CN101916371B - Method for illuminating/normalizing image and method for identifying image by using same - Google Patents

Abstract

The invention relates to a method for illuminating/normalizing an image and a method for identifying an image by using the same, wherein the method for illuminating/normalizing the image comprises the steps of: dividing the input image into a shadow zone and a normal illumination zone; obtaining the small scale part of the image when lambda L1 is valued within 0 to 0.5 according to a TVQI model or LTV model, preferably lambda L1=0.2; obtaining the small scale part of the image when the lambda L1 is valued within 0.6 to 1, preferably lambda L1=0.8; taking the small scale part of the image obtained when the lambda L1 is valued within 0 to 0.5 (preferably 0.2) as the small scale part of the shadow zone of the image, then taking the small scale part of the image obtained when lambda L1 is valued within 0.6 to 1 (preferably 0.8) as the small scale part of the normal illumination zone of the image, and splicing to obtain the small scale part v of the whole image. The invention can enhance the effectiveness of image identification under a complicated illumination condition without knowing the light source information previously.

Description

Image is carried out method that unitary of illumination handles and the image-recognizing method that adopts it

Technical field

The present invention relates to a kind of method that image is carried out the unitary of illumination processing, also relate to a kind of image-recognizing method that adopts it, belong to area of pattern recognition.

Background technology

In recent years, recognition of face research has obtained paying close attention to widely.Illumination, attitude and expression three big problems are the key factors that influences the recognition of face precision always, and wherein illumination factor, the particularly variation of physical environment light are not artificially can control, so photo-irradiation treatment is the step that each face identification system must carry out.Most of face identification systems are made certain limitation to illumination condition usually; Suppose that pending image obtains under basic evenly illumination condition; Illumination condition in they only allow among a small circle changes; Yet actual illumination condition is often inhomogeneous, and bright excessively, dark excessively, shade that polarisation, sidelight, high photoconduction cause all can make effect descend significantly.Therefore, how to reduce numerous researchists of concern illumination has obtained to(for) the influence of recognition of face.

TVQI (Total Variation based quotient image model) model is a kind of use TV-L1 yardstick decomposition model and the human face light normalization model that combines from quotient images (SQI) model.Propositions such as Wang be to use a kind of unitary of illumination model of the balanced outside photechic effect of the smoothness information of image itself from the quotient images model; Wherein the smoothness information of image is the result of weighting Gaussian filter filtering; Therefore but Gaussian filter can not keep the edge details of image HFS, from quotient images models treated lost part image feature information as a result.Proposed the TVQI model for addressing this problem Chen etc., the TVQI model uses TV-L1 to decompose the photechic effect in the large scale characteristics of image balance facial image that obtains.TV-L1 is the PDE of anisotropy diffusion, has the yardstick selectivity characteristic, can keep the edge feature in the image simultaneously.It is simple from quotient images that this method has not only kept, effective characteristics, and can keep more people's face detailed information, improved recognition effect.

Face TV-L1 model and TV-L2 model down and do a brief introduction.

The TV-L1 model:

The TV-L1 model be a kind of come by PDE method development image is carried out the image processing method of smoothing denoising.Its basic thought is to establish a functional J _λ[u]: Try to achieve and make J _λ[u] obtains the u of minimum value.Wherein u is the result of TV-L1 model; The image behind the smoothing denoising just; I is an input picture; Original image just, λ is for regulating the parameter of denoising degree, and

is gradient operator.

J _λFirst in [u] Being used for making image smoothing, second portion λ ∫ | I-u|dx makes u keep the principal character of I.Can regulate the similarity degree of u and I through λ.

From quotient images model (SQI) is a kind of unitary of illumination model, and its basic skills is to utilize low-pass filter to obtain the low-frequency component of facial image, with the computing of discussing of original image pointwise, obtains from quotient images again, is formulated as follows:

$Q(x,y)=\frac{I(x,y)}{S(x,y)}=\frac{I(x,y)}{F*I(x,y)}$

Wherein (x y) is every pixel value in quotient images to Q, and (x y) is every pixel value in the original image to I, and (x is every a pixel value in the low-frequency component of image y) to S, and F is a low-pass filter.Having certain illumination unchangeability from quotient images Q, is the result from the quotient images model.

