CN104933428B - A kind of face identification method and device based on tensor description - Google Patents
A kind of face identification method and device based on tensor description Download PDFInfo
- Publication number
- CN104933428B CN104933428B CN201510437811.2A CN201510437811A CN104933428B CN 104933428 B CN104933428 B CN 104933428B CN 201510437811 A CN201510437811 A CN 201510437811A CN 104933428 B CN104933428 B CN 104933428B
- Authority
- CN
- China
- Prior art keywords
- matrix
- label
- sample
- projection
- training
- 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
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
- G06V40/172—Classification, e.g. identification
Abstract
The invention discloses a kind of face identification method and device based on tensor description, first to thering is the image pattern of label and sample to be sorted without label to carry out similarity-based learning, construct similar neighborhoods figure and normalized weight, for characterizing sample similarity, a class label matrix is manually initialized again, in order to effectively realize directly reduceing for facial image outside sample, the regularization term based on tensor description that the present invention can directly reduce one image outside sample is integrated into existing label propagation model, finally utilize parameter balance similarity measurement, initial category label and the influence based on the regularization term of matrix pattern to recognition of face, complete system modelling, take the maximum of the similarity probabilities in system output, category authentication for facial image, obtain most accurate system identification result.The thought described by introducing tensor, can effectively keep the topological structure between image pixel, and system extension malleability is good outside sample in facial image generalization procedure.
Description
Technical field
The present invention relates to machine learning, pattern-recognition and data mining technology field, and tensor is based on more particularly to one kind
The face identification method and device of description.
Background technology
In the past few decades, face recognition technology becomes research topic all the fashion in computer vision.In recent years,
Face recognition technology development is very fast, and many devices, which have been put into, uses and generate huge social and economic benefits.Recognition of face
Method mainly has geometric properties identification, the identification of feature based face, neural network recognization and support vector machines identification etc., and label passes
It is still less to broadcast application of the technology in field of face identification.
It is a kind of semi-supervised learning method based on similar diagram construction that label, which is propagated, in numerous semi-supervised learning methods,
Label transmission method because its have the advantages that it is quick, effective, simple cause the extensive concern of people.In real world, greatly
Amount can be collected into easily without label data, but have that label data is but very rare, and by can manually be consumed come nominal data
Take plenty of time, manpower;Therefore semi-supervised learning has high practical value and realistic meaning.Label propagate in 2002 by
Zhu et al. is proposed, the extensive concern of domestic and foreign scholars is caused once proposition, and is gradually proved to be a kind of simple, quick, expansion
Malleability is strong, the algorithm that performance is stablized, its application spreads all over multiple fields, social relationships analysis, multimedia messages such as social platform
Searching classification etc..In recent years, it also shows outstanding performance in terms of recognition of face.However, current most of label propagation sides
The vector space model that method uses causes precision not high because of the topological structure between destroying image pixel.
It is therefore proposed that a kind of face identification method and device based on tensor description, keep facial image topological structure
The speed of induction type process is significantly improved at the same time, is strengthened classification performance, is those skilled in the art's urgent problem to be solved.
The content of the invention
In view of this, it is existing to solve the present invention provides a kind of face identification method and device based on tensor description
Vector space model can destroy the topological structure between image pixel in technology, cause accuracy of identification to reduce and model training and pre-
Survey the high technical problem of process complexity.
In order to solve the above technical problems, the present invention provides a kind of face identification method based on tensor description, tensor is utilized
The induction type that description theory carries out facial image semi-supervised label propagation is classified, and this method includes:
The pretreatment of the direct-push classification based on tensor description, including phase of the structure based on training sample are carried out to training set
Scheme like neighbour, reconstructing power computational methods using LLE- obtains weight coefficient matrix, and symmetrization is carried out to the weight coefficient matrix
With normalized, and the classification information using training sample, initialization one is used for the known prison for recording all training samples
Superintend and direct the original tag matrix Y of information;
Label propagation and class prediction model outside the sample based on tensor description are built, output one is based on the soft of training set
Class label matrix F and a linear projection embeded matrix combination;The model updates institute by iteration in the training process
Soft class label matrix F and linear projection embeded matrix combination are stated, the projection embeded matrix combination is utilized to minimize
The fitting mistake between soft label and the soft class label matrix F is predicted in the projection being calculated, and is obtained so that the projection
Predict that the optimal projection embeded matrix of soft label closest to the soft class label matrix F combines;
The embedding operation based on tensor description is carried out to face test sample using the optimal projection embeded matrix combination,
And project the test sample to Label space, the soft class label matrix of all test samples is obtained, determines each test
The corresponding position of most probable value is the prediction classification of the test sample in the soft label vector of sample, obtains recognition result;
Wherein, the projection embeded matrix combination includes the first projection embeded matrix U and the second projection embeded matrix V.
