CN103870820A - Illumination normalization method for extreme illumination face recognition - Google Patents

Abstract

The invention provides an illumination normalization method for extreme illumination face recognition. The method comprises the steps that through a retina model imitating primates, an illumination normalization model based on the retina is built through an improved Naka-Rushton equation; when adaption factors in the Naka-Rushton equation are calculated, the adaption factors are estimated under the condition of considering illumination differences, the influence of light conditions is considered, low-frequency information and high-frequency information in a face image are mainly processed through different algorithms, and therefore the overall robustness of a system is improved. Through experimental verification, the method is rarely influenced by training set selection, and is quite good in stability.

Description

Extreme path is according to the unitary of illumination method of recognition of face

Technical field

The present invention relates to the face recognition technology under foundation and the extreme light conditions of unitary of illumination processing, Modified Retinal Model.

Background technology

Many years in the past, the research of recognition of face made remarkable progress.But under uncontrolled physical environment, the research of recognition of face still faces many challenges.To a certain extent, for an effective face identification system, illumination variation is an open question all the time, and it has seriously reduced the performance of face identification system.

The problem of illumination variation on recognition of face impact, has in recent years caused researchist's attention, and has proposed several different methods.These methods roughly can be classified as 3 classes: illumination three-dimensional model, illumination invariant extract and unitary of illumination method.

Illumination three-dimensional model refers to that by three-dimensional model, facial image being carried out to matching eliminates illumination effect.In order to obtain effective three-dimensional model, these class methods not only need sufficient training sample, and will suppose that facial image is protruding structure, and the complexity of algorithm is higher, and these conditions are often difficult to realize in actual applications.

Illumination invariant extracts the feature that refers to that extraction and illumination condition are irrelevant, eliminates the impact of illumination variation.For addressing this problem, many methods have been proposed, comprise multiple dimensioned retinex (MSR), logarithm total variation (LTV) and gradient face etc.But these methods can not fully be eliminated different lighting effects, halation phenomenon still exists.

Unitary of illumination refers to and utilizes basic image processing techniques to carry out pre-service to light image, obtains the image of illumination robust.The method mainly comprises: log-transformation, histogram equalization, Gamma correction, adaptive histogram equalization and partial histogram equalization etc.Although these class methods can be eliminated the impact of illumination variation on recognition of face to a certain extent, in complex illumination situation, discrimination is difficult to satisfactory.Recently, be subject to human visual system's inspiration, many methods are in the middle of image processing and feature extraction.The people such as Meylan have proposed a kind of improved Naka-Rushton equation, and on this equational basis, use two kinds of continuous nonlinear operations to carry out simulating human vision system, thereby avoid halation, improve overall outward appearance.

Subsequently, by using the difference of above-mentioned two kinds of continuous nonlinear operations and Gaussian filter, a kind of unitary of illumination method based on Modified Retinal Model has been proposed.The method can realize good performance in the recognition of face under different illumination conditions.But, still need the impact of further elimination illumination variation on recognition of face at present badly.

Summary of the invention

In order to solve the effect unsafty problem of said method in extreme light conditions human face identification, the present invention seeks to from human retina model, more effective unitary of illumination method is proposed, to improve the identification accuracy of extreme light conditions human face identification.

The technical solution used in the present invention is in order to achieve the above object:

Extreme path shines a unitary of illumination method for recognition of face,

Analyze the functional layer of primate Modified Retinal Model, consult existing Modified Retinal Model, for the foundation based on amphiblestroid unitary of illumination model provides reference;

Utilize Naka-Rushton equation, for the foundation of unitary of illumination model provides concrete principle foundation;

In the time of the Adaptation factor calculating in Naka-Rushton equation, Adaptation factor is estimated in the situation that considering illumination difference;

Facial image is divided into low frequency and HFS by unitary of illumination processing section, and then take respectively different algorithms to process.

Preferably, set up unitary of illumination model: the Modified Retinal Model that utilizes primate, functional layer comprises photoreceptor layer, outside lamina reticularis OPL and inner lamina reticularis IPL, set up the unitary of illumination model based on Modified Retinal Model, comprise two continuous local image irradiation compression and space bandpass filterings.

Preferably, Naka-Rushton equation: calculate each pixel by carry out gaussian filtering on neighborhood, be defined as follows:

$X_0=X\left(p\right)*G_H+\stackrel{\‾}{X}/2---\left(1\right)$

Wherein, p is the pixel in image, X ₀be to be the Adaptation factor that p is in pixel, X (p) is the intensity of input picture, and * represents convolution algorithm, G _ha 2-d gaussian filters device, corresponding to the mean value of the image pixel intensities of X image.

