CN105678265A - Manifold learning-based data dimensionality-reduction method and device - Google Patents

Abstract

The invention discloses a manifold learning-based data dimensionality-reduction method and a device. The method comprises the steps of firstly, dividing a to-be-detected human face image into sub-images according to an equal-strip rule and converting the image in a corresponding sub-mode; secondly, subjecting the divided to-be-detected human face image to the data dimensionality-reduction treatment; thirdly, classifying the low-dimensional vectors of the dimensionality-reduced image according to K sub-modes in a training set to obtain K recognition results; fourthly, calculating the K recognition results according to the weighting method to obtain a final recognition result of the to-be-detected human face image. Namely, the to-be-detected human face image can be recognized as one human face image in the training set.

Description

Method of Data with Adding Windows and device based on manifold learning

Technical field

The present invention relates to a kind of Method of Data with Adding Windows and device, refer more particularly to a kind of Method of Data with Adding Windows based on manifold learning and device.

Background technology

In recent years, along with developing rapidly of science and technology, the data that people obtain through various channels are greatly increased as compared with the past, therefore, utilize Data Dimensionality Reduction technology that these high dimensional datas are processed, and have just become the requisite important component part of Data processing. Traditional dimension reduction method (such as principal component analysis, independent component analysis, linear discriminant analysis etc.) can process the data set with linear structure effectively. But when data set has nonlinear organization, these methods are just difficult to the inherent low-dimensional information finding to be hidden in high dimensional data. Method of Data with Adding Windows based on manifold learning assumes that higher-dimension observation data are positioned on the low dimensional manifold being embedded into higher-dimension theorem in Euclid space, therefore can effectively find and be maintained in higher dimensional space the inherent geometry presenting distortion collection. Linearisation version as laplacian eigenmaps, locality preserving projections (LPP) algorithm achieves certain success in recognition of face, and this is main precisely due to it can keep the manifold structure at face place effectively in the face of higher-dimension distortion volume human face data collection.

But LPP algorithm is in actual face identification system, particularly in the face of complex environment and magnanimity artificial abortion have the disadvantage that when applying

First, in conventional LPP algorithm, it is that view picture facial image is done as a whole consideration, and it has recently been demonstrated that the change that causes due to the factor such as illumination condition, facial expression of face, often only it is embodied in the subregion of image, namely the scattered situation of local data occurs, and the change of other parts is seldom even unchanged, therefore, if making as a whole by view picture facial image in LPP algorithm, recognition result just will necessarily be produced a very large impact by this localized variation.

Secondly, time LPP algorithm high dimension vector represents view data, when running into singular matrix, computation complexity can be exponentially increased along with the increase of image dimension, and this inevitable consuming in a large number calculates resource and reduces the speed of service of algorithm, causes that whole system performance reduces.

In " based on the Method of Data with Adding Windows of manifold learning and the application in recognition of face thereof; Wang Jianzhong; in June, 2010 " this section of document, although disadvantages mentioned above has been realized, such as by the division of the sizes such as face carries out, then the dimension-reduction treatment to face image data of the LPP algorithm is used, to facial image classification and use the method for weighting that facial image is identified using arest neighbors sorting algorithm, but the some shortcomings of yet suffering from, the application is exactly on this basis, and it is done further improvement.

Summary of the invention

In order to overcome the deficiencies in the prior art, it is an object of the invention to provide a kind of Method of Data with Adding Windows based on manifold learning and device, make face picture retain more key feature information, and be applied in face identification system to improve the degree of accuracy of identification, reduce the consumption of whole system.

For solving the problems referred to above, the technical solution adopted in the present invention is as follows:

The invention discloses a kind of Method of Data with Adding Windows based on manifold learning, comprise the following steps:

S101: be K subimage according to certain regular partition by facial image X to be detected, is then converted into corresponding subpattern by K subimage, the vector of described subpattern is designated as X_i(i=1,2 ..., K);

S102: according to formula Y_i=W_i ^TX_iObtain Y_i; Described Y_iIt is X_iLow-dimensional vector representation, described W_i ^TIt is keep projection algorithm to be derived from by maximum margin criterion and local;

S103: be the facial image X to be detected low dimensional vector Y to its subpattern according to arest neighbors sorting algorithm_iClassify, according to the K in training set sub-set of modes, there are K recognition result; Described training set is based on the set of the facial image preset that described regular partition is K sub-set of modes;

S104: according to the method for weighting, obtaining the described facial image X to be detected probability belonging to c people after being calculated by K recognition result is:

$P_c=\frac{1}{K}\underset{i=1}{\overset{K}{\Σ}}{\mathrm{Wg}}_i*q_i^c,$ Wherein

Then the recognition result drawing facial image X to be detected is: Identity (X)=argmax (p_c);

Wherein said Wg_iBeing the identification weights of i-th subpattern set, its computing formula is:Wherein K_ijIt is for each sample X of i-th subpattern set in described training set_ijNeighbour neutralize its number being in similar sample, described X_ijBeing a sample of i-th subpattern set in described training set, described sample refers to the coordinate representation of the vector of each subpattern set; The value of j is between 1 to N, and N refers to the total number presetting facial image in described training set.

Further, described rule is divided by the stripeds such as facial image.

Further, described W_i ^TIt is by comprising the concrete steps that maximum margin criterion and local maintenance projection algorithm are derived from:

The projection algorithm target formula obtaining new locality preserving projections algorithm that combines is kept to be according to maximum margin criterion and local:

$\left\{\begin{array}{c}\min tr\left(W_i^T\right(X_i{\mathrm{LX}}_i^T-\left(S_b-S_w\right)\left)W_i\right)\\ s.t.W_i^T{X}_i{\mathrm{DX}}_i^T{W}_i=1\end{array}\right.;$ Wherein, $s_b=\underset{i=1}{\overset{l}{\Σ}}{n}_i(m_i-m){(m_i-m)}^t,$ Wherein D is diagonal matrix, and L, D are known; M is the average vector of all samples in all training sets, m_iIt is the average vector of all samples of the i-th class subpattern, S in described training set_bIt is stroll matrix between class, S_wIt is stroll matrix in class, n_iBelong to the number of samples of the i-th class subpattern in described training set; The value of i is between 1 to K; Thus formula can obtain W_i ^TUniquely determine value.

The invention also discloses a kind of Data Dimensionality Reduction device based on manifold learning, it is characterised in that include following device:

Divide module, for being K subimage by facial image X to be detected according to certain regular partition, then K subimage is converted into corresponding subpattern, and the vector of described subpattern is designated as X_i(i=1,2 ..., K);

Data Dimensionality Reduction module, for by formula Y_i=W_i ^TX_iObtain Y_i; Described Y_iIt is X_iLow-dimensional vector representation, described W_i ^TIt is keep projection algorithm to be derived from by maximum margin criterion and local;

Classification and Identification module, being used for according to arest neighbors sorting algorithm is the face figure X to be detected low dimensional vector Y to its subpattern_iClassify, and there are K recognition result according to the K in training set sub-set of modes;Described training set is based on the set of the facial image preset that described regular partition is K sub-set of modes;

Obtaining recognition result module, according to the method for weighting, obtaining the described facial image X to be detected probability belonging to c people after being calculated by K recognition result is:

$P_c=\frac{1}{K}\underset{i=1}{\overset{K}{\Σ}}{\mathrm{Wg}}_i*q_i^c,$ Wherein

Then the recognition result drawing facial image X to be detected is: Identity (X)=argmax (p_c);

Wherein said Wg_iBeing the identification weights of i-th subpattern set, its computing formula is:Wherein K_ijIt is for each sample X of i-th subpattern set in described training set_ijNeighbour neutralize its number being in similar sample, described X_ijBeing a sample of i-th subpattern set in described training set, described sample refers to the coordinate representation of the vector of each subpattern set; The value of j is between 1 to N, and N refers to the total number presetting facial image in described training set.

Further, described rule is divided by the stripeds such as facial image.

Compared to existing technology, the beneficial effects of the present invention is: the mode of the stripeds such as employing of the present invention is to divide facial image, and this dividing mode can greatly retain the texture structure of face each several part, and then remains key feature information. Maximum margin criterion is applied in locality preserving projections algorithm simultaneously, the recognition result of facial image is drawn herein in connection with arest neighbors sorting algorithm, higher than the recognition accuracy of existing traditional LPP algorithm, improve the computational efficiency of tradition LPP algorithm simultaneously, reduce system consumption; The accuracy of identification of system, up to more than 90%, meets the identification requirement under complex environment.

Accompanying drawing explanation

Fig. 1 is provided by the invention one based on the flow chart of the Method of Data with Adding Windows of manifold learning;

Fig. 2 is the flow chart of data processing figure of a recognition of face intelligent safety and defence system provided by the invention.

