CN104899578A - Method and device for face identification - Google Patents

Abstract

The invention discloses a face recognition method, which includes: taking the obtained face image data as a sample to be tested; using a projection transformation matrix to map the sample to be tested into a low-dimensional feature space, and obtaining the projected test sample; in the training sample set, find the standard sample closest to the test sample distance as the target sample; the category of the target sample is determined as the category of the test sample; wherein, the projection transformation matrix is constructed by The intra-class adjacency matrix and the inter-class adjacency matrix are transformation matrices obtained by training multiple samples in the training sample set, so as to maximize the inter-class distance and minimize the intra-class distance. The method and device for face recognition provided by the present invention respectively construct two adjacency matrices for orthogonal discriminant projection: inter-class and intra-class adjacency matrices, which represent intra-class information and inter-class information separately to obtain balanced information , so as to achieve the purpose of the minimum within the class and the maximum among the classes.

Description

Method and device for face recognition

technical field

The invention relates to the field of computer vision, in particular to a face recognition method and device.

Background technique

At present, face recognition technology has developed into a very popular research topic in computer vision, and it is also one of the most successful applications in the field of image analysis. Face data is a typical high-dimensional small-sample data, and dimensionality reduction for face data is a necessary preprocessing step. In recent decades of development, a series of dimensionality reduction techniques have been proposed one after another.

The orthogonal discriminant projection methods proposed so far only construct an adjacency graph, which contains information within and between classes. In the case of unbalanced data distribution, the intra-class and inter-class information will also be unbalanced in the adjacency graph, which will lead to the inability to achieve the purpose of the smallest intra-class distance and the largest inter-class distance.

Contents of the invention

The purpose of the present invention is to provide a method and device for face recognition, aiming to solve the problem that the minimum within a class and the maximum between classes cannot be achieved in the prior art.

In order to solve the above-mentioned technical problems, the present invention provides a method for face recognition, including:

The acquired face image data is used as a sample to be tested;

Using a projection transformation matrix to map the sample to be tested into a low-dimensional feature space to obtain a projected test sample;

In the training sample set, find the standard sample closest to the test sample as the target sample;

determining the class of the target sample as the class of the test sample;

Wherein, the projection transformation matrix is a transformation matrix obtained by training multiple samples in the training sample set through the constructed intra-class adjacency matrix and inter-class adjacency matrix, so that the inter-class distance is the largest and the intra-class distance is the smallest .

Optionally, the projection transformation matrix is an intra-class adjacency matrix and an inter-class adjacency matrix constructed, and a transformation matrix obtained by training a plurality of samples in the training sample set includes:

Through the constructed intra-class ^adjacency matrix ^{F w} and _inter -class ^{adjacency matrix F b} , the _intra ^- class local Scatter matrix S ^w and between-class local scatter matrix S ^b ;

Calculate the projection transformation matrix P through the intra-class local scatter matrix S ^w and the inter-class local scatter matrix S ^b , so that the inter-class distance is the largest and the intra-class distance is the smallest;

in,

t > 0, and are the similar and heterogeneous neighbor sets of x _i respectively, D ^w and D ^b are both diagonal matrices.

Optionally, the determining the projection transformation matrix P through the intra-class local scatter matrix S ^w and the inter-class local scatter matrix S ^b includes:

Perform generalized eigendecomposition on the intra-class local scatter matrix S ^w and the inter-class local scatter matrix S ^b , arrange the obtained eigenvalues in order from large to small, and take the first d eigenvalues corresponding to The feature vector is used as the projection transformation matrix P, where d is the dimension of the space after projection transformation.

Optionally, the training sample set is a pre-established set, and the pre-established process includes:

The obtained multiple face images are used as training samples;

Using the projection transformation matrix to map the training samples into a low-dimensional feature space to obtain projected standard samples;

The standard samples and the known categories of the face images are stored as the training sample set.

Optionally, in the training sample set, finding the standard sample closest to the test sample as the target sample includes:

Using the nearest neighbor classifier, in the training sample set, find the standard sample closest to the test sample as the target sample.

