CN101162502A - Method for removing glasses during human recognition - Google Patents

Abstract

The invention relates to a method for eliminating glasses in human face recognition and is characterized in that the method includes the following steps: (1) by means of the pivot element analysis method and the training feature space of a plurality of images of human faces without glasses, the feature space and an average human face image are stored; (2) a human face image is recreated and input by PCA; (3) a compensation image is recreated by PCA to obtain the recreated image of next iteration and the iteration terminates when the difference of two adjacent iterations is less than the given threshold; (4) the upper part of a final compensation image is combined with the lower part of an input human face image to form a human face image without glasses. Due to adopting the technical proposal, the invention has the advantages that glasses mark or reflection and image tone caused by glasses are eliminated, thereby greatly increasing the recognition rate of human face image with glasses.

Description

The method that glasses in a kind of recognition of face are eliminated

Technical field

The present invention relates to the correlative technology field of recognition of face, the method that the glasses in particularly a kind of recognition of face are eliminated.

Background technology

Glasses are modal shelters in the facial image, and it is very big to the influence of discrimination.In order to improve the discrimination of the facial image of wearing glasses, the vestige of glasses or reflection and the tone that causes by glasses have been proposed to eliminate.

Summary of the invention

The objective of the invention is to have solved in the recognition of face owing to glasses block the problem that people's face influences discrimination for the method for the glasses elimination in a kind of recognition of face is provided.

For achieving the above object, the present invention has adopted following technical scheme:

The invention discloses the method that the glasses in a kind of recognition of face are eliminated, it is characterized in that:

May further comprise the steps:

1) with pca method by several facial image training characteristics spaces of not wearing glasses, and preserve feature space and average facial image;

2) rebuild the input facial image with pca method, algorithm as shown in the formula:

Wherein, Γ ^RBe the facial image of rebuilding,  is average man's face, μ _k(k=1,2 ... be to begin to train the eigenface that obtains, ω D) _kIt is the weight of corresponding k eigenface;

3) rebuild the reconstructed image that compensating images obtains next iteration with pca method, as the difference of adjacent twice iteration ε during less than given thresholding, iteration stops;

The error compensation of iteration:

Γ ^C＝ω·+(1-ω)·Γ?if(t＝1)

${\mathrm{\Γ}}_t^C=\mathrm{\ω}\·{\mathrm{\Γ}}_t^R+(1-\mathrm{\ω})\·\mathrm{\Γ}$ ?if(t＞1)

Wherein:

d(i)＝|Γ(i)-Γ ^R(i)|

D(i)＝(Γ ^R(i)·d(i)) ^1/2

ω(i)＝1 if(D(i)≥T _H)

ω(i)＝1 if(T _L≤D(i)＜T _H)

ω(i)＝0 if(D(i)≥T _L)

Γ is original input picture, and t is an iterations, and stopping criterion for iteration is as follows:

$\left|\right|{\mathrm{\Γ}}_t^C-{\mathrm{\Γ}}_{t-1}^C\left|\right|\≤\mathrm{\ω}$

4) the synthetic glasses-free facial image of the latter half of the first half of final compensating images and input facial image.

Wherein, in the described step 1), covariance matrix and average facial image thereof with the facial image of pca method after according to several normalization, obtain the eigenwert and the proper vector of this covariance, its characteristic of correspondence vector of big or small descending sort according to eigenwert promptly obtains described feature space.In the described step 3), ε is 0.2 during the given thresholding of the difference of adjacent twice iteration.

Because adopted above scheme, the beneficial effect that the present invention is possessed is: can eliminate the vestige of glasses or by reflection and tone that glasses cause, improve the discrimination of the facial image of wearing glasses greatly.

Embodiment

Below in conjunction with drawings and Examples the present invention is further described.