The TV-L2 model:

The TV-L2 model is identical with the basic thought of TV-L1 model, and the constraint condition that only is to use is different, the TV-L2 model uses be the L2 norm as constraint condition, the TV-L1 model uses be the L1 norm as constraint condition, the difference of constraint condition makes functional J _λSecond portion in [u] is different with TV-L1, and the functional in the TV-L2 model is expressed as: Wherein u is the result of TV-L2 model, the image behind the smoothing denoising just, and I is an input picture, original image just, λ be for regulating the parameter of denoising degree,

Be gradient operator.

Summary of the invention

Technology of the present invention is dealt with problems and is: overcome the influence of illumination factor in the input picture; Provide a kind of image is carried out the method that unitary of illumination is handled; A kind of image-recognizing method that adopts it also is provided, and this method can strengthen the validity of image recognition under the complex illumination condition.

Technical solution of the present invention is:

The method that image is carried out the unitary of illumination processing provided by the invention may further comprise the steps:

Input picture is divided into shadow region and illumination normal region;

According to TVQI model or LTV model, obtain λ _L1The small scale part of image when getting the arbitrary value of 0-0.5 scope; Preferably, λ _L1=0.2;

According to TVQI model or LTV model, obtain λ _L1The small scale part of image when getting the arbitrary value of 0.6-1 scope; Preferably, v _λ=0.8

With λ _L1The image small scale part that (preferred 0.2) obtains when getting the arbitrary value of 0-0.5 scope is as the small scale part of image shadow region, with λ _L1The image small scale part that (preferred 0.8) obtains during for arbitrary value of 0.6-1 scope is as the small scale part in image normal illumination zone, and splicing obtains the small scale part v of entire image.

The method that image is carried out the unitary of illumination processing provided by the invention can further include following steps:

With input picture substitution TV-L2 model, obtain the large scale part u of image _L2

Image large scale part

to obtaining is carried out histogram equalization; Blocked histogram equalization preferably;

The image large scale part

that obtains after the small scale part v of said entire image and the equalization is merged.

The method that image is carried out the unitary of illumination processing provided by the invention can further include following steps:

Image large scale part

to obtaining after the equalization is carried out homomorphic filtering;

The image large scale part

that obtains after the small scale part v of said entire image and the homomorphic filtering is merged.

Provided by the invention image is carried out the method that unitary of illumination is handled, can further include following steps: it is 100 * 100 gray level image I that input picture is normalized into pixel.

When said input picture is gray level image, image pixel is revised as 100 * 100; When said input picture is coloured image, it is become gray level image through following formula:

$\mathrm{gray}=\frac{r+g+b}{3}$

In the formula, r, g, b represent in the coloured image red respectively, and be green, the value of blue component, gray represents the gray-scale value of gray level image.

The method that input picture is divided into shadow region and illumination normal region is:

A) the average gray value aver of calculating input image:

$\mathrm{aver}=\left(\underset{i}{\overset{n}{\mathrm{\Σ}}}\underset{j}{\overset{m}{\mathrm{\Σ}}}I(i,j)\right)/(m*n)$

In the formula, m, the length of n difference presentation video and wide;

B) calculate each pixel flag information flag (x, y):

$\mathrm{flag}(x,y)=\left\{\begin{array}{ccc}1& \mathrm{if}& I(x,y)>\mathrm{aver}\\ 0& \mathrm{if}& I(x,y)\≤\mathrm{aver}\end{array}\right.$

Flag (x, y) equaling 1 is that this point of expression is the illumination normal region, (x, y) equal 0 is that this point of expression is the shadow region to flag.

Image-recognizing method provided by the invention is after image being carried out the unitary of illumination processing, uses the subspace analysis method that the high dimensional data that obtains is projected in the lower dimensional space; Carry out feature extraction; Obtain the less vector of dimension, utilize sorter to classify then, obtain recognition result.The preferred arest neighbors method of said sorter.

