CN103198309B - Based on the eyebrow recognition method of rarefaction representation - Google Patents

Abstract

The invention belongs to electronic information technical field, disclose a kind of eyebrow recognition method based on rarefaction representation, comprising: gather original eyebrow image, be partitioned into pure eyebrow image after gray processing, and be normalized; Every pure eyebrow image array is connected by row also transposition and become column vector; Compute vector average; Calculated population Scatter Matrix; Obtain front m eigenwert and the proper vector of overall Scatter Matrix C; Calculate the inner product of eyebrow sample and proper vector; The vector approximation of structure eyebrow sample; Input test eyebrow image, and to its process, obtain the vector approximation testing eyebrow image; Calculate optimum sparse coefficient vector; Calculate the difference of test eyebrow and image; Determine to test the classification belonging to eyebrow.The present invention not only increases the discrimination of eyebrow, also enhances image in the robustness be corroded and identify in contaminated situation.In addition, the present invention can be used for the identification of the other biological features such as face, fingerprint, iris, and extensibility is good.

Description

Based on the eyebrow recognition method of rarefaction representation

Technical field

The invention belongs to electronic information technical field, be specifically related to a kind of eyebrow recognition method based on rarefaction representation.

Background technology

Eyebrow, as the important component part of on face, has relevant literature research to show: the effect of eyebrow in face is even greater than eyes, but eyebrow is not but almost still concerned in recognition of face.In fact, eyebrow has distinct profile, and its feature is simple, is easy to extract.Compared with face, the impact that it is subject to the factor such as expression shape change and age growth is less; And compared with iris, it is easier to gather.Although various evidence shows: the shape of eyebrow is varied, there is good identity specificity, purely utilize eyebrow to carry out the still little of Study of recognition.

Eyebrow recognition is the new biometrics identification technology risen in recent years, has obtained certain cognitive psychology evidence in the world.Single facial characteristics is the most important factor carrying out identifying, but the facial characteristics that people rely on when carrying out face recognition most seems not to be eyes but eyebrow, and that is when identifying, the effect of eyebrow is even greater than the effect of eyes.2007; in the paper " eyebrow recognition method that a kind of feature based string compares " that Li Yu Monitoring and pair flower worked with kingfisher's feathers are delivered on Beijing University of Technology's journal, propose the idea utilizing eyebrow to carry out person identification, and conduct a preliminary study with regard to the method for eyebrow recognition, demonstrate possibility and validity that eyebrow recognition differentiates for personal identification by experiment.Recently, in the paper " Canyoureyebrowstellmewhoyouare " that JuefeiXu.F etc. deliver in SignalProcessingandCommunicationSystems meeting, also been proposed new evidence and prove that eyebrow can as a kind of biological characteristic independently for identifying, compare eyebrow, eyes and the textural characteristics of whole face under different characteristic extracting method, find the average recognition rate only accounting for the eyebrow of whole face 1/6 size, compare the discrimination of the whole face containing more texture and structural information, only have dropped 1/6.A large amount of experiments shows, although eyebrow only accounts for the sub-fraction of face, using it for identification has good effect.

Up to now, the eyebrow recognition method proposed mainly contains: eyebrow recognition method, Based PC A(PrincipalComponentAnalysis based on Discrete HMM (HiddenMarkovModel)) eyebrow recognition method, feature based string compare eyebrow recognition method, based on the eyebrow recognition method of semi-supervised learning and support vector machine and fast Template Matching method.These methods all obtain good recognition correct rate on existing eyebrow database.But they also also exist deficiency: namely when eyebrow image is corroded or is contaminated, recognition effect is unsatisfactory.

Summary of the invention

For the defect existed in the above-mentioned eyebrow recognition mentioned, the present invention proposes one sparse representation method and knowledge method for distinguishing is carried out to eyebrow, enhance discrimination and the robustness of eyebrow recognition.

Ultimate principle of the present invention: utilize test sample book to be that this principle of linear combination of training sample is classified to test sample book.To each eyebrow image zooming-out feature, eyebrow image is expressed as the form of a proper vector, and then test sample book and training sample are expressed as the form of Vector Groups.Then, by separating l ₁norm minimum obtains sparse solution, calculates test eyebrow and sparse solution generation of trying to achieve returns the error of former formula, and minimum that class of the error of calculation is exactly the classification of test belonging to eyebrow.

Based on an eyebrow recognition method for rarefaction representation, it is characterized in that comprising the following steps:

Step one, to M(≤5000) individual everyone gather p (>=2) original eyebrow image, be partitioned into pure eyebrow image after gray processing, and they be normalized to identical size (generally minimum is 10 × 10, maximum be no more than 2000 × 2000).

Step 2, connects by row by every pure eyebrow image array and transposition becomes column vector.

Step 3, is designated as x respectively by all N number of column vectors ₁, x ₂..., x _n, compute vector average μ, formula is:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_i$

Step 4, calculated population Scatter Matrix C, formula is:

$C=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}(x_k-\mathrm{\μ}){(x_k-\mathrm{\μ})}^T$

Step 5, front m the eigenvalue λ of calculated population Scatter Matrix C ₁, λ ₂..., λ _mwith corresponding unit character vector u ₁, u ₂..., u _m.

