CN103984922B - Face identification method based on sparse representation and shape restriction - Google Patents

Abstract

The invention discloses a face identification method based on sparse representation and shape restriction, and belongs the field of image processing, computer vision and mode identification. The method comprises the following steps of firstly, marking the face position and the shape of each image to be identified; secondly, carrying out feature extraction on the image to be identified on the basis of initialized shapes; thirdly, carrying out image matching on the basis of a shape model; fourthly, carrying out sparse representation on the textural features of the image to be identified on the basis of texture features of training images to obtain a coefficient collection corresponding to the training images; fifthly, analyzing obtained coefficients, and taking the identity of the training image which corresponds to the maximum coefficient as an ultimate identification result. Compared with the prior art, the face identification method based on the sparse representation and the shape restriction is relative robust on selection of initial positions, can be applied to face identification under facial shape changing, can improve the face identification precision and the application range greatly, and has good popularization and application value.

Description

It is a kind of based on rarefaction representation and the face identification method of shape constraining

Technical field

The present invention relates to image procossing, computer vision, mode identification technology, specifically a kind of based on sparse Represent the face identification method with shape constraining.

Background technology

Recognition of face is that one kind has recognized a kind of more natural, more direct recognition method compared with fingerprint, retina and iris etc. Jing becomes the study hotspot of living things feature recognition of future generation, be directed to image procossing, computer vision, pattern recognition and Multiple subjects such as neutral net.But when face shape changes, traditional face identification method is difficult to prove effective.

Find by prior art documents, it is mainly logical currently for the face identification method of face shape change Crossing the mode of shape conversion is carried out.H. Mohammadzade and D. Hatzinakos exist《IEEE Transactions on Affective Computing》(affection computation, IEEE magazines, vol.4, no.1,69-82,2013) on deliver " Projection into Expression Subspaces for Face Recognition from Single Sample per Person"（Single image face identification method based on expression subspace projection）.This article is extracted using expression subspace The face characteristic insensitive to expression shape change, solves impact of the expression to recognition of face.But, positioning of this method to face Requirement is very high, needs accurately position of human eye；And the method use training set come obtain expression subspace, it may appear that mistake Fitting phenomenon, i.e., the method failure when new samples are different from training set.The two shortcomings have impact on the performance of the method.

In addition, Wagner, A. et al. exist《IEEE Transactions on Pattern Analysis and Machine Intelligence》（Pattern analyses and machine intelligence IEEE magazine, vol.34, no.2, pp. 372-386, 2012）On deliver " Toward a Practical Face Recognition System: Robust Alignment and Illumination by Sparse Representation”（With regard to actual face identification system：By rarefaction representation reality The alignment and illumination of existing robust）, this article proposes solve the problems, such as that initial alignment is inaccurate using the method for rarefaction representation first. But, the method cannot process facial change of shape, have impact on the use range of the method.

The patent documentation of Publication No. CN101667246A disclose a kind of " face identification method based on rarefaction representation ", A kind of " rarefaction representation recognition of face side based on constrained sampling is disclosed in the patent documentation of Publication No. CN101833654A Method ", these methods are all absorbed in character representation, only in the case of face alignment（Such as positioned according to eyes）Ability Reach good recognition effect.

So far, also nobody proposes that the recognition of face side inaccurately and in the case of facial change of shape can be being positioned Method.

The content of the invention

The technical assignment of the present invention is for above-mentioned the deficiencies in the prior art, there is provided one kind based on rarefaction representation and shape about The face identification method of beam.The method can be applicable to the recognition of face under face shape change, can greatly improve recognition of face Precision and the scope of application.

The present invention is initialized first to the shape and position of face, and to face coarse positioning is carried out, and obtains the devices such as eyes The position of official；Secondly, two width images are carried out with accurately mate using matching algorithm, and adds the shape constraining of face, eliminate shape Shape changes the impact to matching algorithm；Again, the result images of matching are carried out into feature extraction using triangulation, is obtained and shape The unrelated textural characteristics of shape, eliminate impact of the change of shape to recognizing；Finally, it is special to the texture for obtaining using sparse representation theory Levy and be identified, further eliminate the impact that change of shape is produced to textural characteristics.