Preferably, estimation Adaptation factor: I represents original image, and I represents original image, and A represents the low frequency sub-band of I, and DH, DV and DD represent respectively the high-frequency sub-band of I;

Illumination-classification based on Otsu carries out on low frequency sub-band A, supposes that t is the threshold value being obtained by Otsu, and the illumination condition of A can approximately be divided into two classes:

M ₁＝{A(x，y)|A(x，y)≥t} (2)

M ₂＝{A(x，y)|A(x，y)＜t) (3)

Wherein, A (x, y) represents the intensity of A at point (x, y);

The Adaptation factor A of A _{0 ρ}be defined as follows:

$A_{0\mathrm{\ρ}}(x,y)=A(x,y)*G_{\mathrm{\ρ}}+\stackrel{\‾}{{\mathrm{mean}}_A}/2---\left(4\right)$

$\stackrel{\‾}{{\mathrm{mean}}_A}=\left\{\begin{array}{cc}\stackrel{\‾}{M_A^1}& A(x,y)\∈M_1\\ \stackrel{\‾}{M_A^2}& A(x,y)\∈M_2\end{array}\right.---\left(5\right)$

Wherein, G _ρbe gaussian filtering, ρ is standard deviation,

with

respectively M ₁and M ₂mean value.

Preferably, for the unitary of illumination of recognition of face, set up visual model:

Single-stage discrete two-dimensional wavelet transformation is used for realizing facial image is extracted to low frequency and high-frequency information;

Process low-frequency information by imitating retinal information treatment mechanism, and large-scale high frequency coefficient is blocked by a threshold value;

Ask discrete two-dimensional to convert and get its inverse by the low frequency to applicable and high-frequency information and obtain unitary of illumination result.

Preferably, low-frequency information processing:

The first step, sets up photographic layer: the processing equation of low frequency sub-band A is as follows,

$A^p(x,y)=({\max }_A+A_{{0\mathrm{\ρ}}_1}(x,y))A(x,y)/(A(x,y)+A_{{0\mathrm{\ρ}}_1}(x,y))---\left(6\right)$

Wherein, max _athe maximal value of A,

the Adaptation factor of A at point (x, y), ρ ₁standard deviation and the A of gaussian filtering ^p(x, y) is corresponding output;

Second step, sets up outside lamina reticularis: A ^pprocessing as follows,

$A^o(x,y)=({\max }_{A^p}+A_{0_{{\mathrm{\ρ}}_2}}^p(x,y))A^p(x,y)/(A^p(x,y)+A_{0_{{\mathrm{\ρ}}_2}}^p(x,y))---\left(7\right)$

Wherein,

a ^pmaximal value,

a ^pat the Adaptation factor of point (x, y), ρ ₂standard deviation and the A of gaussian filtering ^o(x, y) exports accordingly.

The 3rd step, rim detection: adopt space bandpass filtering to carry out the information processing of simulated interior lamina reticularis, be defined as follows:

$A^e(x,y)=(\exp (-\frac{(x^2+y^2)}{{2\mathrm{\ρ}}_L^2})/2{\mathrm{\π\ρ}}_L^2-\exp (-\frac{(x^2+y^2)}{{2\mathrm{\ρ}}_H^2})/2{\mathrm{\π\ρ}}_H^2)*A^o---\left(8\right)$

Wherein, ρ _land ρ _hthe respective standard that is two Gaussian functions is poor, A ^ethe edge of presentation video.

Preferably, high-frequency information processing:

Make H represent DH, DV or DD, the truncation of high-frequency sub-band is as follows:

$H^{\&prime;}(x,y)=\left\{\begin{array}{cc}t& H(x,y)>t\\ H(x,y)& \left|H(x,y)\right|\&le;t\\ -t& H(x,y)<-t\end{array}\right.---\left(9\right)$

t＝median(|H(x，y)|) (10)

Wherein, median () is used for returning determinant | and the intermediate value of H|, by truncation, high-frequency sub-band is divided into DH ', DV ', DD '.

The invention has the beneficial effects as follows: a kind of extreme path of the present invention is according to the unitary of illumination method of recognition of face, by imitating the Modified Retinal Model of primate, utilize the Naka-Rushton equation after improving to set up based on amphiblestroid unitary of illumination model, in the time of the Adaptation factor calculating in Naka-Rushton equation, Adaptation factor is estimated in the situation that considering illumination difference, consider the impact of illumination condition, in embodiment, adopt different algorithms to process to the low frequency in facial image and high-frequency information emphatically, thereby improve entire system robustness; And through experimental verification, it is minimum that the method is affected by training set selection, has good stability.