The present invention compared with prior art has following advantage:

The method that image is carried out the unitary of illumination processing provided by the invention does not need to know in advance light source information, does not need training set yet, can overcome the influence of illumination factor in the input picture, strengthens the validity of image recognition under the complex illumination condition greatly.

The present invention is specially adapted to facial image identification, and when carrying out recognition of face, the photechic effect in can balanced facial image effectively improves the robustness of recognition of face for illumination effect.The method is very simple and can be applicable to any individual facial image.

Description of drawings

Below will combine the accompanying drawing specific embodiments of the invention to describe.

Fig. 1 is for carrying out the method flow diagram that unitary of illumination is handled according to the present invention to facial image.

Embodiment

In embodiment, be that facial image is that example is elaborated to the present invention with the image of importing.Certainly, the present invention also is applicable to other image.

Embodiment one:

The unitary of illumination of present embodiment is on the TVQI model based, to improve.

As shown in Figure 1, image is carried out the method that unitary of illumination is handled, may further comprise the steps:

(1) input picture is normalized to 100 * 100 gray level image I

If input picture is a gray level image, then image pixel is revised as 100 * 100;

If input picture is a coloured image, then coloured image is become gray level image through following formula:

$\mathrm{gray}=\frac{r+g+b}{3}$

In the formula, r, g, b represent in the coloured image red respectively, and be green, the value of blue component, gray represents the gray-scale value of gray level image.

(2) input picture is divided into shadow region and illumination normal region

The method that input picture I after step (1) normalization is divided into shadow region and illumination normal region is:

A) the average gray value aver of calculating input image:

$\mathrm{aver}=\left(\underset{i}{\overset{n}{\mathrm{\Σ}}}\underset{j}{\overset{m}{\mathrm{\Σ}}}I(i,j)\right)/(m*n)$

In the formula, m, the length of n difference presentation video and wide;

B) calculate each pixel flag information flag (x, y):

$\mathrm{flag}(x,y)=\left\{\begin{array}{ccc}1& \mathrm{if}& I(x,y)>\mathrm{aver}\\ 0& \mathrm{if}& I(x,y)\≤\mathrm{aver}\end{array}\right.$

Flag (x, y) equaling 1 is that this point of expression is the illumination normal region, (x, y) equal 0 is that this point of expression is the shadow region to flag.

(3) obtain λ _L1The small scale part of=0.2 o'clock facial image

Input picture after utilizing the TVQI model to step (1) normalization is handled, and obtains λ _L1The small scale part v of=0.2 o'clock facial image _λ=0.2

The TVQI model is from " Total variation models for variable lighting facerecognition " Chen, Terrence; Yin, Wotao; Zhou, Xiang Sean; Comaniciu, Dorin; Huang, Thomas S.Source:IEEE Transactions on Pattern Analysisand Machine Intelligence, v 28, and n 9, p 1519-1524, September 2006.

Concrete steps are:

A), use λ according to the TVQI model _L1=0.2 o'clock TV-L1 model obtains the facial image u behind the smoothing denoising _λ=0.2, formula is following:

$u_{\mathrm{\λ}=0.2}=\arg {\min }_u{\∫}_{\mathrm{\Ω}}|\▿u|\mathrm{dx}+{\mathrm{\λ}}_{L1}\∫\sqrt{{(I-u)}^2+\mathrm{\ϵ}}\mathrm{dx}$

Wherein u is the result of TV-L1 model; The image behind the smoothing denoising just; I is an input picture; Original image just, λ is for regulating the parameter of smoothing denoising degree, and is gradient operator.

Use the Euler-Lagrange equation Numerical Implementation can obtain u _λ=0.2Steady state solution

B) utilize According to from the quotient images model, with what obtain

As the low-frequency component of image, carry out merchant's computing of pointwise with original image I, obtain λ _L1=0.2 o'clock the small scale part v that comprises people's face detailed information _λ=0.2:

$v_{\mathrm{\λ}=0.2}(x,y)=I(x,y)/{\hat{u}}_{\mathrm{\λ}=0.2}(x,y)$

Wherein (x y) is every pixel value in the original image, v to I _λ=0.2(x y) is every pixel value in the TVQI model result,

Be the steady state solution that obtains In every pixel value.