Detailed description of the invention

Below, in conjunction with accompanying drawing and detailed description of the invention, the present invention is described further:

As it is shown in figure 1, the invention provides a kind of Method of Data with Adding Windows based on manifold learning, comprise the following steps:

S101: by treating that facial image X to be detected is K subimage according to certain regular partition, then K subimage being converted into corresponding subpattern, the vector of described subpattern is designated as: X_i(i=1,2 ..., K).

Described is divided by the stripeds such as facial image according to the mode of certain regular partition, and such dividing mode can greatly retain the texture structure of face various piece, thus remaining more key feature information; The sub-image data simultaneously divided is more low, computed also just lower than other manifold Method of Data with Adding Windows many.

S102: according to formula Y_i=W_i ^TX_iObtain Y_i; Described Y_iIt is X_iLow-dimensional vector representation, wherein said W_i ^TIt is keep projection algorithm to be derived from by maximum margin criterion and local.

In this step, formula Y_i=W_i ^TX_iRefer to the mapping between by High dimensional space data to lower dimensional space data. Object function for locality preserving projections algorithm (LPP) is:

$min\underset{j,k}{\Σ}{(y_{ij}-y_{ik})}^2{s}_{jk}---\left(a\right),$ Wherein S_jkFor X_ijAnd X_ikSimilarity.

Maximum margin criterion (MMC) is to project in lower-dimensional subspace by original sample so that generic data sample is compacter, and different classes of data sample separates as much as possible; Require after its conversion to keep between class distance to maximize in lower dimensional space. Then the object function of MMC is:

J=maxtr (S_b-S_w) (b),

Wherein, $s_b=\underset{i=1}{\overset{l}{\Σ}}{n}_i(m_i-m){(m_i-m)}^t,s_w=\underset{i=1}{\overset{l}{\Σ}}{\underset{j=1}{\overset{n_i}{\Σ}}(x_i-m_i)(x_i-m_i)}^T,$ M is the average vector of all samples in all training sets, m_iIt is the average vector of all samples of the i-th class subpattern, S in described training set_bIt is stroll matrix between class, S_wIt is stroll matrix in class, n_iBelong to the number of samples of the i-th class subpattern in described training set;Described sample refers to a sample in each subpattern in described training set, namely refers to the vectorial coordinate representation in each subpattern, is a sample set for each subpattern; The value of i is between 1 to K; Described sample refers to the coordinate representation of the vector in subimage.

In the method for the invention, MMC is applied in LPP algorithm, that is to say that formula (b) is applied in LPP algorithmic formula (a) the target formula of the LPP that must make new advances:

$\left\{\begin{array}{c}\min tr\left(W_i^T\right(X_i{\mathrm{LX}}_i^T-\left(S_b-S_w\right)\left)W_i\right)\\ s.t.W_i^T{X}_i{\mathrm{DX}}_i^T{W}_i=1\end{array}\right.---\left(c\right)$

Wherein D is diagonal matrix, and L, D are known. Above derivation is that those skilled in the art can be good at deriving, and no longer describes in detail here.

W can be uniquely determined from (c) formula_i ^TValue;

In order to determine W uniquely_i ^T, seek (c) formula with Lagrange Fructus Citri junoris, it may be assumed that

$(X_i{\mathrm{LX}}_i^T-(S_b-S_w\left)\right)W_i={\mathrm{\λ}}_i{X}_i{\mathrm{DX}}_i^T{W}_i---\left(d\right)$

Then can draw W_iIt isWithThe characteristic vector generated, λ_iIt is corresponding subpattern X_iEigenvalue.

Therefore W can uniquely be determined_i ^T, then can determine that low-dimensional mapping function: Y_i=W_i ^TX_i。

S103: the low-dimensional vector representation Y according to the subpattern that arest neighbors sorting algorithm is facial image X to be detected_iClassify, according to the K in training set sub-set of modes, there are K recognition result;

Described training set is the set of facial image set in advance, the wherein N width facial image of total P people, according to the described rule waiting striped to divide, the every width facial image in training set is divided into K subimage, wherein the size of each image is certain, be assumed to be H1*H2 pixel, then the dimension of each subimage translates into the vector of H1*H2/K; After image all of in training set is divided into subimage, subimage co-located in different images is combined into corresponding subpattern set, thus obtain common K different subpattern set, each of which subpattern set is to be composed of a plurality of samples, each subpattern set vector representation, then sample is exactly the data representation of each coordinate in this vector.