Step 6, calculates x _iand u _jinner product , namely .

Step 7, structure x _ivector approximation .

Step 8, the original eyebrow image of input test, is treated to gray scale image, is partitioned into the pure eyebrow image y of test, calculates y and u _jinner product , namely , the vector approximation of structure y. 。

Step 9, calculates optimum sparse coefficient vector s ^*=(s ₁, s ₂..., s _n) ^t, formula is as follows:

$s^{*}={\arg \min }_s{\left|\right|s\left|\right|}_1\mathrm{subject\; to}{\left|\right|\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{s}_i{\tilde{x}}_i-\tilde{y}\left|\right|}_2\≤\mathrm{\ϵ}$

In formula, s=(s _i1, s _i2..., s _iN) ^t, || || ₁for l ₁norm, represent vector each element absolute value and, || || ₂for l ₂norm.

Step 10, calculates the difference of test eyebrow and image , formula is as follows:

$r_i\left(\tilde{y}\right)=\left|\right|{\tilde{y}-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\δ}}_i(s^{*},{\tilde{x}}_j)\left|\right|}_2$

In formula,

Step 11, determine to test the classification belonging to original eyebrow, formula is as follows:

$\mathrm{identity}\left(y\right)=\underset{1\≤i\≤M}{\arg \min }{r}_i\left(\tilde{y}\right)$

In formula, identity(y) for testing the classification belonging to original eyebrow, for minimum that corresponding image category.

The present invention compared with prior art, has following beneficial effect:

The present invention extracts eyebrow proper vector by the method for principal component analysis (PCA) in eyebrow recognition process, eyebrow recognition is carried out by the method for rarefaction representation, not only increase the discrimination of eyebrow recognition, also enhance image in the robustness be corroded and identify in contaminated situation.In addition, the present invention can be used for the identification of the other biological features such as face, fingerprint, iris, has good extensibility.

Accompanying drawing explanation

Fig. 1 is eyebrow recognition module structure drafting involved in the present invention;

Fig. 2 is method flow diagram involved in the present invention;

Fig. 3 is the pure eyebrow image used in application example of the present invention and by the image of noise pollution: (a) original image, (b) 20% is by the image of noise pollution, and (c) 60% is by the image of noise pollution.

Embodiment

Below in conjunction with drawings and the specific embodiments, the invention will be further described.

Hardware device involved in the present invention comprises image capture device and computing machine.The eyebrow image that image capture device obtains is sent into computing machine and is processed, and realizes eyebrow recognition.Concrete eyebrow recognition program module is formed as shown in Figure 1.

Based on the eyebrow recognition method of rarefaction representation process flow diagram as shown in Figure 2.Based on an eyebrow recognition method for rarefaction representation, it is characterized in that comprising the following steps:

Step one, to M people, everyone gathers p (>=2) original eyebrow image, is partitioned into pure eyebrow image, and they are normalized to identical size after gray processing.

Step 2, connects by row by every pure eyebrow image array and transposition becomes column vector.

Step 3, compute vector average.

Step 4, calculated population Scatter Matrix.

Step 5, obtains front 10-50 eigenwert and the proper vector of overall Scatter Matrix C.

Step 6, calculates the inner product of eyebrow sample and proper vector.

Step 7, the vector approximation of structure eyebrow sample.

Step 8, input test eyebrow image, and to its process, obtain the vector approximation testing eyebrow image.

Step 9, with Homotopy solution l ₁norm minimum problem, calculates optimum sparse coefficient vector.

Step 10, calculates the difference of test eyebrow and image.

Step 11, determines to test the classification belonging to eyebrow.

Provide the example that an application the present invention carries out eyebrow recognition below.

Utilize the digital image acquisition apparatus that image pick-up card CG300, CP240 Panasonic's video camera and 75mm high precision Japan import lens group are dressed up, and use Lenovo, Qi Tian, M80000 microcomputer, composition eyebrow acquisition system.System gathers the original eyebrow image of distance about 1 meter of user under general illumination condition.Gather eyebrow image totally 1118 width (everyone at least 10 width) of 109 people, and input in computing machine.The front 5 width eyebrow images of eyebrow recognition experiment employing 109 people are here as training sample, and rear 5 width eyebrow images are as test sample book.

(1) pure eyebrow is as the situation of test sample book

Image is adjusted to respectively the size of 10 × 10,100 × 100,200 × 200, then image is normalized, test according to the process of eyebrow recognition.Experimental result is as shown in table 1.Experiment shows, along with the increasing of pixel of pure eyebrow image zooming-out, discrimination can be improved to some extent.When eyebrow image reaches 200 × 200, the method for rarefaction representation can reach the discrimination of 98.53%; Even if when eyebrow image is 10 × 10, the method still can reach the discrimination of 95.05%.

Under the same conditions, adopt the experimental result of arest neighbors method, recently subspace method and SVM method respectively as shown in table 2,3,4.

