WO2015165227A1

WO2015165227A1 - Human face recognition method

Info

Publication number: WO2015165227A1
Application number: PCT/CN2014/089652
Authority: WO
Inventors: 李俊
Original assignee: 珠海易胜电子技术有限公司
Priority date: 2014-04-28
Filing date: 2014-10-28
Publication date: 2015-11-05
Also published as: CN103902992A; CN103902992B

Abstract

The present invention provides a human face recognition method, comprising the following steps: S1: generating a human face elastic bunch graph; S2: generating a human face recognition model based on appearance, and obtaining, by means of calculation, the cosine similarity between a vector of the human face recognition model based on the appearance and a vector of an existing human face model in a database; S3: generating a human face recognition model based on geometrical characteristics, and obtaining, by means of calculation, the cosine similarity between a vector of the human face recognition model based on the geometrical characteristics and the vector of the existing human face model in the database; S4: using logistic regression mixing according to the similarity level in step S2 and the similarity level in step S3; and S5: judging a human face recognition result according to a result of step S4. The present invention adopts a human face recognition method that mixes on similarity levels a human face recognition method based on appearance and a human face recognition method based on geometrical characteristics, and can be well applied to the actual living environment.

Description

Untitled

Technical field:

The invention relates to a face recognition method.

Face recognition technology has developed rapidly in the past few years. Currently, face recognition technology cannot be satisfactorily handled in real life environments such as outdoor environments, and is only used indoors. The difficulty of face recognition is still illumination change, posture change, age change, occlusion, etc. These affect the face recognition algorithm used by the face recognition system, the degree is different, the classification of face recognition methods and its advantages and disadvantages The difficulty in face recognition that affects this is as follows:

A face recognition method based on appearance uses a pixel value of a face image pixel to generate a face template; a face recognition method based on a geometric feature does not rely on a pixel point, and is based on a feature point of the face (eye, nose, mouth, ear) ...) Generates a face template between geometric positional relationships. In the past, the face recognition method based on appearance compared with the face recognition method based on geometric features can extract very rich face features from each pixel of the image, so it is more than the geometric feature based on several feature points. Face recognition methods bring high face recognition performance, and most successful face recognition methods are based on appearance-based recognition methods.

However, the appearance-based recognition method can not cope well with the illumination changes that affect the pixel points. The geometric feature-based recognition method relies on the geometric positional relationship, which can compensate for the shortcomings of the appearance-based recognition method.

Since relying on the face feature points, it is required to accurately take out the face feature points in the previous stage. The face recognition method is based on the method of generating a face template to examine the overall face image or into a partial field, and is divided into a global face recognition method and a local face recognition method. The global face recognition method for examining the overall face image has the advantages of expressing both the local features and the global features of the face, but it has the disadvantage of not being able to cope with the change of the posture. On the contrary, the method of local face recognition for the change of posture is more than the method of global face recognition. Strong, there is a good response to people The advantage of the local characteristics of the face.

In the past, Elastic Bunch Graph Matching (EBGM) is a feature-based face recognition method. It belongs to the local face recognition method and is one of the most successful face recognition methods, but the local face recognition method. The disadvantage is that it does not reflect the global characteristics of the face. In order to overcome this, a method combining global face recognition method and local face recognition method has emerged, which brings some performance improvement. Both methods are based on the appearance-based face recognition method, which cannot overcome the appearance-based recognition method. Disadvantages.

In the actual living environment, the face image is difficult to face recognition due to changes in illumination, posture, age, and occlusion. Therefore, in the real life environment, face recognition technology can not be completed, for this purpose in-depth study. In recent years, a lot of research has been carried out in this field, and great progress has been made, but still not satisfactory.

Summary of the invention:

The technical problem to be solved by the present invention is to provide a face recognition method that effectively finds facial feature points, is independent of illumination changes, and is stable to posture changes.