Preferably, the similar neighborhoods figure based on training sample is built by following steps, power calculating side is reconstructed using LLE-
Method obtains weight coefficient matrix, and symmetrization processing is carried out to the weight coefficient matrix:
K nearest neighbor search is carried out to training set, finds out K nearest samples of each training sample in training set;
Using the building method of LLE- reconstruct power, calculate and weigh the similitude between the K nearest samples, construct
Similarity measure matrix obtains similar neighborhoods figure;
Symmetrization processing is carried out to the similar neighborhoods figure, obtains the weight coefficient matrix of symmetrization.
Preferably, one is initialized using the classification information of training sample by following steps to be used to record all trained samples
The original tag matrix Y of this known supervision message:
One line number of initialization definitions and columns are respectively the original tag matrix Y of c+1, N, and make the original tag square
The all elements of battle array Y are 0;
To there is label training sample xjIf xjBelong to the i-th class, then make Yi,j=1;
To no label training sample xk, make Y(c+1),k=1, and ensure the sum of element of each column in the original tag matrix Y
It is 1, represents one and only one known or unknown label of each training sample;
Wherein, c is classification sum, and N is total for training sample, class label i ∈ { 1,2 ..., c }.
Preferably, the soft label of prediction is obtained closest to the soft class label matrix F by following steps
Optimal projection embeded matrix combination:
Using embeded matrix combination is projected, bilinearity image characteristics extraction is carried out to each training sample, it is pre- to obtain projection
Survey soft label;
Using the difference between the projection soft label of prediction and the soft class label matrix F as fitting mistake;
Make the fitting mistake minimum by iteration optimization, it is embedding to obtain the optimal projection based on tensor most accurate description
Enter matrix combination;
Wherein, for described F, U, V, the method that one of value is updated using fixed other values is iterated.
Preferably, further include and l2 is introduced to the first projection embeded matrix U, 1 norm, introduces the second projection embeded matrix V
L2 norms, are specially:
Some multirows or element for making described U, V are zero, so as to reduce the mixed signal and noise in original input space
Influence to exporting soft label, ensure that U, V's is openness.
Preferably, further include the method weighed by parameter and adjust influence degree of the parameters to model.
Preferably, face test sample is based on using the optimal projection embeded matrix combination by following steps
The embedding operation of tensor description, and the test sample is projected to Label space, obtain the soft classification mark of all test samples
Sign matrix:
The unfiled sample in the test sample is projected using the optimal projection embeded matrix combination, by institute
State unfiled sample and be embedded in low-dimensional Label space from dimensional images tensor space, and predict and obtain the soft classification mark of test sample
Sign matrix.
Present invention also offers a kind of face identification device based on tensor description, using tensor description theory to face figure
Induction type classification as carrying out semi-supervised label propagation, the device include:
Training pretreatment module, for carrying out the pretreatment of the direct-push classification based on tensor description to training set, including
The similar neighborhoods figure based on training sample is built, reconstructing power computational methods using LLE- obtains weight coefficient matrix, to the power
Weight coefficient matrix carries out symmetrization and normalized, and initializes one for recording institute using the classification information of training sample
There is the original tag matrix Y of the known supervision message of training sample;
Training module, is propagated and class prediction model, output one for building label outside the sample based on tensor description
Soft class label matrix F and a linear projection embeded matrix combination based on training set;The model is in the training process
The soft class label matrix F is updated by iteration and the linear projection embeded matrix combines, the throwing is utilized to minimize
The fitting mistake between soft label and the soft class label matrix F is predicted in the projection that the combination of shadow embeded matrix is calculated, and is obtained
Combined to the optimal projection embeded matrix for causing the projection to predict soft label closest to the soft class label matrix F;
Test module, for face test sample retouch based on tensor using the optimal projection embeded matrix combination
The embedding operation stated, and the test sample is projected to Label space, the soft class label matrix of all test samples is obtained,
Determine that the corresponding position of most probable value is the prediction classification of the test sample in the soft label vector of each test sample, is obtained
To recognition result;
Wherein, the projection embeded matrix combination includes the first projection embeded matrix U and the second projection embeded matrix V.
It can be seen via above technical scheme that compared with prior art, the invention discloses a kind of based on tensor description
Face identification method and device, first carry out training set the pretreatment of the direct-push classification based on tensor description, include the use of
LLE- reconstruct power builds each training sample similar neighborhoods figure, symmetrization and normalized weight matrix is obtained, for characterizing sample
This similitude, and one original tag matrix Y of initialization, in order to effectively realize directly reduceing for facial image outside sample, this hair
It is bright that a regularization term based on tensor description that can directly reduce facial image outside sample is integrated into existing label propagation
Model, finally knows face using parameter balance similarity measurement, original class label and based on the regularization term of matrix pattern
Other influence, completes system modelling, takes the similarity probabilities in system output (i.e. the soft class label matrix of test sample)
Maximum, for the category authentication of facial image, obtains most accurate system identification result.Described by introducing based on tensor
Regularization term, can effectively keep the topological structure between image pixel, system extension exhibition outside sample in facial image generalization procedure
Property is good.