Specifically, technical assignment of the invention is realized in the following manner：One kind based on rarefaction representation and shape about The face identification method of beam, it is characterised in that the method is comprised the following steps：

First, shape initialization

Face location in each width images to be recognized is marked with shape；

2nd, feature extraction is carried out to images to be recognized based on initialized shape

Shape based on images to be recognized extracts texture, forms images to be recognized texture feature vector, images to be recognized stricture of vagina Reason characteristic vector is uniquely corresponding with the identity of images to be recognized；

3rd, the images match based on shape

Training image is matched based on textural characteristics to be identified, obtains the textural characteristics of each training image, Images to be recognized texture feature vector is uniquely corresponding with the identity of training image；

Four, based on the training image textural characteristics obtained by step 3, the images to be recognized texture that step 2 is obtained Feature carries out rarefaction representation, obtains the coefficient sets corresponding with training image；

Five, the coefficient that step 4 is obtained is analyzed, the training picture identity corresponding to the coefficient of maximum is taken as most Whole recognition result.

Furthermore, it is understood that step one includes：

A, to a width images to be recognized, extract face feature point using labeling method, carry out initialization operation；

B, determine shape, the shape representation for making each image in shape is, each image is expressed as one Average shapeWithThe linear combination of individual shape vector.

Above-mentioned labeling method of stating is preferably active shape model or active apparent model.

The concrete grammar of step 2 is：The human face region of images to be recognized is carried out using Delaunay Triangulation drawing Point, then the textural characteristics of images to be recognized are mapped to by average shape according to point correspondence, obtain unrelated with shape treating The textural characteristics of identification image.

Step 3 includes：

The shape that a, the images to be recognized textural characteristics and step one realization that are obtained based on step 2 are trained, it is right Each width training image carries out images match, obtains its shape；

B, the human face region of training image is divided using Delaunay Triangulation, then according to point correspondence The textural characteristics of training image are mapped to into average shape, the textural characteristics unrelated with shape are obtained, the spy of training image is realized Levy expression.

The concrete grammar of step 4 is：Images to be recognized is expressed as linear group of training image using sparse expression theory Close, i.e.,

WhereinCoefficient corresponding to all training images；

Solve optimumProblem, L-1 norm optimization problems can be expressed as, i.e.,

And optimal solution is obtained by Augmented Lagrange Multiplier algorithms.

The analysis method of step 5 is to take maximum and take minimum residual method

Compared with prior art, the face identification method based on rarefaction representation and shape constraining of the invention is to initial position Selection compare robust, it is not necessary to expression is identified or is converted, error is reduced, higher discrimination is realized.

Description of the drawings

Accompanying drawing 1 is shape initialization schematic diagram in embodiment；

Accompanying drawing 2 is feature extraction schematic diagram in embodiment；

Accompanying drawing 3 is the images match figure in embodiment based on shape；

Accompanying drawing 4 is that sparse expression recognition methodss schematic diagram is based in embodiment.

Specific embodiment

The recognition of face based on rarefaction representation and shape constraining with reference to Figure of description with specific embodiment to the present invention Method is described in detail below.

Embodiment：

The present invention's is comprised the following steps based on the face identification method of rarefaction representation and shape constraining：

The first step, shape initialization

1. a pair width images to be recognized, using ASM methods face feature point, such as eyes, nose, mouth, face mask are extracted Deng as shown in Figure 1.This patent does not limit the quantity and extracting method of face feature point, and the present embodiment uses active shape mould Type carries out initialization operation.

2. the determination of shape.The shape representation for making each image in shape is .Each image can be expressed as an average shapeWithThe linear combination of individual shape vector：

（1）

Second step, feature extraction is carried out based on initialized shape to images to be recognized

The human face region of images to be recognized is divided using Delaunay Triangulation, then according to point correspondence The textural characteristics of images to be recognized are mapped to into average shape, the feature unrelated with shape is obtained, as shown in Figure 2.Treat knowledge The result that other image carries out feature extraction is expressed as。

3rd step, the images match based on shape

1st, the textural characteristics for being obtained based on second step and the first step realize the shape for training, and each width is trained Image carries out images match, obtains its shape, as shown in Figure 3.