(4) obtain λ _L1The small scale part of=0.8 o'clock facial image

Input picture after utilizing the TVQI model to step (1) normalization is handled, and obtains λ _L1The small scale part v of=0.8 o'clock facial image _λ=0.8Concrete steps are:

A), use λ according to the TVQI model _L1=0.8 o'clock TV-L1 model obtains the facial image u behind the smoothing denoising _λ=0.8, formula is following:

$u_{\mathrm{\λ}=0.2}=\arg {\min }_u{\∫}_{\mathrm{\Ω}}|\▿u|\mathrm{dx}+{\mathrm{\λ}}_{L1}\∫\sqrt{{(I-u)}^2+\mathrm{\ϵ}}\mathrm{dx}$

Wherein u is the result of TV-L1 model; The image behind the smoothing denoising just; I is an input picture; Original image just, λ is for regulating the parameter of smoothing denoising degree, and is gradient operator.

Use the Euler-Lagrange equation Numerical Implementation to obtain u _λ=0.8Steady state solution

B) utilize According to from the quotient images model, with what obtain

As the low-frequency component of image, carry out merchant's computing of pointwise with original image I, obtain λ _L1=0.8 o'clock the small scale part v that comprises people's face detailed information _λ=0.8:

$v_{\mathrm{\λ}=0.2}(x,y)=I(x,y)/{\hat{u}}_{\mathrm{\λ}=0.2}(x,y)$

Wherein (x y) is every pixel value in the original image, v to I _λ=0.8(x y) is every pixel value in the TVQI model result, Be the steady state solution that obtains

In every pixel value.

(5) multiple dimensioned splicing obtains the small scale part of entire image

The λ that step (3) is obtained _L1The small scale part v of=0.2 o'clock facial image _λ=0.2As the small scale part of the shadow region in the facial image, the λ that step (4) is obtained _L1The small scale part v of=0.8 o'clock facial image _λ=0.8Small scale part as the zone of the normal illumination in the facial image:

$v(x,y)=\left\{\begin{array}{ccc}{v}_{\mathrm{\λ}=0.8}(x,y)& \mathrm{if}& \mathrm{flag}(x,y)==1\\ v_{\mathrm{\λ}=0.2}(x,y)& \mathrm{if}& \mathrm{flag}(x,y)==0\end{array}\right.$

Obtain the small scale part v of entire image.

(6) utilize the TV-L2 model to obtain the large scale part u of facial image _L2

Input picture substitution TV-L2 model after the normalization that step (1) is obtained, carry out smoothing denoising:

$u_{L2}=\arg {\min }_u\∫|\▿u|\mathrm{dx}+{\mathrm{\λ}}_{L2}\∫{(I-u)}^2+\mathrm{\ϵdx}$

Wherein u is the image behind the smoothing denoising, and I is an original image, λ _L2Be parameter, in order to guarantee (I-u) ²Can little adding distracter ε.

Use the Euler-Lagrange equation Numerical Implementation to obtain u _L2Steady state solution

(7) the large scale image is carried out the blocked histogram equalization

The basic thought of histogram equalization is to be the histogram transformation of original graph equally distributed form, has so just increased the dynamic range of pixel gray-scale value, thereby can reach the effect that strengthens the integral image contrast.

(the histogram equalization of piecemeal is not good to carry out the blocked histogram equalization for the facial image large scale part

that step (6) is obtained; Be blocked histogram equalization better effects if); Strengthen local contrast, balanced photechic effect.Concrete steps are:

A, to image block

${\hat{u}}_{L2}^1(x,y)={\hat{u}}_{L2}(x,y)$ 1≤x≤50，1≤y≤50

${\hat{u}}_{L2}^2(x-50,y)={\hat{u}}_{L2}(x,y)$ 51≤x≤100，1≤y≤50

${\hat{u}}_{L2}^3(x,y-50)={\hat{u}}_{L2}(x,y)$ 1≤x≤50，51≤y≤100

${\hat{u}}_{L2}^4(x,y)={\hat{u}}_{L2}(x-50,y-50)$ 51≤x≤100，51≤y≤100

B, each piece is carried out histogram equalization

Utilize the cumulative distribution functional based method to carry out histogram equalization, method is following:

A, calculating gray scale Density Distribution GP

GP(i)＝σ _i/(m*n) i＝1…256

σ wherein _iBe that gray-scale value is the number of the pixel of i in the image;

B, calculating accumulative histogram distribution S

$S\left(i\right)=\underset{k=1}{\overset{i}{\mathrm{\Σ}}}\mathrm{GP}\left(k\right)$ i＝1…256

C, cumulative distribution round, and set up the mapping relations R () of gray level

R(i)＝floor(256*S(i)+0.5) i＝1…256

D, according to the mapping relations of setting up; Original image is carried out grey scale mapping, obtain the histogram equalization result of each piece image

${\hat{u}}_{L2}^{j*}=R\left({\hat{u}}_{L2}^j\right)$ j＝1，2，3，4

C, be spliced into entire image, obtain the result

of the blocked histogram equalization of entire image

${\hat{u}}_{L2}^{*}(x,y)={\hat{u}}_{L2}^{1*}(x,y)$ 1≤x≤50，1≤y≤50

${\hat{u}}_{L2}^{*}(x-50,y)={\hat{u}}_{L2}^{2*}(x,y)$ 51≤x≤100，1≤y≤50

${\hat{u}}_{L2}^{*}(x,y-50)={\hat{u}}_{L2}^{3*}(x,y)$ 1≤x≤50，51≤y≤100

${\hat{u}}_{L2}^{*}(x,y)={\hat{u}}_{L2}^{4*}(x-50,y-50)$ 51≤x≤100，51≤y≤100

(8) homomorphic filtering

that step (7) is obtained carries out homomorphic filtering, and concrete steps are:

A) image is carried out logarithm operation

$z(x,y)=\ln {\hat{u}}_{L2}^{*}(x,y)$

B) image is carried out the Fourier conversion, image is transformed into frequency domain from the spatial domain

where

is the Fourier transform

C) (u v), carries out filtering to image to utilize Gauss's Hi-pass filter H

S(u，v)＝H(u，v)Z(u，v)

D) image is carried out contrary Fourier conversion, image is transformed into the spatial domain from frequency domain

where

is the inverse Fourier transform

E) image is carried out exponent arithmetic, obtain image is carried out the result

after the homomorphic filtering

${\hat{u}}_{L2}^{**}(x,y)=e^{g(x,y)}$

(9) multiple dimensioned fusion

The large scale part

that small scale part v that step (5) is obtained and step (8) obtain is carried out addition and is merged:

$y(x,y)=\mathrm{\α}{u}_{L2}^{**}(x,y)+\mathrm{\βv}(x,y)$

Wherein two kinds of different scales are used different weighting coefficients: α=0.6, β=0.4.Y is the facial image behind the unitary of illumination.

Need to prove; Above-mentioned steps (2), (3), (4) and (5) are that facial image is carried out the steps necessary that unitary of illumination is handled, and step (1), (6), (7), (8) and (9) are the steps that better and further comprises for the unitary of illumination treatment effect that makes image.

According to image-recognizing method of the present invention; Be after image being carried out the unitary of illumination processing; Use the subspace analysis method that the high dimensional data that obtains is projected in the lower dimensional space, carry out feature extraction, obtain the less vector of dimension; Utilize sorter (preferred arest neighbors method) to classify then, obtain recognition result.

Image after unitary of illumination is handled can not be applied to recognition of face because dimension is higher, so need to use the subspace analysis method to carry out feature extraction, high dimensional data is projected to than in the lower dimensional space.The method that high dimensional data is become low dimension data can be selected following several kinds for use:

Principal component analysis (PCA) (PCA:Principal Component Analysis), this method is extracted the major component characteristic;

Independent component analysis (ICA:Independent Component Correlation Algorithm), this method is extracted the independent component characteristic;

Linear discriminant analysis (LDA:linearity distinction analysis), this method extract the optimum diagnostic characteristics of Fisher.