Brief description of the drawings
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
There is attached drawing needed in technology description to be briefly described, it should be apparent that, drawings in the following description are only this
The embodiment of invention, for those of ordinary skill in the art, without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is a kind of flow chart of face identification method based on tensor description provided in an embodiment of the present invention;
Fig. 2 is a kind of structure diagram signal of face identification device based on tensor description provided in an embodiment of the present invention
Figure;
Fig. 3 is a kind of identification schematic diagram of face identification method based on tensor description provided in an embodiment of the present invention.
Embodiment
Below in conjunction with the attached drawing in the embodiment of the present invention, the technical solution in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is only part of the embodiment of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art are obtained every other without making creative work
Embodiment, belongs to the scope of protection of the invention.
The core of the present invention is to provide a kind of face identification method and device based on tensor description, to solve the prior art
Middle vector space model can destroy the topological structure between image pixel, cause accuracy of identification reduction and model training and predicted
The high technical problem of journey complexity.
The invention discloses a kind of face identification method and device based on tensor description, in order to effectively realize sample stranger
Face image is directly reduceed, and the present invention is by the integrated tensor regularization term that can be used for directly reduceing facial image outside sample to marking
Propagation model is signed, and using parameter balance similarity measurement, initial category label and based on the regularization term of matrix pattern to people
The influence of face identification, completes system modelling, takes the maximum of the similarity probabilities in system output, the classification for facial image
Identification, obtains most accurate system identification result.By introducing the regularization term described based on tensor, sample of the present invention is being inherited
The topological structure between image pixel is effectively maintained while label propagation model advantage, makes the generalization procedure of facial image quick
And it is accurate, system extension malleability is good.
In order to make those skilled in the art more fully understand the present invention program, with reference to the accompanying drawings and detailed description
The present invention is described in further detail.
The present invention has used 2 real human face image data sets altogether, is Caltech Faces and Yale-B respectively.It is based on
The high efficiency of calculating considers that face images data unify boil down to 32 × 32.Wherein, Caltech Faces totally 450
Front face image, totally 27 people, i.e. 27 label classifications;Yale-B totally 2414 front face images, totally 38 people, i.e., 38 marks
Sign classification.
Present invention is primarily based on the image classification method of induction type, i.e., test set is lured using existing training pattern
Conduction Tag Estimation, therefore new method proposed by the present invention is mainly illustrated with the performance of test set.Wherein, LNP is proposed
The induction type method of test set data is handled by the soft label of data outside reconstructed sample, and GFHF, LLGC, SLP do not have
Corresponding induction type method, therefore to the above method, handled in an experiment by unified using the abductive approach of LNP formulas outside sample
Data, but performance of the above method in test set, are still mainly determined by the performance of its training pattern.And ELP is proposed using throwing
The mode of shadow handles the outer data of sample, and ELP will use this method to carry out induction type classification to test set in experiment.In order to fair and
The facility of printenv, all methods construct similar neighborhoods figure using LLE- reconstruct power by unified, establish weight matrix.
In the induction type classification experiments of recognition of face, 2 face data sets are divided into training set and test by us
Collection;For every a kind of sample, wherein 30% is taken to be used as training set, remaining 70% is used as test set.The result of experiment is using classification
Accuracy is shown;1,2,3 are taken at random in training set is per a kind of data to Yale-B ..., 10 data, which are used as, mark
Signed-off sample sheet;1,2,3 are taken at random in training set is per a kind of data to Caltech Faces ..., 6 data, which are used as, label
Sample.It is each have exemplar quantity under, take the average value of 15 grab sample results as experimental result.
The Caltech Faces human face data collection concentrated using truthful data is embodied as embodiment below
Mode illustrates.
With reference to figure 1, a kind of face identification method embodiment 1 based on tensor description provided in an embodiment of the present invention is shown
Flow chart, the induction type that this method carries out facial image using tensor description theory semi-supervised label propagation classifies, specifically
It may include steps of:
Step S100, training pretreatment is completed according to above-mentioned experimental setup, i.e., training set is carried out based on tensor description
The pretreatment of direct-push classification, including similar neighborhoods figure of the structure based on training sample, weight effective range are set as k nearest neighbor,
Power computational methods are reconstructed using LLE- and obtain weight coefficient matrix, and symmetrization and normalized are carried out to weight coefficient matrix;
And the classification information using training sample, the original mark of one known supervision message for being used to record all training samples of initialization
Sign matrix Y;
The step mainly completes the pretreatment of facial image classification based training process, is substantially carried out correlated variables, parameter is set
Put and initialize, i.e., similarity-based learning is carried out to training set data, construct reconstruction coefficients matrix, and manually calibration initialization one
Original tag matrix Y.