The present invention also provides a device for face recognition, including:

An acquisition module, configured to use the acquired face image data as a sample to be tested;

A mapping module, configured to map the sample to be tested to a low-dimensional feature space using a projection transformation matrix to obtain a projected test sample;

A search module, configured to search for the standard sample closest to the test sample in the training sample set as the target sample;

a determination module, configured to determine the category of the target sample as the category of the test sample;

Wherein, the projection transformation matrix is a transformation matrix obtained by training multiple samples in the training sample set through the intra-class adjacency matrix and the inter-class adjacency matrix constructed by the training module, so that the inter-class distance is the largest and the intra-class Minimum distance.

Optionally, the training module includes:

A training acquisition unit, configured to use the acquired multiple face images as training samples;

A training mapping unit, configured to use the projection transformation matrix to map the training samples into a low-dimensional feature space to obtain projected standard samples;

The training storage unit is configured to store the standard samples and the known categories of the face images as the training sample set.

Optionally, the search module is used to find the standard sample closest to the test sample in the training sample set as the target sample including:

The search module is specifically configured to use the nearest neighbor classifier to find the standard sample closest to the test sample in the training sample set as the target sample.

The method and device for face recognition provided by the present invention use a projection transformation matrix to map the acquired samples to be tested into a low-dimensional feature space to obtain projected test samples. Then, in the training sample set, find the standard sample closest to the test sample as the target sample, and determine the category of the target sample as the category of the test sample, so as to achieve the purpose of face recognition. The method and device for face recognition provided by the present invention respectively construct two adjacency matrices for orthogonal discriminant projection: inter-class and intra-class adjacency matrices, which represent intra-class information and inter-class information separately to obtain balanced information , so as to achieve the purpose of the minimum within the class and the maximum among the classes.

Description of drawings

Fig. 1 is the method flowchart of a kind of embodiment of the method for face recognition provided by the present invention;

Fig. 2 is a flow chart of the process of determining the projection transformation matrix in another embodiment of the method for face recognition provided by the present invention;

Fig. 3 is the flowchart of the process of pre-establishing the training sample set in another embodiment of the method for face recognition provided by the present invention;

Figure 4 is a curve diagram of the classification accuracy of the three algorithms as the dimension changes;

Fig. 5 is a structural block diagram of a specific embodiment of the device for face recognition provided by the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Apparently, the described embodiments are only some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

A method flow chart of a specific embodiment of the method for face recognition provided by the present invention is shown in Figure 1, and the method includes:

Step S101: taking the acquired face image data as samples to be tested;

Step S102: using a projection transformation matrix to map the sample to be tested into a low-dimensional feature space to obtain a projected test sample;

Step S103: In the training sample set, find the standard sample closest to the test sample as the target sample;

Step S104: determining the category of the target sample as the category of the test sample;

Wherein, the projection transformation matrix is a transformation matrix obtained by training multiple samples in the training sample set through the constructed intra-class adjacency matrix and inter-class adjacency matrix, so that the inter-class distance is the largest and the intra-class distance is the smallest .

The face recognition method provided by the present invention uses a projection transformation matrix to map the obtained samples to be tested into a low-dimensional feature space to obtain projected test samples. Then, in the training sample set, find the standard sample closest to the test sample as the target sample, and determine the category of the target sample as the category of the test sample, so as to achieve the purpose of face recognition. The face recognition method provided by the present invention constructs two adjacency matrices respectively for the orthogonal discriminant projection: inter-class and intra-class adjacency matrices, which separately represent intra-class information and inter-class information to obtain balanced information, thereby To achieve the purpose of the minimum within the class and the maximum among the classes.

It should be pointed out that the intra-class in the present invention refers to the relationship between samples of the same class; the inter-class refers to the relationship between samples of different classes.