Accompanying drawing explanation

Fig. 1 is that embodiment is to a comparative result comparison diagram that people uses distinct methods to obtain under different illumination conditions in Yale B face database.

Embodiment

Describe the preferred embodiments of the present invention in detail below in conjunction with accompanying drawing.

For improving the accuracy of extreme illumination condition human face identification, a kind of unitary of illumination method of new recognition of face is proposed: by imitating the Modified Retinal Model of primate, utilize the Naka-Rushton equation after improving to set up based on amphiblestroid unitary of illumination model, consider the impact of illumination condition, in embodiment, adopt different algorithms to process to the low frequency in facial image and high-frequency information emphatically, thereby improve entire system robustness.

The core technology of embodiment is estimation and the unitary of illumination processing section of Adaptation factor.First analyze the functional layer of primate Modified Retinal Model, consult existing Modified Retinal Model, for the foundation based on amphiblestroid unitary of illumination model provides reference; Utilize improved Naka-Rushton equation, for the foundation of unitary of illumination model provides concrete principle foundation; In the time of the Adaptation factor calculating in Naka-Rushton equation, to consider under complex illumination condition, the mean value of piece image pixel, as part Adaptation factor, is but applied to whole pixel domain, and this is inappropriate.For obtaining Adaptation factor accurately, Adaptation factor must be to estimate in the situation that considering illumination difference; Facial image is divided into low frequency and HFS by unitary of illumination processing section, and then take respectively different algorithms to process.

The detailed step of embodiment is as follows:

Unitary of illumination model: the Modified Retinal Model that utilizes primate, comprise 3 functional layers: photoreceptor layer, outside lamina reticularis OPL and inner lamina reticularis IPL, set up the unitary of illumination model based on Modified Retinal Model, comprise two continuous local image irradiation compression and space bandpass filterings.

Naka-Rushton equation: utilize a kind of improved Naka-Rushton equation, mainly calculate each pixel by carry out gaussian filtering on its neighborhood, be defined as follows:

$X_0=X\left(p\right)*G_H+\stackrel{\&OverBar;}{X}/2---\left(1\right)$

Wherein, p is the pixel in image, X ₀be to be the Adaptation factor that p is in pixel, X (p) is the intensity of input picture, and * represents convolution algorithm, G _ha 2-d gaussian filters device,

corresponding to the mean value of the image pixel intensities of X image.

Unitary of illumination based on improvement type can be avoided halation phenomenon, improves overall outward appearance.Still there is defect.Such as, use gaussian filtering to estimate that local intensity of illumination often can produce inaccurate result to the edge of image to original image.Particularly under complex illumination condition, the mean value of piece image

as part Adaptation factor, be but applied to whole pixel domain, this is inappropriate.In unitary of illumination is after a while processed, these produce error possibly.Therefore in embodiment, for obtaining Adaptation factor accurately, be, mainly to complete unitary of illumination processing in the low-frequency band of image by two-dimensional wavelet transformation.In addition Adaptation factor X, ₀to estimate in the situation that considering illumination difference.

Estimation Adaptation factor: I represents original image, and A represents the low frequency sub-band of I, and DH, DV and DD represent respectively the high-frequency sub-band of I.First, based on the illumination-classification of Otsu, adopt the maximum difference between various classification, on low frequency sub-band A, carry out.Suppose that t is the threshold value being obtained by Otsu, the illumination condition of A can approximately be divided into two classes:

M ₁＝{A(x，y)|A(x，y)≥t} (2)

M ₂＝{A(x，y)|A(x，y)＜t) (3)

Wherein, A (x, y) represents the intensity of A at point (x, y).

The Adaptation factor A of A _{0 ρ}be defined as follows:

$A_{0\mathrm{\&rho;}}(x,y)=A(x,y)*G_{\mathrm{\&rho;}}+\stackrel{\&OverBar;}{{\mathrm{mean}}_A}/2---\left(4\right)$

$\stackrel{\&OverBar;}{{\mathrm{mean}}_A}=\left\{\begin{array}{cc}\stackrel{\&OverBar;}{M_A^1}& A(x,y)\&Element;M_1\\ \stackrel{\&OverBar;}{M_A^2}& A(x,y)\&Element;M_2\end{array}\right.---\left(5\right)$

Wherein, G _ρbe gaussian filtering, ρ is standard deviation,

with respectively M ₁and M ₂mean value.