Step S101, label is propagated outside sample of the structure based on tensor description and class prediction model, output one are based on
The soft class label matrix F (also referred to as training soft label matrix F) of training set and a linear projection embeded matrix combination;It is above-mentioned
Model updates soft class label matrix F by iteration in the training process and linear projection embeded matrix combines, to minimize profit
The fitting mistake between soft label and soft class label matrix F is predicted in the projection being calculated with projection embeded matrix combination, is obtained
Combined to the optimal projection embeded matrix for causing projection to predict soft label closest to soft class label matrix F;
Wherein, projecting embeded matrix combination includes the first projection embeded matrix U and the second projection embeded matrix V;
Introduce projection embeded matrix combination (U and V) and bilinearity image characteristics extraction carried out to original face image data,
And establish induction type label propagation model by way of the soft label of iterative learning fitting training.It is straight according to being described based on tensor
Pushing-type Tag Estimation model, label matrix F soft to training, projection embeded matrix U and V are iterated solution, and it is soft to export training
Optimal projection the embeded matrix U and V of the label matrix F and closest F of projection result to training set.
Step S102, the insertion based on tensor description is carried out to face test sample using optimal projection embeded matrix combination
Operation, and test sample is projected to Label space, the soft class label matrix of all test samples is obtained, determines each test
The corresponding position of most probable value is the prediction classification of test sample in the soft label vector of sample, obtains recognition result.
I.e. according to optimal projection embeded matrix U, the V obtained in step 101, test set data are carried out to describe based on tensor
Dimensionality reduction projection, by raw image data projection be embedded in Label space, obtain be left all unlabeled exemplars the soft mark of prediction
Label;Maximum in the corresponding soft label of test set is worth to the corresponding prediction classification of test set sample.
Based on the face identification method based on tensor description disclosed in above-described embodiment, step is realized by following steps
In S100, the similar neighborhoods figure based on training sample is built, reconstructing power computational methods using LLE- obtains weight coefficient matrix, right
Weight coefficient matrix carries out symmetrization processing:
Step S200, k nearest neighbor search is carried out to training set, finds out K arest neighbors of each training sample in training set
Sample;
Step S201, using the building method of LLE- reconstruct power, calculate and weigh similar between K nearest samples
Property, construction similarity measure matrix obtains similar neighborhoods figure;
Step S202, symmetrization processing is carried out to similar neighborhoods figure, obtains the weight coefficient matrix of symmetrization.
It is more specifically, as follows:
Randomly select a certain proportion of data from Caltech Faces human face datas concentration first (has label and fits comprising a small amount of
Measure unlabeled exemplars, training set ratio and have the number of exemplar with reference to above-mentioned experimental setup) be used as training set, other data work
For test set (being unlabeled exemplars);If the training dataset selected isWherein m × n is
The size of the original dimension, i.e. facial image of data.L+u=NTFor the number of training sample,
To have label data collection in training set,For in training sample without label data
Collection.Assuming that total class label number is c, and every kind of label has all been contained in label data collection XLIn, and have label data collection XL
In each sample one and only one belong to the label of tally set { 1,2 ..., c }.Therefore, training set isNTFor the sample size of training set.In the present embodiment Caltech Faces human face datas
Concentrate, data dimension m × n is 32 × 32, total number of samples amount=450, label classification number c=27., will according to above-mentioned experimental setup
Caltech Faces face databases random division opens facial image for training set 135, and test set 315 is opened.
Based on this, power structure similar neighborhoods figure is reconstructed using LLE-, builds symmetrization weight matrix.Concretely comprise the following steps:
In training set, to each training sample xiIts K arest neighbors in training set is found, forms x accordinglyiK it is near
Neighbour collection N (xi).Concentrated in the present embodiment Caltech Faces human face datas, K=7.Weight between neighbour's sample is reconstructed by LLE-
Power is characterized, and the weight between non-neighbors sample is arranged to zero.Any training set image pattern xiAnd xjBetween similarity use
LLE- reconstruct power is characterized, specific as follows:
Wherein xjAnd xrIt is target image sample xiNeighbour, i.e.,
xj,xr∈N(xi).Initial reconstitution coefficient matrix isIt is and rightSymmetrization processing is carried out, is specially:OrderFor a diagonal matrix, wherein
In the present invention, realized by following steps in step S100, initialize one using the classification information of training sample
Original tag matrix Y for the known supervision message for recording all training samples:
Step S300, one line number of initialization definitions and columns are respectively the original tag matrix Y of c+1, N, and make original
The all elements of label matrix Y are 0;
Step S301, to there is label training sample xjIf xjBelong to the i-th class, then make Yi,j=1;
Step S302, to no label training sample xk, make Y(c+1),k=1, and ensure the member of each column in original tag matrix Y
The sum of element is 1, represents one and only one known or unknown label of each training sample;
Wherein, c is classification sum, and N is total for training sample, class label i ∈ { 1,2 ..., c }.