S104: according to the method for Nearest Neighbor with Weighted Voting, obtaining the described facial image X to be detected probability belonging to c people after being calculated by K recognition result is:

$P_c=\frac{1}{K}\underset{i=1}{\overset{K}{\Σ}}{\mathrm{Wg}}_i*q_i^c---\left(e\right),$

WhereinWherein said Wg_iIt is the identification weights of i-th subpattern set;

So the recognition result for facial image X to be detected is:

Identity (X)=argmax (p_c), that is to say the recognition result taking that maximum classification of probability as X.

It is be calculated obtaining by category label between each sample and its k nearest neighbor sample in this subpattern set to the final weights identified for each subpattern in each face; Described k nearest neighbor sample refers to by utilizing what Euclidean distance calculated to be in the sample in same subpattern set with it.

Then for i-th subpattern set, its identification weights Wg_iComputing formula is:

${\mathrm{Wg}}_i=\frac{1}{N*K}\underset{j=1}{\overset{N}{\Σ}}{K}_{ij}---\left(f\right),$

Wherein K_ijIt is for each sample X of i-th subpattern set in described training set_ijNeighbour neutralize its number being in similar sample, described X_ijBeing a sample of i-th subpattern set in described training set, described sample refers to the coordinate representation of the vector of each subpattern set; The value of j is between 1 to N, and N refers to the total number presetting facial image in described training set.

The present invention does training set by Yake face database and facial image to be detected is carried out Classification and Identification, obtain a result to be substantially better than and existing there is LPP (locality preserving projections) and SPLPP (subpattern locality preserving projections) algorithm, improve the computational efficiency of tradition LPP, also superior to other subpattern algorithm in the accuracy rate identified.Described Yake face database is created with control centre by Yake university computer vision. Specifically detailed introduction may refer to the document in background technology.

As in figure 2 it is shown, present invention also offers a kind of recognition of face intelligent safety and defence system, this system applies to the heretofore described Method of Data with Adding Windows based on manifold learning. This system includes with lower module:

Search module: search for for face, deploy to ensure effective monitoring and control of illegal activities in real time and the contrast of face.

Detection module: for the Face datection based on the hsv model space.

Processing module: the extraction of the rectification for face, the pretreatment to facial image and face characteristic.

Data Dimensionality Reduction module: for using herein described method to carry out the Data Dimensionality Reduction of facial image.

Identification module: be used for drawing recognition result.

This system uses a kind of Method of Data with Adding Windows based on manifold learning to be identified face so that the accuracy of identification of system, up to more than 90%, has substantially met the identification requirement under complex environment. Simultaneity factor real-time is good, carries aspect.

Present invention also offers a kind of and described a kind of device corresponding based on the Method of Data with Adding Windows of manifold learning, including following device:

Divide module, for being K subimage by facial image X to be detected according to certain regular partition, then K subimage is converted into corresponding subpattern, and the vector of described subpattern is designated as X_i(i=1,2 ..., K);

Data Dimensionality Reduction module, for by formula Y_i=W_i ^TX_iObtain Y_i; Described Y_iIt is X_iLow-dimensional vector representation, described W_i ^TIt is keep projection algorithm to be derived from by maximum margin criterion and local;

Classification and Identification module, being used for according to arest neighbors sorting algorithm is the face figure X to be detected low dimensional vector Y to its subpattern_iClassify, and there are K recognition result according to the K in training set sub-set of modes; Described training set is based on the set of the facial image preset that described regular partition is K sub-set of modes;

Obtain recognition result module, according to the method for weighting, be calculated K recognition result obtaining the described facial image X to be detected probability belonging to c people be:

$P_c=\frac{1}{K}\underset{i=1}{\overset{K}{\Σ}}{\mathrm{Wg}}_i*q_i^c,$ Wherein

Then the recognition result drawing facial image X to be detected is: Identity (X)=argmax (p_c);

Wherein said Wg_iBeing the identification weights of i-th subpattern set, its computing formula is:Wherein K_ijIt is for each sample X of i-th subpattern set in described training set_ijNeighbour neutralize its number being in similar sample, described X_ijBeing a sample of i-th subpattern set in described training set, described sample refers to the coordinate representation of the vector of each subpattern set; The value of j is between 1 to N, and N refers to the total number presetting facial image in described training set.

Further, described rule is divided by the stripeds such as facial image.

It will be apparent to those skilled in the art that can technical scheme as described above and design, make other various corresponding changes and deformation, and all these change and deformation all should belong within the protection domain of the claims in the present invention.