The images to be recognized is made to be, training image is, shape conversion function representation is, shape conversion parameter is, The object function of matching is

（2）

The solution of the object function is an iterative process.First to parameterLinearisation is carried out, i.e.,

（3）

This is a least square problem and with closed solutions：

（4）

WhereinIt isJacobian matrix, i.e.,

（5）

For the gradient of training image,For derivative of the deformation function to deformation parameter,For Hessian matrixes Gauss-Newton is estimated：

（6）

Finally,（2）Middle parameterRenewal carry out in the following way：

（7）

Shape per piece image can pass through（1）Try to achieve.

2nd, the feature representation of training image.It is similar with images to be recognized, using Delaunay Triangulation to training figure The human face region of picture is divided, and then the textural characteristics of training image is mapped to into average shape according to point correspondence, is obtained Obtain the feature unrelated with shape.Feature obtained by training image collection can be expressed as

（8）

4th step, images to be recognized is expressed as the linear combination of training image using sparse expression theory, i.e.,

（9）

WhereinCoefficient corresponding to all training images.Solve optimumProblem, L-1 norms can be expressed as most Optimization problem, i.e.,

（10）

And optimal solution can be obtained by Augmented Lagrange Multiplier algorithms.

5th step, takes identity of the image identity corresponding to coefficient maximum in coefficient vector for images to be recognized, i.e.,

（11）

The process is as shown in Figure 4.

Claims (5)

1. it is a kind of based on rarefaction representation and the face identification method of shape constraining, it is characterised in that the method is comprised the following steps：

First, shape initialization

Face location in each width images to be recognized is marked with shape, including：

A, to a width images to be recognized, extract face feature point using labeling method, carry out initialization operation；

B, determine shape, the shape representation for making each image in shape is s, and each image is expressed as an average shape Shape s₀With the linear combination of n shape vector；

2nd, feature extraction is carried out to images to be recognized based on initialized shape

Shape based on images to be recognized extracts texture, forms images to be recognized texture feature vector, and images to be recognized texture is special Levy vector uniquely corresponding with the identity of images to be recognized；

3rd, the images match based on shape

Training image is matched based on textural characteristics to be identified, obtains the textural characteristics of each training image, wait to know Other image texture characteristic vector is uniquely corresponding with the identity of training image；

Four, based on the training image textural characteristics obtained by step 3, the images to be recognized textural characteristics that step 2 is obtained Rarefaction representation is carried out, the coefficient sets corresponding with training image are obtained；

Five, the coefficient that step 4 is obtained is analyzed, the training picture identity corresponding to the coefficient of maximum is taken as final knowledge Other result.

2. it is according to claim 1 based on rarefaction representation and the face identification method of shape constraining, it is characterised in that described Labeling method is active shape model or active apparent model.

3. it is according to claim 1 based on rarefaction representation and the face identification method of shape constraining, it is characterised in that step Two concrete grammar is：The human face region of images to be recognized is divided using Delaunay Triangulation, then according to point The textural characteristics of images to be recognized are mapped to average shape by corresponding relation, obtain the texture of the images to be recognized unrelated with shape Feature.

4. it is according to claim 1 based on rarefaction representation and the face identification method of shape constraining, it is characterised in that step Three include：

The shape that a, the images to be recognized textural characteristics and step one realization that are obtained based on step 2 are trained, to each Width training image carries out images match, obtains its shape；

B, the human face region of training image is divided using Delaunay Triangulation, then will be instructed according to point correspondence The textural characteristics for practicing image are mapped to average shape, obtain the textural characteristics unrelated with shape, realize the mark sheet of training image Reach.

5. it is according to claim 1 based on rarefaction representation and the face identification method of shape constraining, it is characterised in that step Four concrete grammar is：Images to be recognized is expressed as the linear combination of training image using sparse expression theory, i.e.,

T=A α (1)

Wherein, T is images to be recognized；A is training image；α is the coefficient corresponding to all training images；

The problem of optimum α is solved, L-1 norm optimization problems can be expressed as, i.e.,

$\begin{array}{c}{\mathrm{\α}}^{*}={\mathrm{argmin}}_x\left|\right|\mathrm{\α}|{|}_1+|\left|e\right|{|}_1\\ s.t.y=A\mathrm{\α}+e.\end{array}---\left(2\right)$

And optimal solution is obtained by Augmented Lagrange Multiplier algorithms.