Table 1 sparse representation method experimental result

Table 2 arest neighbors methods experiment result

Table 3 is subspace method experimental result recently

Table 4 support vector machine method experimental result

(2) eyebrow image is by situation during noise pollution

For the ease of the number percent that control noises accounts in the picture, what add is salt-pepper noise herein, as schemed shown in attached 3.

When image size is adjusted to 100 × 100, adopt the experimental result of distinct methods as shown in table 5.Experiment shows, when image size is adjusted to 100 × 100, the tested contaminated degree of eyebrow image from 0% to 90%, along with the increase discrimination of contaminated degree declines all to some extent, but when pollution reaches 40%, the method for rarefaction representation still can reach the discrimination of 90.3%.Under the same conditions, compare with nearest neighbor method, recently subspace method and support vector machine method, the method for rarefaction representation has better discrimination, and that is, the method for rarefaction representation has stronger robustness.

By the recognition effect after noise pollution when table 5 eyebrow image is 100 × 100 size

Experimental result when image size is adjusted to 10 × 10 is as shown in table 6.The tested contaminated degree of eyebrow image from 0% to 50%, along with the increase discrimination of contaminated degree can decline to some extent, but when pollute reach 30% time rarefaction representation method still can reach 79.8% discrimination.

By the recognition effect after noise pollution when table 6 eyebrow image is 10 × 10 size

As can be seen from above experimental result, Beijing Polytechnical University's eyebrow database carries out identification and obtains 98.5% discrimination, compare with recognition of face and reach suitable recognition effect at its best, the possibility that the eyebrow demonstrating the mankind again uses as a kind of biological characteristic and feasibility.Sorting technique based on rarefaction representation is applied in eyebrow recognition has certain robustness.For by the eyebrow image of noise pollution and the eyebrow image that is at least partially obscured, the method for rarefaction representation has good recognition effect.According to current bibliographical information, the face identification rate based on the sorting technique of rarefaction representation is 92.0% ~ 100%.Therefore eyebrow likely reaches suitable level with recognition of face in identity verify, thus may replace recognition of face in some applications. the validity of sufficient proof the method and superiority.

Above embodiment is only in order to illustrate the present invention, and and unrestricted technical scheme described in the invention.Therefore, all do not depart from technical scheme and the improvement thereof of the spirit and scope of the present invention, all should be encompassed in right of the present invention.

Claims (1)

1., based on an eyebrow recognition method for rarefaction representation, it is characterized in that comprising the following steps:

Step one, to M (≤5000) individual, everyone gathers p (>=2) original eyebrow image, and be partitioned into pure eyebrow image after gray processing, and they are normalized to identical size, minimum is 10 × 10, is maximumly no more than 2000 × 2000;

Step 2, connects by row by every pure eyebrow image array and transposition becomes column vector;

Step 3, is designated as x respectively by all N number of column vectors ₁, x ₂..., x _n, compute vector average μ, formula is:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{x}_i$

Step 4, calculated population Scatter Matrix C, formula is:

$C=\frac{1}{N}\underset{k=1}{\overset{N}{\Σ}}(x_k-\mathrm{\μ}){(x_k-\mathrm{\μ})}^T$

Step 5, front m the eigenvalue λ of calculated population Scatter Matrix C ₁, λ ₂..., λ _mwith corresponding unit character vector u ₁, u ₂..., u _m;

Step 6, calculates x _iand u _jinner product namely

Step 7, structure x _ivector approximation

Step 8, the original eyebrow image of input test, is treated to gray scale image, is partitioned into the pure eyebrow image y of test, calculates y and u _jinner product namely ${\tilde{y}}_j=<y,u_j>,$ The vector approximation of structure y $\tilde{y}={({\tilde{y}}_1,{\tilde{y}}_2,...,{\tilde{y}}_m)}^T;$

Step 9, calculates optimum sparse coefficient vector s ^*=(s ₁, s ₂..., s _n) ^t, formula is as follows:

$s^{*}={\mathrm{argmin}}_s\left|\right|s\left|{|}_{1}subjectto\right||\underset{i=1}{\overset{N}{\&Sigma;}}{s}_i{\tilde{x}}_i-\tilde{y}|{|}_2\&le;\mathrm{\&epsiv;}$

In formula, s=(s _i1, s _i2..., s _iN) ^t, || || ₁for l ₁norm, represent vector each element absolute value and, || || ₂for l ₂norm;

Step 10, calculates the difference of test eyebrow and image i=1,2 ..., M, formula is as follows:

$r_i\left(\tilde{y}\right)=\left|\right|\tilde{y}-\underset{j=1}{\overset{N}{\&Sigma;}}{\mathrm{\&delta;}}_i(s^{*},{\tilde{x}}_j)|{|}_2$

In formula,

Step 11, determine to test the classification belonging to original eyebrow, formula is as follows:

$identity\left(y\right)=\underset{1\&le;i\&le;M}{\arg \min }{r}_i\left(\tilde{y}\right)$

In formula, the classification of identity (y) belonging to the original eyebrow of test, for that minimum r _ithe image category that (y ~) is corresponding.