the steps (7a) and (7b) respectively to obtain the classification result of each test sample on the K nearest neighbor classifiers.
And 8, integrating the classification result of each channel of the test sample obtained in the step 7 to obtain the final classification result of the test sample.
(8a) Respectively counting the number CN of the correctly classified test samples on each nearest neighbor classifier according to the classification result of each test sample on the K nearest neighbor classifiers obtained in the step 7kAnd calculating the recognition rate of each nearest neighbor classifier:
Figure BDA0001349019230000069
wherein, CNkNumber of correctly classified test samples on the k-th nearest neighbor classifier, okThe identification rate of the kth nearest neighbor classifier;
(8b) respectively calculating the linear weight coefficients α of the K nearest neighbor classifiers according to the identification rates of the K nearest neighbor classifiers obtained in the step (8a)k：
Figure BDA00013490192300000610
(8c) Linear weight coefficient α obtained from step (8b)kCalculating K channel projection coefficient vectors of the test sample
Figure BDA00013490192300000611
Training sample K channels low-rank robust feature hj(k) Weighted distance between:
Figure BDA0001349019230000071
to obtain a weighted distance set { d }1i,d2i,...,dji,...,dni}；
(8d) According to the weighted distance set { d) obtained in the step (8c)1i,d2i,...,dji,...,dniD, weighting the minimum value d in the distance setωiAnd taking the corresponding category of the omega training sample as the classification result of the test sample, wherein omega ∈ {1, 2.
The foregoing description is only an example of the present invention and should not be construed as limiting the invention, as it will be apparent to those skilled in the art that various modifications and variations in form and detail can be made without departing from the principle and structure of the invention after understanding the present disclosure and the principles, but such modifications and variations are considered to be within the scope of the appended claims.

Claims (6)