The present invention provides another specific implementation of the method for face recognition. Compared with the previous embodiment, this embodiment increases the process of determining the projection transformation matrix, as shown in Figure 2:

Step S201: Through the constructed intra-class adjacency matrix F ^w and inter-class adjacency matrix F ^b , calculate according to S _w =X(D ^w -F ^w )X ^T and S _b =X(D ^b -F ^b )X ^T Intra-class local scatter matrix S ^w and between-class local scatter matrix S ^b ;

Step S202: Calculate the projection transformation matrix P through the intra-class local scatter matrix S ^w and the inter-class local scatter matrix S ^b , so that the inter-class distance is the largest and the intra-class distance is the smallest;

in,

t > 0, and are the similar and heterogeneous neighbor sets of x _i respectively, D ^w and D ^b are both diagonal matrices.

As a preferred implementation, the determination of the projection transformation matrix P through the internal local scatter matrix S ^w and the inter-class local scatter matrix S ^b can be further specifically:

Perform generalized eigendecomposition on the intra-class local scatter matrix S ^w and the inter-class local scatter matrix S ^b , arrange the obtained eigenvalues in order from large to small, and take the eigenvectors corresponding to the first d eigenvalues As the projection transformation matrix P, where d is the dimension of the space after projection transformation.

After determining the projection transformation matrix, this embodiment also provides a process of pre-establishing a training sample set, as shown in Figure 3:

Step S301: using the obtained multiple face images as training samples;

Step S302: using the projection transformation matrix to map the training samples into a low-dimensional feature space to obtain projected standard samples;

Step S303: Store the standard samples and the known categories of the face images as the training sample set.

The present invention also provides another specific implementation of the face recognition method. In this embodiment, the ORL face database contains 400 face images of 40 people; 10 images for each person. Some of the images of faces were taken at different times. People's facial expressions and facial details have different degrees of changes, such as opening or closing eyes, wearing glasses or not wearing glasses, smiling or not smiling; facial postures also have considerable changes, depth rotation and plane rotation can reach 200; the size of each image is 32×32 pixels, and each pixel has 256 gray levels. Randomly select 50% from the database as training samples, and the remaining 50% as testing samples, repeat random sampling 10 times, and report the average result.

Specifically, this embodiment includes a process of establishing a face training data set after training and classifying images through the face training data set.

Let the existing face training data set be Where x _i ∈ R ^D is a certain face data, y _i ={1,2,...,c} indicates the category label of x _i , c indicates the number of categories, N indicates the total number of training samples, and D indicates the number of training samples dimension.

In this embodiment, N=200, c=40, D=1024. Of course, it can also be other numerical values, which do not affect the realization of the present invention.

In order to consider maintaining the geometric features of low-dimensional coordinates and training point information at the same time, find an optimal transformation P, and convert the data set Mapped to a relatively low-dimensional feature space, such as a d-dimensional space, and d<<D. In this low-dimensional feature space, the inter-class distance is maximized and the intra-class distance is minimized, namely:

$\begin{array}{cc}\underset{\mathrm{<span\; class="notranslate">\; P</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; x</span>}}& \mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; b</span>}}\mathrm{<span\; class="notranslate">\; P</span>}}{{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span>\end{array}$

Where trace is the matrix trace function, S _b is the inter-class local scatter matrix, and S _w is the intra-class local scatter matrix. To compute these two local scatter matrices, we construct two adjacency matrices, the within-class adjacency matrix F ^w and the between-class adjacency matrix F ^b . Then S _w ＝X(D ^w -F ^w )X ^T and S _b ＝X(D ^b -F ^b )X ^T , where D ^w and D ^b are both diagonal matrices, and F ^w and F ^b are defined as follows:

and

Where t>0 is the parameter of the function, and are the similar and heterogeneous neighbor sets of _xi , respectively. In this embodiment, t=8.

To obtain P, we perform generalized _{eigendecomposition} on Sb and _Sw . Sort the obtained eigenvalues in descending order, and take the eigenvectors corresponding to the first d eigenvalues to form a matrix P=[p ₁ ,p ₂ ,…,p _d ], where p _i is the eigenvectors of .