Claims (5)

1. based on the Method of Data with Adding Windows of manifold learning, it is characterised in that comprise the following steps:

S101: be K subimage according to certain regular partition by facial image X to be detected, is then converted into corresponding subpattern by K subimage, the vector of described subpattern is designated as X_i(i=1,2 ..., K);

S102: according to formulaObtain Y_i; Described Y_iIt is X_iLow-dimensional vector representation, described inIt is keep projection algorithm to be derived from by maximum margin criterion and local;

S103: be the facial image X to be detected low dimensional vector Y to its subpattern according to arest neighbors sorting algorithm_iClassify, according to the K in training set sub-set of modes, there are K recognition result; Described training set is based on the set of the facial image preset that described regular partition is K sub-set of modes;

S104: according to the method for weighting, obtaining the described facial image X to be detected probability belonging to c people after being calculated by K recognition result is:

$P_c=\frac{1}{K}\underset{i=1}{\overset{K}{\Σ}}{\mathrm{Wg}}_i*q_i^c,$ Wherein

Then the recognition result drawing facial image X to be detected is: Identity (X)=argmax (p_c);

Wherein said Wg_iBeing the identification weights of i-th subpattern set, its computing formula is:Wherein K_ijIt is for each sample X of i-th subpattern set in described training set_ijNeighbour neutralize its number being in similar sample, described X_ijBeing a sample of i-th subpattern set in described training set, described sample refers to the coordinate representation of the vector of each subpattern set; The value of j is between 1 to N, and N refers to the total number presetting facial image in described training set.

2. as claimed in claim 1 based on the Method of Data with Adding Windows of manifold learning, it is characterised in that described rule is divided by the stripeds such as facial image.

3. as claimed in claim 1 based on the Method of Data with Adding Windows of manifold learning, it is characterised in that described inIt is by comprising the concrete steps that maximum margin criterion and local maintenance projection algorithm are derived from:

The projection algorithm target formula obtaining new locality preserving projections algorithm that combines is kept to be according to maximum margin criterion and local:

$\left\{\begin{array}{c}\min tr\left(W_i^T\right(X_i{\mathrm{LX}}_i^T-\left(S_b-S_w\right)\left)W_i\right)\\ s.t.W_i^T{X}_i{\mathrm{DX}}_i^T{W}_i=1\end{array}\right.;$ Wherein, $s_b=\underset{i=1}{\overset{l}{\Σ}}{n}_i(m_i-m){(m_i-m)}^t,$ Wherein D is diagonal matrix, and L, D are known; M is the average vector of all samples in all training sets, m_iIt is the average vector of all samples of the i-th class subpattern, S in described training set_bIt is stroll matrix between class, S_wIt is stroll matrix in class, n_iBelong to the number of samples of the i-th class subpattern in described training set; The value of i is between 1 to K;

Thus formula can obtainUniquely determine value.

4. based on the Data Dimensionality Reduction device of manifold learning, it is characterised in that include following device:

Divide module, for being K subimage by facial image X to be detected according to certain regular partition, then K subimage is converted into corresponding subpattern, and the vector of described subpattern is designated as X_i(i=1,2 ..., K);

Data Dimensionality Reduction module, is used for passing through formulaObtain Y_i; Described Y_iIt is X_iLow-dimensional vector representation, described inIt is keep projection algorithm to be derived from by maximum margin criterion and local;

Classification and Identification module, being used for according to arest neighbors sorting algorithm is the facial image X to be detected low dimensional vector Y to its subpattern_iClassify, and there are K recognition result according to the K in training set sub-set of modes; Described training set is based on the set of the facial image preset that described regular partition is K sub-set of modes;

Obtaining recognition result module, according to the method for weighting, obtaining the described facial image X to be detected probability belonging to c people after being calculated by K recognition result is:

$P_c=\frac{1}{K}\underset{i=1}{\overset{K}{\Σ}}{\mathrm{Wg}}_i*q_i^c,$ Wherein

Then the recognition result drawing facial image X to be detected is: Identity (X)=argmax (p_c);

Wherein said Wg_iBeing the identification weights of i-th subpattern set, its computing formula is:Wherein K_ijIt is for each sample X of i-th subpattern set in described training set_ijNeighbour neutralize its number being in similar sample, described X_ijBeing a sample of i-th subpattern set in described training set, described sample refers to the coordinate representation of the vector of each subpattern set;The value of j is between 1 to N, and N refers to the total number presetting facial image in described training set.

5. as claimed in claim 4 based on the Data Dimensionality Reduction device of manifold learning, it is characterised in that described rule is divided by the stripeds such as facial image.