1. The face recognition method based on deep learning multilayer non-negative matrix decomposition comprises the following steps:
(1) inputting each channel data of a training sample into a VGG-Face deep convolution neural network to obtain characteristic data X (K) of each channel data of the training sample, wherein K is 1, 2.
(2) Respectively carrying out the characteristic extraction processes of normalization, nonlinear transformation and matrix decomposition on the characteristic data X (k) obtained in the step (1) to obtain a coefficient matrix H (k);
(3) repeating the feature extraction process in the step (2) for L times to obtain the low-rank robust feature hj(k) Wherein j is 1,2, and n is the total number of training samples; the method comprises the following concrete steps:
(3a) processing the characteristic data X (k) of each channel of the training sample according to the step (2) to obtain a layer 1 basis matrix Z1(k) And a layer 1 coefficient matrix H1(k) Wherein, K is 1, 2.., K;
(3b) according to the step (2), the 1 st layer coefficient matrix H obtained in the step (3a)1(k) Processing to obtain a layer 2 base matrix Z2(k) And a layer 2 coefficient matrix H2(k)；
(3c) Continuing to repeat the same steps according to steps (3a) and (3b), according to the l-1 th layer coefficient matrix Hl-1(k) To obtain the first layer basis matrix Zl(k) And the first layer coefficient matrix Hl(k) Until the repetition number l is L, a L th layer basis matrix Z is obtainedL(k) And L th layer coefficient matrix HL(k) Wherein l 2.. L is the number of layers of the multilayer non-negative matrix decomposition;
(3d) l th layer coefficient matrix H obtained according to the step (3c)L(k) Obtaining the low-rank robust features h of each channel of the training samplej(k) Wherein j is 1, 2.. times.n;
(4) obtaining a low-rank robust feature h according to the step (3)j(k) Constructing K nearest neighbor classifiers;
(5) inputting each channel data of the test sample into a VGG-Face deep convolution neural network to obtain characteristic data Y (k) of each channel data of the test sample;
(6) performing a projection process according to the characteristic data Y (k) obtained in the step (5) to obtain a projection coefficient vector
Figure FDA0002458444860000011
(7) The projection coefficient vector obtained in the step (6) is processed
Figure FDA0002458444860000012
Inputting the test samples into K nearest neighbor classifiers to obtain a classification result of each channel of the test samples, wherein i is 1, 2.
(8) And (5) integrating the classification result of each channel of the test sample obtained in the step (7) to obtain the classification result of the test sample.
2. The method of claim 1, wherein the step (2) is implemented as follows:
(2a) normalizing the characteristic data X (k) by using L2 norm;
(2b) carrying out nonlinear transformation on the result subjected to normalization processing in the step (2a) by using a sigmoid function to obtain a transformed result B (k);
(2c) performing matrix decomposition on the result B (k) after the nonlinear transformation in the step (2b) by using soft constraint non-negative matrix decomposition to obtain B (k) approximately equal to Z (k) A (k) F (k), wherein B (k) is a matrix of m × n order, Z (k) is a base matrix of m × phi order, A (k) is an auxiliary matrix of phi × c order, F (k) is a prediction label matrix of c × n order, m is an original feature dimension, phi is a decomposition dimension, c is a category number, and n is the total number of training samples;
(2d) obtaining a coefficient matrix according to the auxiliary matrix A (k) and the prediction label matrix F (k) obtained after the soft constraint non-negative matrix decomposition in the step (2 c): h (k) ═ a (k) f (k).
3. The method of claim 2, wherein the soft constrained non-negative matrix factorization in step (2c) is used to matrix-decompose the result b (k) after the non-linear transformation in step (2b) by:
(2c1) random initialization base matrix Z(1)(k) Auxiliary matrix A(1)(k) And a predictive label matrix F(1)(k) As a result of iteration 1, where the basis matrix Z(1)(k) Any element in (1) satisfies
Figure FDA0002458444860000021
Figure FDA0002458444860000022
Is a basis matrix Z(1)(k) Row p and column q elements; auxiliary matrix A(1)(k) Any element in (1) satisfies
Figure FDA0002458444860000023
Figure FDA0002458444860000024
As an auxiliary matrix A(1)(k) α row β column element, predictive tag matrix F(1)(k) Any element in (1) satisfies
Figure FDA0002458444860000025
Figure FDA0002458444860000026
For predicting label matrix F(1)(k) Row y of (2)
Figure FDA0002458444860000028
Column elements, p 1,2,., m, q 1,2,., phi, α 1,2, …, phi, β 1,2,., c, gamma 1,2, …, c,
Figure FDA0002458444860000029
(2c2) for the element Z in the base matrix Z, the following formula is usedp,qUpdating:
Figure FDA0002458444860000027
where T is the number of iterations, T2., iter, iter being the maximum number of iterations, T being the matrix transpose,
Figure FDA0002458444860000031
for non-normalized basis matrix Z obtained after t iterations(t)' (k) row p and column q elements;
(2c3) for the base matrix Z obtained in step (2c2)(t)' (k) carrying out normalization processing to obtain a base matrix Z iterated for t times(t)(k)；
(2c4) For the element A in the auxiliary matrix A (k), the following formula is usedα,β(k) Updating:
Figure FDA0002458444860000032
wherein the content of the first and second substances,
Figure FDA0002458444860000033
for the auxiliary matrix A obtained after t iterations(t)(k) α row β column element A(t)(k) The auxiliary matrix is obtained after t iterations;
(2c5) according to the following formula, the elements in the prediction label matrix F (k)
Figure FDA0002458444860000037