After obtaining the projection matrix P, the samples of the original sample space are projected to the low-dimensional feature space by projection, z _i =P ^T _xi , where z _i is the projection of x _i in the low-dimensional space, z _i ∈ R ^d . make is the projected training sample set. In this example, the value of d varies from 1 to 50.

For a test sample x∈R ^D , use the projection transformation P to map it to the low-dimensional feature space, and obtain the projected test sample z=P ^T x∈R ^d .

Using the nearest neighbor classifier, classify the projected test sample z in the low-dimensional feature space. That is to say, in the training sample set , find the sample closest to the test sample, and then assign the category of the sample to the projected test sample z. This completes the classification of x. In this embodiment, there are 200 samples to be tested, and the classification module is repeated 200 times.

Figure 4 shows the curves of the classification accuracy of the three algorithms as the dimension changes. The three comparison methods are: Orthogonal Discriminant Projection (ODP), Discriminant Neighbor Embedding (DNE) and the present invention. It can be seen that the recognition rate of the present invention is higher than the other two methods. Table 1 shows the comparison of the best performance of the three methods when the dimensionality reduction is between 1 and 50, and the corresponding best dimensions are in parentheses. The present invention achieves the best performance at lower dimensions.

Table 1

A structural block diagram of a specific embodiment of the device for face recognition provided by the present invention is shown in Figure 5, the device includes:

Acquisition module 100, is used for using the face image data that obtains as the sample to be tested;

A mapping module 200, configured to use a projection transformation matrix to map the sample to be tested into a low-dimensional feature space to obtain a projected test sample;

A search module 300, configured to search for a standard sample closest to the test sample in the training sample set as a target sample;

A determining module 400, configured to determine the category of the target sample as the category of the test sample;

Wherein, the projection transformation matrix is a transformation matrix obtained by training multiple samples in the training sample set through the intra-class adjacency matrix and the inter-class adjacency matrix constructed by the training module 500, so that the inter-class distance is the largest and the class The inner distance is the smallest.

The device for face recognition provided by the present invention uses a projection transformation matrix to map the obtained samples to be tested into a low-dimensional feature space to obtain projected test samples. Then, in the training sample set, find the standard sample closest to the test sample as the target sample, and determine the category of the target sample as the category of the test sample, so as to achieve the purpose of face recognition. The device for face recognition provided by the present invention constructs two adjacency matrices respectively for the orthogonal discriminant projection: an inter-class adjacency matrix and an intra-class adjacency matrix, which separately represent intra-class information and inter-class information to obtain balanced information, thereby To achieve the purpose of the minimum within the class and the maximum among the classes.

The training module 500 in the face recognition device provided by the present invention may further include:

A training acquisition unit 501, configured to use acquired multiple face images as training samples;

A training mapping unit 502, configured to use the projection transformation matrix to map the training samples into a low-dimensional feature space to obtain projected standard samples;

The training storage unit 503 is configured to store the standard samples and the known categories of the face images as the training sample set.

As a specific implementation, the search module is used to find the standard sample closest to the test sample in the training sample set as the target sample including:

The search module utilizes the nearest neighbor classifier to find the standard sample closest to the test sample in the training sample set as the target sample.

Other specific settings and methods of the face recognition device provided by the present invention are similar and will not be repeated here.

The device for face recognition provided by the present invention uses a projection transformation matrix to map the obtained samples to be tested into a low-dimensional feature space to obtain projected test samples. Then, in the training sample set, find the standard sample closest to the test sample as the target sample, and determine the category of the target sample as the category of the test sample, so as to achieve the purpose of face recognition. The device for face recognition provided by the present invention constructs two adjacency matrices respectively for the orthogonal discriminant projection: an inter-class adjacency matrix and an intra-class adjacency matrix, which separately represent intra-class information and inter-class information to obtain balanced information, thereby To achieve the purpose of the minimum within the class and the maximum among the classes. Compared with the orthogonal discriminant projection algorithm, the present invention can deal with the problem of unbalanced distribution of data samples, and has higher recognition rate.