The y that v that step (5) is obtained through the subspace analysis method or step (9) obtain projects in the lower dimensional space, obtains the less vector of dimension, utilizes the arest neighbors method to classify then, obtains recognition result.

The arest neighbors method is the simplest a kind of sorter, except this method, also has some comparatively complicated classifier methods to combine with the subspace analysis method, for example BAYESIAN NETWORK CLASSIFIER, cascade classifier, BP neural network classifier etc.

Embodiment two:

The unitary of illumination of present embodiment is on the LTV model based, to improve.

The LTV model is from " Total variation models for variable lighting facerecognition " Chen, Terrence; Yin, Wotao; Zhou, Xiang Sean; Comaniciu, Dorin; Huang, Thomas S.Source:IEEE Transactions on Pattern Analysisand Machine Intelligence, v 28, and n 9, p 1519-1524, September 2006.

LTV model and TVQI model come from same piece of writing article, and the difference of the two is: what TVQI used is merchant's computing, and what LTV used is logarithm operation.Use LTV model and the difference of using the TVQI model only to be step (3) and step (4).

Only just describe below with the difference of embodiment one.

Step 3: utilize the input picture after the normalization that the LTV model obtains step 1 to handle, obtain λ _L1The small scale part v of=0.2 o'clock facial image _λ=0.2Concrete steps are:

A, input picture is carried out logarithm operation

f(x，y)＝lnI(x，y)

B, use λ _L1=0.2 o'clock TV-L1 model obtains the u behind the f smoothing denoising _λ=0.2, formula is following:

$u_{\mathrm{\λ}=0.2}=\arg {\min }_u{\∫}_{\mathrm{\Ω}}|\▿u|\mathrm{dx}+{\mathrm{\λ}}_{L1}\∫\sqrt{{(f-u)}^2+\mathrm{\ϵ}}\mathrm{dx}$

Wherein u is the result of TV-L1 model, and I is an input picture, λ _L1Be the parameter of adjusting smoothing denoising degree,

Be gradient operator.

Use the Euler-Lagrange equation Numerical Implementation can obtain u _λ=0.2Steady state solution

C, obtain the essential part ρ of image

$\mathrm{\ρ}(x,y)=f(x,y)-{\hat{u}}_{\mathrm{\λ}=0.2}(x,y)$

D, the essential part ρ of image is carried out exponent arithmetic obtain λ _L1=0.2 o'clock the small scale part v that comprises people's face detailed information _λ=0.2

v _λ＝0.2(x，y)＝e ^ρ(x，y)

Step 4: utilize the LTV model that the normalized input picture of step (1) is handled, obtain λ _L1The small scale part v of=0.8 o'clock facial image _λ=0.8Concrete steps are:

A, input picture is carried out logarithm operation

f(x，y)＝lnI(x，y)

B, use λ _L1=0.8 o'clock TV-L1 model obtains the u behind the f smoothing denoising _λ=0.8, formula is following:

$u_{\mathrm{\λ}=0.2}=\arg {\min }_u{\∫}_{\mathrm{\Ω}}|\▿u|\mathrm{dx}+{\mathrm{\λ}}_{L1}\∫\sqrt{{(f-u)}^2+\mathrm{\ϵ}}\mathrm{dx}$

Wherein u is the result of TV-L1 model, and I is an input picture, λ _L1Be the parameter of adjusting smoothing denoising degree,

Be gradient operator.

Use the Euler-Lagrange equation Numerical Implementation can obtain u _λ=0.8Steady state solution

C, obtain the essential part ρ of image

$\mathrm{\ρ}(x,y)=f(x,y)-{\hat{u}}_{\mathrm{\λ}=0.8}(x,y)$

D, the essential part ρ of image is carried out exponent arithmetic obtain λ _L1=0.8 o'clock the small scale part v that comprises people's face detailed information _λ=0.8

v _λ＝0.8(x，y)＝e ^ρ(x，y)。

The content of not doing in the instructions of the present invention to describe in detail belongs to this area professional and technical personnel's known technology.

The present invention is not limited to the content that claim and the foregoing description are addressed, so long as any invention of creating out according to design of the present invention all should belong within protection scope of the present invention.