The invention provides a face recognition method, which comprises the following steps:

S1: elastic face diagram of adult face;

S2: generating an appearance-based face recognition model, and calculating a cosine similarity between the appearance-based face recognition model and the existing face model vector in the database;

S3: generating a face recognition model based on geometric features, and calculating a cosine similarity between the obtained geometric feature based face recognition model and the existing face model vector in the database;

S4: using logical regression mixing based on the similarity level based on step S2 and step S3;

S5: Determine the face recognition result based on the result of step S4.

Further, a face elastic beam map is generated, and the face feature points are extracted according to the Haar feature in the detected face region.

Further, generating a facial elastic beam map, first extracting four points in the detected face region, respectively, the left and right eyeball midpoints, the mouth midpoints, and the lower jaw points, forming an initial partial face model, having 30 The template map of the feature points analyzes the relationship between each feature point and the four points of the initial part of the face model, generates a two-dimensional affine transformation, and applies this transformation on the 30 feature points of the template map to obtain The eigenvalues corresponding to the 30 feature points are used to obtain the initial global face model; all 30 feature points of the initial global face model seek the correct convergence point, and the face elastic beam is generated as the feature point. Figure.

Further, an appearance-based face recognition model is generated, and Gabor Jet is extracted from 30 feature points of the face elastic beam map, and the obtained vector is used as an initial model of the appearance-based face model, and the Gabor Jet complex number is extracted. Straight, composed of 40 vectors that are straight elements; the initial model of the face model uses PCA and LDA to obtain a face recognition model based on appearance.

Further, a face recognition model based on geometric features is generated, and the distance between the extracted face feature points is calculated, and PCA and LDA are applied to the feature vectors whose elements are proportional to the horizontal axis and the vertical axis direction component, and the basis is obtained. A face recognition model of geometric features.

The invention adopts a face recognition method which combines the appearance-based face recognition method and the geometric feature-based face recognition method at the similarity level, can be satisfactorily applied in the actual living environment, and proposes a more effective search. The method of face feature points also proposes a face recognition method based on geometric features that is independent of illumination changes and stable to posture changes.

BRIEF DESCRIPTION OF THE DRAWINGS:

The drawings described herein are intended to provide a further understanding of the invention, and are intended to be a part of the invention. In the drawing:

FIG. 1 is a schematic flow chart showing a face recognition method according to an embodiment of the present invention.

detailed description:

The invention will be described in detail below with reference to the drawings in conjunction with the embodiments.

Embodiments of the present invention provide a method for recognizing a face, including the following steps:

First, generate a face elastic beam diagram

In the present invention, firstly, four points are extracted in the detected face region, which are the left and right eyeball midpoints, the mouth midpoint and the lower jaw point, and form an initial partial face model, and a template having 30 feature points. In the figure, the relationship between each feature point and the four points of the initial part of the face model is analyzed, and a two-dimensional affine transformation is generated. The transformation is applied to the 30 feature points of the template image to obtain 30 feature points and The corresponding eigenvalues result in an initial global face model.

The transformation formula is as follows:

[Correct according to Rule 26 04.12.2014]
The change that needs to be made without giving up universality is.

The first item is the association with the extrusion and rotation, and the latter is the translation association.

The translational association is simply calculated as the difference between the focus of the two-part model.

[Correct according to Rule 26 04.12.2014]
which is

[Correct according to Rule 26 04.12.2014]

The focus of the initial part of the face model consisting of 4 points obtained at the beginning;

[Correct according to Rule 26 04.12.2014]

The focus of the partial face model consisting of 4 points in the prepared face model;

The four pixels of the matrix of rotation and stretching can be obtained from the relationship between two corresponding points of the two-part face model.

Part of the face model consists of 4 points, so that it has the smallest error, so that the corresponding 4 points are close to each other, and the linear regression transformation is obtained to obtain the rotation and stretching transformation matrix.

The obtained affine transformation is applied to the template image to generate an initial global face model.