It is more specifically, as follows:
OrderFor the original tag matrix of training set, it have recorded training set sample
Known label information.Wherein, yiFor column vector, corresponding i-th of sample xi.(c+1) dimension table shows increases by one on the basis of c classes
Foreign peoples's detection dimensions, if classification results show that certain sample belongs to c+1 classes, it is foreign peoples to illustrate the sample, that is, is not belonging to known
C label classification.For original tag matrix Y, there is exemplar in training set, if it is known that xjLabel belong to
I classes, then assignment yi,j=1 (1≤i≤c), same column other elements are assigned a value of zero, i.e. yi,j=0 (1≤i≤c+1);For training
Concentrate the sample without label, the assignment y at i=c+1i,j=1, same column other elements are 0, i.e. yi,j=0 (1≤i≤c).Institute
It is original tag matrix to obtain Y, and the sum of every column element of Y is 1, i.e., each having label training sample, one and only one is true
Fixed label.Concentrated in the present embodiment Caltech Faces human face datas, class label sum c=27.
Based on the face identification method based on tensor description disclosed in above-described embodiment, step is obtained by following steps
Optimal projection embeded matrix combination in S101:
Step S400, using embeded matrix combination is projected, bilinearity image characteristics extraction is carried out to each training sample, is obtained
Soft label is predicted to projection;
Step S401, it will project and predict that the difference between soft label and soft class label matrix F (is also referred to as fitting mistake
Projection error);
Step S402, make fitting mistake minimum by iteration optimization, obtain the optimal projection based on tensor most accurate description
Embeded matrix combines;
Wherein, for F, U, V in direct-push label propagation model, one of value is updated using fixed other values
Method is iterated.
Step S101 is based primarily upon carries out direct-push label iterative diffusion to training set, that is, is carried out at the same time to the soft label of training
The iteration renewal of matrix F and the iterative learning to projection matrix U, V;Wherein U, V will be used for F to the middle projection result of training set
Renewal.
Further, it is specific as follows:
Described based on tensor, introduce the direct-push label propagation model of matrix pattern regularization term, be defined as follows:
Wherein Wi,jRepresent xiNeighbour xjTo xiCollaboration reconstruction weights.fiFor a column vector of the soft label F of training, its
Maximum element position corresponds to training sample xiPrediction label classification.Define projection matrix UTxiV is to training sample x using projection embeded matrix U, ViBilinearity image characteristics extraction is carried out, it projects knot
Fruit (characteristics of image extracted) and the soft label f of the training sampleiGap be projection error, projection error is made by iteration
Minimum, so as to obtain the projection embeded matrix combination of most accurate description.
Meanwhile model introduces l to projection embeded matrix U2,1Norm, l is introduced to projection embeded matrix V2Norm, effectively really
The openness of U, V is protected;Specifically, make many rows (or element) vanishing of projection embeded matrix U, V, so as to reduce original defeated
Enter the influence of mixed signal and noise to the soft label of output in space, ensure that U, V's is openness.
In addition, the method also weighed by parameter adjusts influence degree of the parameters to direct-push label propagation model,
Disturbance degree i.e. between factor is weighed by parameter alpha, β and γ.
During actual iterative solution, usually by model conversion into following form:
Q in model is diagonal matrix, i-th of diagonal element QiiIn corresponding above-mentioned model for μi;For there is label sample
This Qii+ ∞ is arranged to, for unlabeled exemplars QiiIt is arranged to zero.Wherein
For training sample xiProjection result.In each iteration, for F, U, V in model, we are using fixed other values come more
The method of new one of value.Concentrated in the present embodiment Caltech Faces human face datas, we are provided with label in an experiment
The Q of sampleiiFor 1010Carry out approximate+∞.
Based on the face identification method based on tensor description disclosed in above-described embodiment, step is realized by following steps
The embedding operation based on tensor description is carried out to face test sample in S102, using optimal projection embeded matrix combination, and will
Test sample is projected to Label space, obtains the soft class label matrix of all test samples:
The unfiled sample in test sample is projected using optimal projection embeded matrix combination, by unfiled sample
Low-dimensional Label space is embedded in from dimensional images tensor space, and predicts and obtains the soft class label matrix of test sample.
The step mainly carries out induction type Tag Estimation to test sample, i.e., using projecting embeded matrix U, V to test set
In projected without label data, by the unfiled sample of test set from dimensional images tensor space be embedded in low-dimensional label sky
Between, and predict and obtain the soft label matrix of unfiled sample.According to the maximum respectively arranged in soft label matrix F, corresponding survey is obtained
The prediction label classification of sample sheet, completes induction type semi-supervision image classification process.Further, it is specific as follows:
Test set data are projected using embeded matrix U, V is projected, it is specific as follows:
Wherein xi∈XTest, i.e. xiFor any test sample in test set.Induce and predict for each unlabeled exemplars
GainedThe label classification that the position correspondence of wherein maximum is predicted, i.e.,Thus
The prediction label classification of test set face images sample is obtained, completes the semi-supervised face figure of induction type based on tensor description
As assorting process.