The method that glasses in a kind of recognition of face are eliminated is characterized in that: may further comprise the steps:

1) utilizes pivot analysis software PCA (principle componet analysis) to use pca method, and preserve feature space and average facial image by several facial image training characteristics spaces of not wearing glasses; Covariance matrix and the average facial image thereof by the facial image of PCA after wherein according to several normalization, obtain the eigenwert and the proper vector of this covariance, its characteristic of correspondence vector of big or small descending sort according to eigenwert promptly obtains described feature space;

2) rebuild the input facial image with pca method, algorithm as shown in the formula:

Wherein, Γ ^RBe the facial image of rebuilding,  is average man's face, μ _k(k=1,2 ... be to begin to train the eigenface that obtains, ω D) _kIt is the weight of corresponding k eigenface;

3) rebuild the reconstructed image that compensating images obtains next iteration with pca method, as the difference of adjacent twice iteration ε during less than given thresholding, iteration stops, and wherein ε is 0.2;

The error compensation of iteration:

Γ ^C＝ω·+(1-ω)·Γ if(t＝1)

${\mathrm{\Γ}}_t^C=\mathrm{\ω}\·{\mathrm{\Γ}}_t^R+(1-\mathrm{\ω})\·\mathrm{\Γ}$ ?if(t＞1)

Wherein:

d(i)＝|Γ(i)-Γ ^R(i)|

D(i)＝(Γa ^R(i)·d(i)) ^1/2

ω(i)＝1 if(D(i)≥T _H)

ω(i)＝1 if(T _L≤D(i)＜T _H)

ω(i)＝0 if(D(i)≥T _L)

Γ is original input picture, and t is an iterations, and stopping criterion for iteration is as follows:

$\left|\right|{\mathrm{\Γ}}_t^C-{\mathrm{\Γ}}_{t-1}^C\left|\right|\≤\mathrm{\ϵ}$

4) the synthetic glasses-free facial image of the latter half of the first half of final compensating images and input facial image.

Claims (3)

1. the method eliminated of the glasses in the recognition of face is characterized in that: may further comprise the steps:

1) with pca method by several facial image training characteristics spaces of not wearing glasses, and preserve feature space and average facial image;

2) rebuild the input facial image with pca method, algorithm as shown in the formula:

Wherein, Γ ^RBe the facial image of rebuilding,  is average facial image, μ _k(k=1,2 ... be to begin to train the eigenface that obtains, ω D) _kIt is the weight of corresponding k eigenface;

3) rebuild the reconstructed image that compensating images obtains next iteration with pca method, as the difference of adjacent twice iteration ε during less than given thresholding, iteration stops;

The error compensation of iteration:

Γ ^C＝ω·+(1-ω)·Γ?if(t＝1)

${\mathrm{\Γ}}_t^C=\mathrm{\ω}\·{\mathrm{\Γ}}_t^R+(1-\mathrm{\ω})\·\mathrm{\Γ}$ if(t＞1)

Wherein:

d(i)＝|Γ(i)-Γ ^R(i)|

D(i)＝(Γ ^R(i)·d(i)) ^1/2

ω(i)＝1 if(D(i)≥T _H)

ω(i)＝1 if(T _L≤D(i)＜T _H)

ω(i)＝0 if(D(i)≥T _L)

Γ is original input picture, and t is an iterations, and stopping criterion for iteration is as follows:

$\left|\right|{\mathrm{\Γ}}_t^C-{\mathrm{\Γ}}_{t-1}^C\left|\right|\≤\mathrm{\ϵ}$

4) the synthetic glasses-free facial image of the latter half of the first half of final compensating images and input facial image.

2. the method that the glasses in a kind of recognition of face according to claim 1 are eliminated, it is characterized in that: in the described step 1), covariance matrix and average facial image thereof with the facial image of pca method after according to several normalization, obtain the eigenwert and the proper vector of this covariance, its characteristic of correspondence vector of big or small descending sort according to eigenwert promptly obtains described feature space.

3. the method that glasses in a kind of recognition of face according to claim 1 are eliminated is characterized in that: in the described step 3), ε is 0.2 during the given thresholding of the difference of adjacent twice iteration.