Each embodiment in this specification is described in a progressive manner, each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (8)

1. A method of face recognition, comprising:

taking the obtained face image data as a sample to be detected;

mapping the sample to be tested to a low-dimensional feature space by using a projection transformation matrix to obtain a projected test sample;

in a training sample set, searching a standard sample closest to the test sample as a target sample;

determining a category of the target sample as a category of the test sample;

the projective transformation matrix is a transformation matrix obtained by training a plurality of samples in the training sample set through a constructed intra-class adjacency matrix and an inter-class adjacency matrix, so that the inter-class distance is maximum and the intra-class distance is minimum.

2. The method of claim 1, wherein the projective transformation matrix is a transformation matrix obtained by training a plurality of samples in the training sample set through a constructed intra-class adjacency matrix and an inter-class adjacency matrix, and comprises:

by constructed intra-class adjacency matrix F^wAnd inter-class adjacency matrix F^bAccording to S_w＝X(D^w-F^w)X^TAnd S_b＝X(D^b-F^b)X^TCalculating to obtain an intra-class local divergence matrix S^wAnd inter-class local divergence matrix S^b；

By said intra-class local divergence matrix S^wAnd inter-class local divergence matrix S^bCalculating to obtain a projection transformation matrix P so as to enable the inter-class distance to be maximum and the intra-class distance to be minimum;

wherein,

t＞0，andare each x_iSet of homogeneous and heterogeneous neighbors of D^wAnd D^bAre all diagonal matrices.

3. The method of claim 2, wherein the passing of the intra-class local divergence matrix S^wAnd the inter-class local divergence matrix S^bDetermining the projective transformation matrix P comprises:

for the intra-class local divergence matrix S^wAnd the inter-class local divergence matrix S^bAnd carrying out generalized eigen decomposition, arranging the obtained eigenvalues in a descending order, and taking eigenvectors corresponding to the first d eigenvalues as the projection transformation matrix P, wherein d is the dimension of the space after projection transformation.

4. A method for face recognition as claimed in any one of claims 1 to 3, wherein the set of training samples is a pre-established set, and the pre-established process comprises:

taking the obtained multiple face images as training samples;

mapping the training sample to a low-dimensional feature space by using the projective transformation matrix to obtain a projected standard sample;

and storing the standard sample and the known class of the face image as the training sample set.

5. The method of any one of claims 1 to 3, wherein the searching for the standard sample closest to the test sample as the target sample in the training sample set comprises:

and searching a standard sample closest to the test sample in the training sample set by using a nearest neighbor classifier as a target sample.

6. An apparatus for face recognition, comprising:

the acquisition module is used for taking the acquired face image data as a sample to be detected;

the mapping module is used for mapping the sample to be tested to a low-dimensional feature space by using a projection transformation matrix to obtain a projected test sample;

the searching module is used for searching a standard sample closest to the test sample in the training sample set as a target sample;

a determining module for determining the category of the target sample as the category of the test sample;

the projection transformation matrix is obtained by training a plurality of samples in the training sample set through a constructed intra-class adjacent matrix and an inter-class adjacent matrix by a training module so as to enable the inter-class distance to be maximum and the intra-class distance to be minimum.

7. The apparatus for face recognition as defined in claim 6, wherein the training module comprises:

the training acquisition unit is used for taking the acquired multiple face images as training samples;

the training mapping unit is used for mapping the training sample to a low-dimensional feature space by using the projection transformation matrix to obtain a projected standard sample;

and the training storage unit is used for storing the standard sample and the known class of the face image as the training sample set.

8. The apparatus for face recognition according to claim 6, wherein the searching module is configured to search, as the target sample, a standard sample closest to the test sample in the training sample set, and includes:

the searching module is specifically configured to search, in the training sample set, a standard sample closest to the test sample as a target sample by using a nearest neighbor classifier.