Unitary of illumination: for the unitary of illumination of recognition of face, set up a kind of new effective visual model.First, single-stage discrete two-dimensional wavelet transformation is used for realizing facial image is extracted to low frequency and high-frequency information.Then, consider illumination condition, process low-frequency information by imitating retinal information treatment mechanism, and large-scale high frequency coefficient is blocked by a threshold value.Finally, ask discrete two-dimensional to convert and get its inverse by the low frequency to applicable and high-frequency information and obtain unitary of illumination result.The detailed process of low frequency and high-frequency information is as follows:

Low-frequency information processing:

The first step, sets up photographic layer: the processing equation of low frequency sub-band A is as follows,

$A^p(x,y)=({\max }_A+A_{{0\mathrm{\&rho;}}_1}(x,y))A(x,y)/(A(x,y)+A_{{0\mathrm{\&rho;}}_1}(x,y))---\left(6\right)$

Wherein, max _athe maximal value of A,

the Adaptation factor of A at point (x, y), ρ ₁standard deviation and the A of gaussian filtering ^p(x, y) is corresponding output.

Second step, sets up outside lamina reticularis: A ^pprocessing as follows,

$A^o(x,y)=({\max }_{A^p}+A_{0_{{\mathrm{\&rho;}}_2}}^p(x,y))A^p(x,y)/(A^p(x,y)+A_{0_{{\mathrm{\&rho;}}_2}}^p(x,y))---\left(7\right)$

Wherein,

a ^pmaximal value,

a ^pat the Adaptation factor of point (x, y), ρ ₂standard deviation and the A of gaussian filtering ^o(x, y) exports accordingly.

The 3rd step, rim detection: adopt space bandpass filtering to carry out the information processing of simulated interior lamina reticularis.Be defined as follows:

$A^e(x,y)=(\exp (-\frac{(x^2+y^2)}{{2\mathrm{\&rho;}}_L^2})/2{\mathrm{\&pi;\&rho;}}_L^2-\exp (-\frac{(x^2+y^2)}{{2\mathrm{\&rho;}}_H^2})/2{\mathrm{\&pi;\&rho;}}_H^2)*A^o---\left(8\right)$

Wherein ρ _land ρ _hthe respective standard that is two Gaussian functions is poor, A ^ethe edge of presentation video.

High-frequency information processing:

Because concerning the facial image under different illumination conditions, a large amount of wavelet coefficients of high-frequency sub-band are unsettled, can have a negative impact in the situation that not processing for the high-frequency information of recognition of face.Therefore,, for improving the illumination robustness of high-frequency information, large high-frequency parameter is carried out to truncation.

Make H represent DH, DV or DD, the truncation of high-frequency sub-band is as follows:

$H^{\&prime;}(x,y)=\left\{\begin{array}{cc}t& H(x,y)>t\\ H(x,y)& \left|H(x,y)\right|\&le;t\\ -t& H(x,y)<-t\end{array}\right.---\left(9\right)$

t＝median(|H(x，y)|) (10)

Wherein median () is used for returning determinant | and the intermediate value of H|, by truncation, high-frequency sub-band is divided into DH ', DV ', DD '.

Experimental verification

In order to verify validity of the present invention, select the Yale B illumination face database of expansion.This storehouse comprises 38 personage's samples, nine attitudes, and 64 kinds of different illumination, amount to 21888 width images.Original image is of a size of 640 × 480, and all images, by artificial cutting, only comprise face's part, and image size is adjusted to 192 × 168 as far as possible.At cognitive phase, adopt classical PCA method for reducing dimension, embodiment retains 90% energy in experiment, and uses nearest neighbor classifier based on Euclidean distance to classify to facial image, then carries out discrete two-dimensional wavelet transformation with wavelet function db1.

In Fig. 1, top line is the original image of a people from subset one to subset five; Middle row is the design sketch that adopts Vu ' s method to obtain; Last column is the design sketch that adopts the present invention to obtain.

Table 1 is everyone discrimination (%) of selecting a width random image to reach as training set from subset one

Method	Subset one	Subset two	Subset three	Subset four	Subset five	All subsets
							Original image	2.63	2.78	3.06	4.13	3.33	3.25
LTV	92.06	80.22	92.18	85.99	91.46	88.50
							Gradient face	99.47	98.36	89.15	77.71	90.25	90.04
Vu	96.05	90.15	95.65	90.84	94.11	93.26
							Embodiment	97.75	96.00	96.33	95.24	95.22	95.86

In order to assess the performance of embodiment institute put forward the methods, test.From subset one, everyone selects a width random image as training set, and all the other are as test set, because above-mentioned training set and test set are random, have carried out 30 emulation and has got their mean value, and table 1 has been listed the average recognition rate of distinct methods.In sum, in table 1, data show that embodiment of the present invention algorithm used is subject to training set to select affect minimum and have good stability.