Claims (10)

1. image is carried out the method that unitary of illumination is handled, it is characterized in that, may further comprise the steps:

Input picture is divided into shadow region and illumination normal region;

According to TVQI model or LTV model, obtain λ _L1The small scale part of image when getting the arbitrary value of 0-0.5 scope;

According to TVQI model or LTV model, obtain λ _L1The small scale part of image when getting the arbitrary value of 0.6-1 scope;

With λ _L1The image small scale part that obtains when getting the arbitrary value of 0-0.5 scope is as the small scale part of image shadow region, with λ _L1The image small scale part that obtains during for arbitrary value of 0.6-1 scope is as the small scale part in image normal illumination zone, and splicing obtains the small scale part v of entire image.

2. the method that image is carried out the unitary of illumination processing according to claim 1 is characterized in that, further may further comprise the steps:

With input picture substitution TV-L2 model, obtain the large scale part u of image _L2

To the image large scale part u that obtains _L2The steady state solution of using the Euler-Lagrange equation Numerical Implementation to obtain Carry out histogram equalization;

The image large scale part

that obtains after the small scale part v of said entire image and the equalization is merged.

3. the method that image is carried out the unitary of illumination processing according to claim 2 is characterized in that, further may further comprise the steps:

Image large scale part

to obtaining after the equalization is carried out homomorphic filtering;

The image large scale part

that obtains after the small scale part v of said entire image and the homomorphic filtering is merged.

4. according to claim 1 image is carried out the method that unitary of illumination is handled, it is characterized in that, further may further comprise the steps: it is 100 * 100 gray level image I that input picture is normalized into pixel.

5. the method that image is carried out the unitary of illumination processing according to claim 4 is characterized in that, when said input picture is gray level image, image pixel is revised as 100 * 100; When said input picture is coloured image, it is become gray level image through following formula:

$\mathrm{gray}=\frac{r+g+b}{3}$

In the formula, r, g, b represent in the coloured image red respectively, and be green, the value of blue component, gray represents the gray-scale value of gray level image.

6. the method that image is carried out the unitary of illumination processing according to claim 1 is characterized in that the method that input picture is divided into shadow region and illumination normal region is:

A) the average gray value aver of calculating input image:

$\mathrm{aver}=\left(\underset{i}{\overset{n}{\mathrm{\Σ}}}\underset{j}{\overset{m}{\mathrm{\Σ}}}I(i,j)\right)/(m*n)$

In the formula, m, the length of n difference presentation video and wide;

B) calculate each pixel flag information flag (x, y):

$\mathrm{flag}(x,y)=\left\{\begin{array}{ccc}1& \mathrm{if}& I(x,y)>\mathrm{aver}\\ 0& \mathrm{if}& I(x,y)\≤\mathrm{aver}\end{array}\right.$

Flag (x, y) equaling 1 is that this point of expression is the illumination normal region, (x, y) equal 0 is that this point of expression is the shadow region to flag.

7. the method that image is carried out the unitary of illumination processing according to claim 1 is characterized in that,

According to TVQI model or LTV model, obtain λ _L1The small scale part v of=0.2 o'clock image _λ=0.2

According to TVQI model or LTV model, obtain λ _L1The small scale part v of=0.8 o'clock image _λ=0.8

With λ _L1The small scale part v of=0.2 o'clock image _λ=0.2As the small scale part of image shadow region, with λ _L1The small scale part v of=0.8 o'clock image _λ=0.8As the small scale part in image normal illumination zone, splicing obtains the small scale part v of entire image.

8. the method that image is carried out the unitary of illumination processing according to claim 2 is characterized in that, to the image large scale part u that obtains _L2The steady state solution of using the Euler-Lagrange equation Numerical Implementation to obtain

Carry out the blocked histogram equalization.

9. image-recognizing method; It is characterized in that,, use the subspace analysis method that high dimensional data is projected in the lower dimensional space according to the unitary of illumination result that above-mentioned arbitrary claim obtains; Carry out feature extraction; Obtain the less vector of dimension, utilize sorter to classify then, obtain recognition result.

10. image-recognizing method according to claim 9 is characterized in that, said image is a facial image.

Images