G _I -T(G _T );

G _I : face elastic beam diagram;

G _T : template map;

After obtaining the initial global face model, the confidence level of the partial field centered on each feature point is obtained, and the point next to the point is also sought to be sure, and the corresponding feature point is updated with the point of high certainty.

If there is no point of great confidence, it will terminate.

This way each feature point can no longer be updated.

For all 30 feature points of the initial global face model, the correct convergence point is sought, and the face elastic beam map which is used as the feature point is generated.

In the present invention, the haar feature is used instead of the Gabor feature, and the face feature points are extracted according to the Haar feature in the detected face region, and the Haar feature is to replace the pixel values in each field to examine the pixel values in a certain field. That is, the feature is to detect the difference or combination of pixel values of various modes in the candidate domain; in order to improve the detection performance of the object, such Haar features must be rich, which is trained by the two-dimensional classifier-cascade classifier.

The detector using Haar features is faster and more accurate than other detectors, and is mostly used for object detection. Especially the face detector based on the Haar feature of viola-jones is the most successful detector; in the present invention, each is the most successful detector. The feature points are extracted from the large-capacity face database and the model centered on the point, and the viola-jones detector is used for training.

Second, generate appearance-based face recognition model and match

Gabor Jet is extracted from the 30 feature points of the face elastic beam map, and the vector obtained by the connection is used as the initial model of the appearance-based face model.

Gabor Jet is convolution of the Gabor filter for the pixel of interest. Got it.

The convolution of the Gabor filter and image is calculated using the following formula:

[Correct according to Rule 26 04.12.2014]
The Gabor filter is as follows:

[Correct according to Rule 26 04.12.2014]

Vector basis

The type of Gabor filter is determined. In the invention, a total of 40 Gabor filters are formed for 5 frequencies and 8 directions.

[Correct according to Rule 26 04.12.2014]

That is, the Gabor filter is a set of 40 complex coefficients:

[Correct according to Rule 26 04.12.2014]

In the present invention, the magnitude of the Gabor Jet complex is taken out, and the vector consists of 40 straight elements.

Appearance-based face recognition model is obtained by applying PCA and LDA to the initial model of the face model.

In the matching phase of the two face models, the cosine similarity between the appearance-based face recognition model obtained above and the existing face model vector in the database is calculated.

[Correct according to Rule 26 04.12.2014]

Third, generate a face recognition model based on geometric features and matching.

In the past, as a representative face recognition method based on geometric features, a face recognition method based on distance and a face recognition method based on scale can be mentioned. The original scale-based face recognition method uses the distance between face feature points. Proportion, but this method is also different for the corresponding distance of the stereoscopically rotated image, and has the disadvantage of being unstable to the posture change.

In order to overcome such disadvantages, in the present invention, the horizontal direction component and the vertical direction component are independent according to the distance. The site was inspected.

For two line segments of the same plane, when the plane rotates, the ratio of the components of the rotation direction of the two line segments is the same, and the length and proportion of the component in the direction orthogonal to the direction of rotation are the same.

The human head rotates and generally can only rotate in the vertical and horizontal directions.

That is, a line segment connecting feature points of almost the same plane in the face model is extracted, and the face templates are extracted based on the fact that the ratio of the horizontal and vertical axis components of the corresponding line segment is constant for the same person in the two face models.

The face template generation phase is as follows:

In the feature points obtained above, except for the points where the nose and the ear are mainly curved, the points which are almost in the same plane are selected, and the order between them is not considered, and the possible double (combination) is made. Find the distance between each pair of vertical and horizontal axes.

[Correct according to Rule 26 04.12.2014]

DH _i : ith horizontal distance

DV _i : ith vertical distance

(X _i , Y _i ): coordinates of the jth node

n: the number of feature points

Then, according to the order of the direction axes, each direction is made into a possible double (combination), and for each double correspondence, the ratio of the distances included in the double is included.