With the embodiments of the present invention disclosed in a kind of face identification method based on tensor description it is corresponding, the present invention is real
Apply example and additionally provide a kind of face identification device based on tensor description, with reference to figure 2, which can include following content:
Training pretreatment module 101, for carrying out the pretreatment of the direct-push classification based on tensor description, bag to training set
Similar neighborhoods figure of the structure based on training sample is included, reconstructing power computational methods using LLE- obtains weight coefficient matrix, to weight
Coefficient matrix carries out symmetrization and normalized;It is and all for recording using the classification information initialization one of training sample
The original tag matrix Y of the known supervision message of training sample;
Training module 102, is propagated and class prediction model, output one for building label outside the sample based on tensor description
A soft class label matrix F and a linear projection embeded matrix combination based on training set;Model leads in the training process
Cross iteration and update soft class label matrix F and the combination of linear projection embeded matrix, embeded matrix combination is projected to minimize to utilize
The fitting mistake between soft label and soft class label matrix F is predicted in the projection being calculated, and is obtained so that projecting the soft mark of prediction
The optimal projection embeded matrix of label closest to soft class label matrix F combines;
Test module 103, for face test sample retouch based on tensor using optimal projection embeded matrix combination
The embedding operation stated, and test sample is projected to Label space, the soft class label matrix of all test samples is obtained, is determined
The corresponding position of most probable value is the prediction classification of test sample in the soft label vector of each test sample, obtains identification knot
Fruit;
Wherein, projecting embeded matrix combination includes the first projection embeded matrix U and the second projection embeded matrix V.
On the selection of experiment parameter, all methods uniformly reconstruct power to build similar neighborhoods figure, wherein to K using LLE-
The parameter K of neighbour is uniformly arranged to 7.Parameter alpha in recognition of face disaggregated model based on tensor description, β and γ will pass through net
The mode of lattice search makes choice, and the value range of parameter is { 10-6,10-4,10-2,100,102,104,106}.Experiment every time
In, the process for dividing training set and test set is random, and to having exemplar in any one every class training set sample
Quantity, will be repeated 15 times random experiments to obtain average result.Table 1 is referred to, for the present invention and 4 classical label propagation algorithms
(i.e. GFHF, LLGC, SLP, LNP) and embedded label propagation algorithm ELP, the induction type point on 6 real image data collection
Class Comparative result.In experiment, data set is divided training set and test set, and training set includes tally set and non label set, surveys
It is unlabeled exemplars to try collection.Task is to classify to test set using induction type assorting process after establishing training pattern.
Table 1 gives the average result and best result of 15 experiments.The label propagation algorithm for participating in comparing is each using the ginseng of acquiescence
Number is tested.
The present invention of table 1. and the contrast of the accuracy of conventional labels propagation algorithm and embedded label propagation algorithm
Please refer to Fig.3, show for a kind of identification of the face identification method based on tensor description disclosed by the embodiments of the present invention
It is intended to.
By experimental result we can see that the image classification positive effect of the present invention is passed better than the label of traditional classical
Algorithm is broadcast, the applicability and robustness with higher.
It should be noted that each embodiment in this specification is described by the way of progressive, each embodiment weight
Point explanation is all difference with other embodiments, between each embodiment identical similar part mutually referring to.
For device class embodiment, since it is substantially similar to embodiment of the method, so describe fairly simple, related part ginseng
See the part explanation of embodiment of the method.
Detailed Jie has been carried out to a kind of face identification method and device based on tensor description provided by the present invention above
Continue.Specific case used herein is set forth the principle of the present invention and embodiment, and the explanation of above example is only
It is the method and its core concept for being used to help understand the present invention.It should be pointed out that for those skilled in the art
For, without departing from the principle of the present invention, some improvement and modification can also be carried out to the present invention, these improve and repair
Decorations are also fallen into the protection domain of the claims in the present invention.
Claims (7)
1. it is a kind of based on tensor description face identification method, it is characterised in that using tensor description theory to facial image into
The induction type classification that the semi-supervised label of row is propagated, this method include:
The pretreatment of the direct-push classification based on tensor description is carried out to training set, including is built similar near based on training sample
Neighbour's figure, reconstructs power computational methods using LLE- and obtains weight coefficient matrix, symmetrization is carried out to the weight coefficient matrix with returning
One change is handled, and the classification information using training sample, and initialization one is used for the known supervision letter for recording all training samples
The original tag matrix Y of breath;
Label outside the sample based on tensor description is built to propagate and class prediction model, one soft classification based on training set of output
Label matrix F and a linear projection embeded matrix combination, wherein, it is embedding that the projection embeded matrix combination includes the first projection
Enter matrix U and the second projection embeded matrix V;The model is in the training process using embeded matrix combination is projected, to each instruction
Practice sample and carry out bilinearity image characteristics extraction, obtain projection and predict soft label;By the projection predict soft label with it is described soft
Difference between class label matrix F is as fitting mistake;Make the fitting mistake minimum by iteration optimization, obtain being based on opening
Measure the optimal projection embeded matrix combination of most accurate description;Wherein, for described F, U, V, it is updated using fixed other values
In the method for a value be iterated;
The embedding operation based on tensor description is carried out to face test sample using the optimal projection embeded matrix combination, and will
The test sample is projected to Label space, is obtained the soft class label matrix of all test samples, is determined each test sample
Soft label vector in the corresponding position of most probable value be the test sample prediction classification, obtain recognition result.