Updating:
Figure FDA0002458444860000034
wherein the content of the first and second substances,
Figure FDA0002458444860000035
predicting label matrix F after t iterations(t)(k) Row y of (2)
Figure FDA0002458444860000038
A column element; f(t)(k) Predicting a label matrix after iterating for t times; λ is a regular term coefficient;
Figure FDA00024584448600000310
is the gamma row of the predefined local label matrix C (k)
Figure FDA0002458444860000039
A column element;
(2c6) judging whether the iteration time t reaches the maximum iteration time iter: if so, stopping iteration, and performing the basic matrix Z obtained by the iter iteration(iter)(k) Auxiliary matrix A(iter)(k) And a predictive label matrix F(iter)(k) As final basis matrix z (k), auxiliary matrix a (k) and predictive label matrix f (k); otherwise, return to step (2c 2).
4. The method of claim 1, wherein the step (6) is implemented as follows:
(6a) carrying out projection processing procedures of normalization processing, nonlinear transformation and projection transformation on the characteristic data Y (k) of the test sample to obtain a projection matrix
Figure FDA0002458444860000036
Wherein K ∈ {1, 2.., K }, and K is the number of sample channels;
(6a1) normalizing the characteristic data Y (k) of the test sample by using L2 norm;
(6a2) performing nonlinear transformation on the result obtained after the normalization processing in the step (6a1) by using a Sigmoid function to obtain a transformation result f (Y (k)) after the nonlinear transformation, wherein f (·) represents that the nonlinear transformation is performed by using the Sigmoid function;
(6a3) respectively carrying out the nonlinear transformation on the result f (Y (k)) obtained in the step (6a2) in the step (3a) to obtain the layer 1 base matrix Z1(k) Carrying out projection transformation to obtain a1 st layer projection matrix:
Figure FDA0002458444860000041
wherein the content of the first and second substances,
Figure FDA00024584448600000414
representing a generalized inverse operation;
(6b) the layer 1 projection matrix obtained according to the step (6a)
Figure FDA0002458444860000042
And a layer 2 basis matrix Z2(k) Performing the same processing procedure to obtain the 2 nd layer projection matrix
Figure FDA0002458444860000043
(6c) Continuing to repeat the same steps according to steps (6a) and (6b), projecting the matrix according to layer l-1
Figure FDA0002458444860000044
And the l-th layer basis matrix Zl(k) Obtaining the first layer projection matrix
Figure FDA0002458444860000045
Until the repetition time l is L, a L layer projection matrix is obtained
Figure FDA0002458444860000046
Wherein l 2.. L is the number of layers of the multilayer non-negative matrix decomposition;
(6d) l th layer projection matrix obtained according to the step (6c)
Figure FDA0002458444860000047
Obtaining a projection coefficient vector of each test sample
Figure FDA0002458444860000048
Wherein, i is 1, 2.
5. The method of claim 1, wherein the step (7) is performed as follows:
(7a) computing low-rank robust features h of training samplesj(k) Projection coefficient vector with test sample
Figure FDA0002458444860000049
Low dimension euclidean distance therebetween
Figure FDA00024584448600000410
Obtaining a set of distances
Figure FDA00024584448600000411
Where j 1, 2., n, K ∈ {1, 2., K }, i ∈ {1, 2., e }, | | |, | u |, n, K ∈ {1, 2., (K) }, i ∈ {1, 2., (e) |, |2Represents a2 norm;
(7b) the distance set obtained according to step (7a)
Figure FDA00024584448600000412
Minimum value in distance set
Figure FDA00024584448600000413
A corresponding ξ th training sample class is used as a classification result of the ith test sample on the kth nearest classifier, wherein ξ∈ {1, 2.. multidot.n };
(7c) and (5) classifying the K channels of each test sample according to the steps (7a) and (7b) respectively to obtain the classification result of each test sample on the K nearest neighbor classifiers.
6. The method of claim 1, wherein said step (8) is performed by:
(8a) according to the steps(7) The obtained classification result of each test sample on K nearest neighbor classifiers is respectively counted to obtain the number CN of correctly classified test samples on each nearest neighbor classifierkAnd calculating the recognition rate of each nearest neighbor classifier:
Figure FDA0002458444860000051
wherein, CNkNumber of correctly classified test samples on the k-th nearest neighbor classifier, okThe identification rate of the kth nearest neighbor classifier;
(8b) respectively calculating the linear weight coefficients α of the K nearest neighbor classifiers according to the identification rates of the K nearest neighbor classifiers obtained in the step (8a)k：
Figure FDA0002458444860000052
(8c) Linear weight coefficient α obtained from step (8b)kCalculating K channel projection coefficient vectors of the test sample
Figure FDA0002458444860000054
Training sample K channels low-rank robust feature hj(k) Weighted distance between:
Figure FDA0002458444860000053
obtain a set of weighted distances d1i,d2i,...,dji,...,dniJ ═ 1, 2.., n, i ∈ {1, 2.., e };
(8d) according to the weighted distance set { d) obtained in the step (8c)1i,d2i,...,dji,...,dniD, weighting the minimum value d in the distance setωiAnd taking the corresponding category of the omega training sample as the classification result of the test sample, wherein omega ∈ {1, 2.
CN201710568578.0A 2017-07-13 2017-07-13 Face recognition method based on deep learning multi-layer non-negative matrix decomposition Active CN107451537B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710568578.0A CN107451537B (en) 2017-07-13 2017-07-13 Face recognition method based on deep learning multi-layer non-negative matrix decomposition