RH _i =DH _j /DH _k ,j,(≠k)=1,...,m,i=1,...,mC2

RV _i = DV _j / DV _k , j, (≠k) = 1, ..., m, i = 1, ..., mC2

RH _i : the i-th horizontal ratio

RV _i : ith vertical ratio

m: the number of distances on one axis

[Correct according to Rule 26 04.12.2014]

Generates a vector whose scale is obtained by the direction axis and connects them to each other to generate the original template.

V _ο -(RH ₁ , RH ₂ ,...,RH _t ,RV ₁ ,RV ₂ ,...,RV _t ),t-mC2

The original template vector thus obtained contains many unnecessary features and the recognition ability is not high.

That is, the PCA (alone by the direction axis) is applied here to remove the unnecessary components and then perform vector reduction.

The LDA is obtained for the obtained reduced vector, and a geometric feature-based face recognition model with high recognition power is generated.

In the matching phase of the two face models, the cosine similarity between the geometric feature-based face recognition model obtained from above and the existing face model vector in the database is calculated, and the cosine similarity calculation method is based on the appearance. The face recognition method is the same.

Fourth, mixing

The similarity level between the appearance-based face recognition method and the geometric feature-based face recognition method is mixed using logistic regression, and the formula is as follows:

[Correct according to Rule 26 04.12.2014]

X ₁ ; appearance-based method distance X ₂ ; geometric feature-based method distance,

β ₀ , β ₁ , β ₂ : Logistic regression coefficient.

The invention adopts a face recognition method which combines the appearance-based face recognition method and the geometric feature-based face recognition method at the similarity level, and can be satisfactorily applied to the actual living environment, and proposes It is more effective to find the method of facial feature points. It also proposes a geometric feature-based face recognition method that is independent of illumination changes and stable to posture changes.

The above description is only the preferred embodiment of the present invention, and is not intended to limit the present invention, and various modifications and changes can be made to the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and scope of the present invention are intended to be included within the scope of the present invention.

Claims

A face recognition method, comprising the steps of:

S1: generating a facial elastic beam map;

S2: generating an appearance-based face recognition model, and calculating a cosine similarity between the appearance-based face recognition model and the existing face model vector in the database;

S3: generating a face recognition model based on geometric features, and calculating a cosine similarity between the obtained geometric feature based face recognition model and the existing face model vector in the database;

S4: using logical regression mixing based on the similarity level based on step S2 and step S3;

S5: Determine the face recognition result based on the result of step S4.
The face recognition method according to claim 1, wherein a face elastic beam map is generated, and a face feature point is extracted based on the Haar feature in the detected face region.
The face recognition method according to claim 2, wherein the facial elastic beam map is generated, and firstly, four points are extracted in the detected face region, which are the left and right eyeball midpoints, the mouth midpoint, and The lower point constitutes the initial part of the face model, and the relationship between each feature point and the four points of the initial part of the face model is analyzed in a template map with 30 feature points to generate a two-dimensional affine transformation in the template diagram. Applying this transformation to the 30 feature points, the eigenvalues corresponding to 30 feature points are obtained, and the initial global face model is obtained. For all 30 feature points of the initial global face model, the correct convergence is sought. Point, generate a face elastic beam map using this as a feature point.
The face recognition method according to claim 2, wherein a face recognition model based on appearance is generated, and Gabor Jet is extracted from 30 feature points of the face elastic beam map, and the vector obtained by connecting is based on the appearance. The initial model of the face model, taking out the width of the Gabor Jet complex, consisting of 40 vectors that are straight elements; applying the PCA and LDA to the initial model of the face model. A face recognition model based on appearance.
The face recognition method according to claim 4, wherein a face recognition model based on geometric features is generated, and a distance between the extracted face feature points is calculated, and between the components in the horizontal axis and the vertical axis direction The feature vector of the feature is applied to PCA and LDA to obtain a face recognition model based on geometric features.