Claims (7)

1. extreme path, according to a unitary of illumination method for recognition of face, is characterized in that:

Analyze the functional layer of primate Modified Retinal Model, consult existing Modified Retinal Model, for the foundation based on amphiblestroid unitary of illumination model provides reference;

Utilize Naka-Rushton equation, for the foundation of unitary of illumination model provides concrete principle foundation;

In the time of the Adaptation factor calculating in Naka-Rushton equation, Adaptation factor is estimated in the situation that considering illumination difference;

Facial image is divided into low frequency and HFS by unitary of illumination processing section, and then take respectively different algorithms to process.

2. extreme path as claimed in claim 1 is according to the unitary of illumination method of recognition of face, it is characterized in that: set up unitary of illumination model: the Modified Retinal Model that utilizes primate, functional layer comprises photoreceptor layer, outside lamina reticularis OPL and inner lamina reticularis IPL, set up the unitary of illumination model based on Modified Retinal Model, comprise two continuous local image irradiation compression and space bandpass filterings.

3. extreme path as claimed in claim 1, according to the unitary of illumination method of recognition of face, is characterized in that: Naka-Rushton equation: calculate each pixel by carry out gaussian filtering on neighborhood, be defined as follows:

Wherein, p is the pixel in image, X ₀be to be the Adaptation factor that p is in pixel, X (p) is the intensity of input picture, and * represents convolution algorithm, G _ha 2-d gaussian filters device,

corresponding to the mean value of the image pixel intensities of X image.

4. extreme path as claimed in claim 1, according to the unitary of illumination method of recognition of face, is characterized in that: estimation Adaptation factor: I represents original image, and A represents the low frequency sub-band of I, and DH, DV and DD represent respectively the high-frequency sub-band of I;

Illumination-classification based on Otsu carries out on low frequency sub-band A, supposes that t is the threshold value being obtained by Otsu, and the illumination condition of A can approximately be divided into two classes:

M ₁＝{A(x，y)|A(x，y)≥t} (2)

M ₂＝{A(x，y)|A(x，y)＜t} (3)

Wherein, A (x, y) represents the intensity of A at point (x, y);

The Adaptation factor A of A _{o ρ}be defined as follows:

Wherein, G _ρbe gaussian filtering, ρ is standard deviation,

with

respectively M ₁and M ₂mean value.

5. the extreme path as described in claim 1-4 any one, according to the unitary of illumination method of recognition of face, is characterized in that: for the unitary of illumination of recognition of face, set up visual model:

Single-stage discrete two-dimensional wavelet transformation is used for realizing facial image is extracted to low frequency and high-frequency information;

Process low-frequency information by imitating retinal information treatment mechanism, and large-scale high frequency coefficient is blocked by a threshold value;

Ask discrete two-dimensional to convert and get its inverse by the low frequency to applicable and high-frequency information and obtain unitary of illumination result.

6. extreme path as claimed in claim 5, according to the unitary of illumination method of recognition of face, is characterized in that:

Low-frequency information processing:

The first step, sets up photographic layer: the processing equation of low frequency sub-band A is as follows,

Wherein, max _a-the maximal value of A,

the Adaptation factor of A at point (x, y), ρ ₁standard deviation and the A of gaussian filtering ^p(x, y) is corresponding output;

Second step, sets up outside lamina reticularis: A ^pprocessing as follows,

Wherein,

a ^pmaximal value, a ^pat the Adaptation factor of point (x, y), ρ ₂standard deviation and the A of gaussian filtering ^o(x, y) exports accordingly.

The 3rd step, rim detection: adopt space bandpass filtering to carry out the information processing of simulated interior lamina reticularis, be defined as follows:

Wherein, ρ _land ρ _hthe respective standard that is two Gaussian functions is poor, A ^ethe edge of presentation video.

7. extreme path as claimed in claim 5, according to the unitary of illumination method of recognition of face, is characterized in that:

High-frequency information processing:

Make H represent DH, DV or DD, the truncation of high-frequency sub-band is as follows:

t＝median(|H(x，y)|) (10)

Wherein, median () is used for returning determinant | and the intermediate value of H|, by truncation, high-frequency sub-band is divided into DH ', DV ', DD '.