2. the method as described in claim 1, it is characterised in that build the similar neighborhoods based on training sample by following steps
Figure, reconstructs power computational methods using LLE- and obtains weight coefficient matrix, symmetrization processing is carried out to the weight coefficient matrix:
K nearest neighbor search is carried out to training set, finds out K nearest samples of each training sample in training set;
Using the building method of LLE- reconstruct power, calculate and weigh the similitude between the K nearest samples, construction is similar
Metric matrix obtains similar neighborhoods figure;
Symmetrization processing is carried out to the similar neighborhoods figure, obtains the weight coefficient matrix of symmetrization.
3. method as claimed in claim 2, it is characterised in that initial using the classification information of training sample by following steps
Change the original tag matrix Y of a known supervision message for being used to record all training samples:
One line number of initialization definitions and columns are respectively the original tag matrix Y of c+1, N, and make the original tag matrix Y
All elements be 0;
To there is label training sample xjIf xjBelong to the i-th class, then make Yi,j=1;
To no label training sample xk, make Y(c+1),k=1, and ensure that the sum of element of each column is in the original tag matrix Y
1, represent one and only one known or unknown label of each training sample;
Wherein, c is classification sum, and N is total for training sample, class label i ∈ { 1,2 ..., c }.
4. the method as described in claim 1, it is characterised in that further include and l2 is introduced to the first projection embeded matrix U, 1 norm,
L2 norms are introduced to the second projection embeded matrix V, are specially:
Some multirows or element for making described U, V are zero, so as to reduce mixed signal in original input space and noise to defeated
Go out the influence of soft label, ensure that U, V's is openness.
5. the method as described in Claims 1-4 any one, it is characterised in that further include the method tune weighed by parameter
Save influence degree of the parameters to model.
6. the method as described in claim 1, it is characterised in that utilize the optimal projection embeded matrix group by following steps
Close and the embedding operation based on tensor description is carried out to face test sample, and the test sample is projected to Label space, obtain
To the soft class label matrix of all test samples:
The unfiled sample in the test sample is projected using the optimal projection embeded matrix combination, by described in not
Classification samples are embedded in low-dimensional Label space from dimensional images tensor space, and predict and obtain the soft class label square of test sample
Battle array.
7. it is a kind of based on tensor description face identification device, it is characterised in that using tensor description theory to facial image into
The induction type classification that the semi-supervised label of row is propagated, the device include:
Training pretreatment module, for carrying out the pretreatment of the direct-push classification based on tensor description, including structure to training set
Similar neighborhoods figure based on training sample, reconstructs power computational methods using LLE- and obtains weight coefficient matrix, to the weight system
Matrix number carries out symmetrization and normalized, and initializes one for recording all instructions using the classification information of training sample
Practice the original tag matrix Y of the known supervision message of sample;
Training module, for building, label outside the sample based on tensor description is propagated and class prediction model, output one are based on
The soft class label matrix F of training set and a linear projection embeded matrix combination, wherein, the projection embeded matrix combination
Including the first projection embeded matrix U and the second projection embeded matrix V;The model is in the training process using projecting embeded matrix
Combination, bilinearity image characteristics extraction is carried out to each training sample, is obtained projection and is predicted soft label;The projection is predicted soft
Difference between label and the soft class label matrix F is as fitting mistake;The fitting mistake is made most by iteration optimization
It is small, obtain the optimal projection embeded matrix combination based on tensor most accurate description;Wherein, for described F, U, V, using fixing it
He is iterated value to update the method for one of value;
Test module, for being carried out using the optimal projection embeded matrix combination to face test sample based on tensor description
Embedding operation, and the test sample is projected to Label space, the soft class label matrix of all test samples is obtained, is determined
The corresponding position of most probable value is the prediction classification of the test sample in the soft label vector of each test sample, is known
Other result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510437811.2A CN104933428B (en) | 2015-07-23 | 2015-07-23 | A kind of face identification method and device based on tensor description |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510437811.2A CN104933428B (en) | 2015-07-23 | 2015-07-23 | A kind of face identification method and device based on tensor description |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104933428A CN104933428A (en) | 2015-09-23 |