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710568578.0A CN107451537B (en) 2017-07-13 2017-07-13 Face recognition method based on deep learning multi-layer non-negative matrix decomposition

Publications (2)

Publication Number Publication Date
CN107451537A CN107451537A (en) 2017-12-08
CN107451537B true CN107451537B (en) 2020-07-10

Family

ID=60488656

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710568578.0A Active CN107451537B (en) 2017-07-13 2017-07-13 Face recognition method based on deep learning multi-layer non-negative matrix decomposition

Country Status (1)

Country Link
CN (1) CN107451537B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108256569B (en) * 2018-01-12 2022-03-18 电子科技大学 Object identification method under complex background and used computer technology

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102254328A (en) * 2011-05-17 2011-11-23 西安电子科技大学 Video motion characteristic extracting method based on local sparse constraint non-negative matrix factorization
CN103345624A (en) * 2013-07-15 2013-10-09 武汉大学 Weighing characteristic face recognition method for multichannel pulse coupling neural network
CN105469034A (en) * 2015-11-17 2016-04-06 西安电子科技大学 Face recognition method based on weighted diagnostic sparseness constraint nonnegative matrix decomposition
CN106355138A (en) * 2016-08-18 2017-01-25 电子科技大学 Face recognition method based on deep learning and key features extraction

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9721202B2 (en) * 2014-02-21 2017-08-01 Adobe Systems Incorporated Non-negative matrix factorization regularized by recurrent neural networks for audio processing

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102254328A (en) * 2011-05-17 2011-11-23 西安电子科技大学 Video motion characteristic extracting method based on local sparse constraint non-negative matrix factorization
CN103345624A (en) * 2013-07-15 2013-10-09 武汉大学 Weighing characteristic face recognition method for multichannel pulse coupling neural network
CN105469034A (en) * 2015-11-17 2016-04-06 西安电子科技大学 Face recognition method based on weighted diagnostic sparseness constraint nonnegative matrix decomposition
CN106355138A (en) * 2016-08-18 2017-01-25 电子科技大学 Face recognition method based on deep learning and key features extraction

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
"基于NMF与BP神经网络的人脸识别方法研究";熊培；;《中国硕士学位论文全文数据库 信息科技辑》;20101215(第12期);I138-281 *
"基于深度迁移学习的人脸识别方法研究";余化鹏等；;《成都大学学报( 自然科学版)》;20170630;第36卷(第2期);第151-156页 *
"深度非负矩阵分解算法研究";曲省卫；;《中国硕士学位论文全文数据库 信息科技辑》;20170315(第3期);I138-85 *
正交指数约束的平滑非负矩阵分解方法及应用;同鸣;《系统工程与电子技术》;20130704;第35卷(第10期);第2221-2228页 *

Also Published As

Publication number Publication date
CN107451537A (en) 2017-12-08

Similar Documents

Publication Publication Date Title
Yuan et al. Projective nonnegative matrix factorization for image compression and feature extraction
He et al. Data-dependent label distribution learning for age estimation
Zhang et al. Deep cascade model-based face recognition: When deep-layered learning meets small data
Patel et al. Latent space sparse and low-rank subspace clustering
He et al. Online robust subspace tracking from partial information
Yang et al. Adaptive method for nonsmooth nonnegative matrix factorization
CN108171279B (en) Multi-view video adaptive product Grassmann manifold subspace clustering method
CN108564061B (en) Image identification method and system based on two-dimensional pivot analysis
CN112115881B (en) Image feature extraction method based on robust identification feature learning
Hu et al. Single sample face recognition under varying illumination via QRCP decomposition
CN111324791B (en) Multi-view data subspace clustering method
Lu et al. Feature fusion with covariance matrix regularization in face recognition
Sharma et al. Efficient face recognition using wavelet-based generalized neural network
Li et al. Online low-rank representation learning for joint multi-subspace recovery and clustering
He et al. Blind source separation using clustering-based multivariate density estimation algorithm
Li et al. Weighted error entropy-based information theoretic learning for robust subspace representation
Chen et al. Semi-supervised dictionary learning with label propagation for image classification
Li et al. Incoherent dictionary learning with log-regularizer based on proximal operators
CN107423697B (en) Behavior identification method based on nonlinear fusion depth 3D convolution descriptor
Chen et al. Sparse general non-negative matrix factorization based on left semi-tensor product
CN111126169A (en) Face recognition method and system based on orthogonalization graph regular nonnegative matrix decomposition
CN107451537B (en) Face recognition method based on deep learning multi-layer non-negative matrix decomposition
Kobayashi Generalized mutual subspace based methods for image set classification
CN108734206B (en) Maximum correlation principal component analysis method based on deep parameter learning
Xu et al. Nonlinear component analysis based on correntropy

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