CN104933428B true CN104933428B (en) | 2018-05-01 |
Family
ID=54120588
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510437811.2A Active CN104933428B (en) | 2015-07-23 | 2015-07-23 | A kind of face identification method and device based on tensor description |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104933428B (en) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105335756B (en) * | 2015-10-30 | 2019-06-11 | 苏州大学 | A kind of image classification method and image classification system based on Robust Learning model |
CN105354595B (en) * | 2015-10-30 | 2019-09-27 | 苏州大学 | A kind of robust visual pattern classification method and system |
CN105469117B (en) * | 2015-12-03 | 2019-01-01 | 苏州大学 | A kind of image-recognizing method and device extracted based on robust features |
CN105608471B (en) * | 2015-12-28 | 2020-01-14 | 苏州大学 | Robust direct-push label estimation and data classification method and system |
AU2017230184B2 (en) * | 2016-03-11 | 2021-10-07 | Magic Leap, Inc. | Structure learning in convolutional neural networks |
CN107122745B (en) * | 2017-04-28 | 2020-10-20 | 深圳市茁壮网络股份有限公司 | Method and device for identifying person track in video |
CN109740018B (en) * | 2019-01-29 | 2021-03-02 | 北京字节跳动网络技术有限公司 | Method and device for generating video label model |
CN110458078B (en) * | 2019-08-05 | 2022-05-06 | 高新兴科技集团股份有限公司 | Face image data clustering method, system and equipment |
CN111091064A (en) * | 2019-11-22 | 2020-05-01 | 天津理工大学 | Face recognition identity authentication method based on affinity nearest neighbor classification algorithm |
CN111723759B (en) * | 2020-06-28 | 2023-05-02 | 南京工程学院 | Unconstrained face recognition method based on weighted tensor sparse graph mapping |
CN112685590B (en) * | 2020-12-29 | 2022-10-14 | 电子科技大学 | Image retrieval method based on convolutional neural network regularization processing |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6990217B1 (en) * | 1999-11-22 | 2006-01-24 | Mitsubishi Electric Research Labs. Inc. | Gender classification with support vector machines |
CN101436250B (en) * | 2008-11-19 | 2010-12-01 | 西安电子科技大学 | Multi-view angle human face recognizing method based on non-linear tensor resolution and view angle manifold |
CN103886329B (en) * | 2014-03-21 | 2017-03-08 | 西安电子科技大学 | Polarization image classification method based on resolution of tensor dimensionality reduction |
CN104463202B (en) * | 2014-11-28 | 2017-09-19 | 苏州大学 | A kind of multiclass image semisupervised classification method and system |
CN104794489B (en) * | 2015-04-23 | 2019-03-08 | 苏州大学 | A kind of induction type image classification method and system based on deep tag prediction |
-
2015
- 2015-07-23 CN CN201510437811.2A patent/CN104933428B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN104933428A (en) | 2015-09-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104933428B (en) | A kind of face identification method and device based on tensor description | |
CN109961089A (en) | Small sample and zero sample image classification method based on metric learning and meta learning | |
Metcalf et al. | The strong gravitational lens finding challenge | |
CN104794489B (en) | A kind of induction type image classification method and system based on deep tag prediction | |
CN104463202B (en) | A kind of multiclass image semisupervised classification method and system | |
CN106874688B (en) | Intelligent lead compound based on convolutional neural networks finds method | |
CN109934293A (en) | Image-recognizing method, device, medium and obscure perception convolutional neural networks | |
CN105354595B (en) | A kind of robust visual pattern classification method and system | |
CN110414432A (en) | Training method, object identifying method and the corresponding device of Object identifying model | |
CN109299716A (en) | Training method, image partition method, device, equipment and the medium of neural network | |
CN109800811A (en) | A kind of small sample image-recognizing method based on deep learning | |
CN105184260B (en) | A kind of image characteristic extracting method and pedestrian detection method and device | |
Zhang et al. | A GANs-based deep learning framework for automatic subsurface object recognition from ground penetrating radar data | |
CN106971200A (en) | A kind of iconic memory degree Forecasting Methodology learnt based on adaptive-migration | |
CN108846413A (en) | A kind of zero sample learning method based on global semantic congruence network | |
CN105930834A (en) | Face identification method and apparatus based on spherical hashing binary coding | |
CN107392254A (en) | A kind of semantic segmentation method by combining the embedded structural map picture from pixel | |
CN112927266B (en) | Weak supervision time domain action positioning method and system based on uncertainty guide training | |
CN113128369A (en) | Lightweight network facial expression recognition method fusing balance loss | |
CN110390347A (en) | Conditions leading formula confrontation for deep neural network generates test method and system | |
CN114332545B (en) | Image data classification method and device based on low-bit pulse neural network | |
CN110135505A (en) | Image classification method, device, computer equipment and computer readable storage medium | |
CN110110663A (en) | A kind of age recognition methods and system based on face character | |
CN110502694A (en) | Lawyer's recommended method and relevant device based on big data analysis | |
CN109816030A (en) | A kind of image classification method and device based on limited Boltzmann machine |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |