CN105046202A - Adaptive face identification illumination processing method - Google Patents

Abstract

The invention discloses an adaptive face identification illumination processing method comprises: positioning a human face by means of an Adaboost method to obtain a human face part in a picture; performing illumination complexity computation on the human face after the human face is positioned; and performing illumination processing on the human face in virtue of different illumination processing methods by setting a complexity threshold value. The method processes an input image by means of the above three steps so as to obtain a final human face with eliminated illumination influence. The method may effectively suppress an influence of a light change on later face identification, prevents lack of robustness due to the same illumination processing method used in different environments, and may fast and accurately acquire a normalized human face image subjected to illumination processing.

Description

Adaptive recognition of face photo-irradiation treatment method

Technical field

The invention belongs to recognition of face preconditioning technique, specifically a kind of adaptive recognition of face photo-irradiation treatment method.

Background technology

Recent years, the research of controllable environment human face identification makes great progress, but under natural situation, in imaging process, the change of various uncontrollable factor always causes same face to produce very large change.Wherein the impact of illumination has been called one of bottleneck of face identification system undoubtedly.

The difference of same facial image under different illumination conditions may be greater than the difference of different facial image under same light is shone.Therefore, the solution of researchist's growing interest lighting issues, constantly combines existing method and improves, and proposes the method that many discriminations are higher, universality is stronger.These methods or utilize digital image processing techniques to carry out pre-service to illumination, or utilize mathematical theory combining image technology to convert image, to obtaining the good image of illumination, carrying out modeling process again or to facial image, attempting the information such as shape, attitude of restoring face; Again face is identified on this basis, mostly obtain good recognition performance.

Three-type-person's face identification photo-irradiation treatment method of current main flow has: histogram equalization, gradient face and illumination pretreatment chain.Histogram equalization method is that one carries out statistical study in the overall situation, the intensity profile in adjustment histogram, then regenerates histogrammic method.The method is mainly in order to suppress some information (gray scale that pixel proportion is less) selectively, and strengthen some information (gray scale that pixel proportion is more) needing performance in addition, namely the object of equalization is reached, when light conditions is good, to the recognition effect robust that the later stage carries out, and very simple and fast (1.StephenM.Pizer, E.PhilipAmburn, JohnD.Austin, RobertCromartie, AriGeselowitz.Adaptivehistogramequalizationanditsvariati ons, ComputerVision, Graphics, andImageProcessing, 1987, vol.39, pp.355-368.).Illumination pretreatment chain (TT) is by forming an illumination pretreatment chain by combine with technique such as gamma transformation, difference of Gaussian filtering, the adverse effect that light change brings can be eliminated well, achieve very high discrimination, be applicable to the face image processing that illumination effect is comparatively serious, but process is comparatively complicated.The processing procedure of gradient face (GRF) is: the gradient component first asking for image X and Y-direction, and doing convolution by Gauss's first order derivative and image can be in the hope of; Secondly, obtain the gradient component of image at X and Y-direction gradient component separately according to image, it is a kind of method of illumination robust, higher (the 2.StephenM.Pizer of face identification rate under illumination condition very complex situations, E.PhilipAmburn, JohnD.Austin, RobertCromartie, AriGeselowitz.FaceRecognitionUnderVaryingIlluminationUsi ngGradientfaces, ImageProcessing, IEEETransactionson, 2009, vol.18, pp.2599-2606).Three kinds of methods have their respective specific aims, all can not carry out the process of robust to each light conditions.

Summary of the invention

The object of the present invention is to provide one to provide different face complexity to judge, thus select different algorithm collocation, reach the method that can realize the photo-irradiation treatment effect of robust under any photoenvironment.

The technical solution realizing the object of the invention is: a kind of adaptive recognition of face photo-irradiation treatment method, and step is as follows:

The first step, locating human face, namely carries out the location of face to input picture by Adaboost.First a sample set is organized, obtain the Haar feature (face and non-face) of sample set, training classifier is carried out: whether correctly revise the weights of sample according to sample classification in every one deck sorter by feature, and send into lower one deck sorter training, then every one deck Multiple Classifier Fusion is risen and be used as final Adaboost Decision Classfication device.Then detect facial image with sorter, realize Face detection.

Second step, sets up illumination subspace, for three kinds of possible light conditions, organizes multiple sample set through Gaussian Blur process respectively, sample set is trained to the strong classifier obtaining several Weak Classifier cascades.Facial image through first step process is put in sorter, obtains the illumination complexity of face.

3rd step, according to the judgement of second step classification, if illumination effect is serious, then selects gradient face to carry out photo-irradiation treatment, then selects illumination pretreatment chain, if Irradiance well then can directly select histogram equalization to carry out photo-irradiation treatment if moderate.

Compared with prior art, its remarkable advantage: (1) has merged the human face light disposal route of multiple existing excellent, can to the photo-irradiation treatment normalization carrying out robust containing facial image in the present invention.(2) by judging different light conditions, selection difference targetedly photo-irradiation treatment method carries out unitary of illumination process to containing facial image.Photo-irradiation treatment is made to have very strong robustness and anti-noise ability.(3) introduce the concept of human face light complexity, merged the thought of Face datection cascade sorter, can select photo-irradiation treatment algorithm according to different face light conditions.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the present invention's adaptive recognition of face photo-irradiation treatment method.

Fig. 2 is conventional Haar feature appearance schematic diagram.

Fig. 3 is Gaussian Blur smoothed image schematic diagram.

Fig. 4 is cascade illumination complexity sorter schematic diagram.

Fig. 5 is gradient face photo-irradiation treatment method schematic diagram.

Fig. 6 is illumination pretreatment chain method schematic diagram.

Fig. 7 is histogram equalization schematic diagram.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described in further detail.

Composition graphs 1, the present invention's adaptive recognition of face photo-irradiation treatment method, step is as follows:

The first step, Face detection process is carried out to input picture: organize a sample set, extract the Haar feature of picture, pick out suitable Haar feature by Adaboost algorithm subsequently and form detecting device, Face datection is carried out to the facial image that contains of input, obtains facial image.

(1) Haar feature appearance is as shown in Figure 2 conventional Haar feature extraction, and these features value on face with non-face position characterizes different.By the pixel in white portion and the pixel deducted in black region and the value obtaining a masterplate.In image, the integrated value ii (x, y) of certain 1 A (x, y) is defined as definition s (x, y) is A (x, y) and y direction thereof upwards all pixel sums, like this, in the image of m*n size, each pixel calculates ii (x, y) with s (x, y), m*n*2 calculating can obtain whole integrogram matrix.

(2) Weak Classifier is chosen.

Wherein f is feature, and θ is threshold value, and p is the direction of the sign of inequality, and x is one and detects subwindow.

$q_{t,i}=\frac{w_{t,i}}{{\mathrm{\Σ}}_{j=1}^n{w}_{t,j}}$

Wherein w _t,ibe the weight of t sorter i-th feature, q _t,iit is the weight after normalization.

For each feature f, train a Weak Classifier h, calculate characteristic weighting fault rate ε _f,

ε _f＝Σ _iq _i|h(x _i,f,p,θ)-y _i|

Choose and there is minimal error rate ε _fweak Classifier h _i, and adjust the weight of sample

$w_{t+1,i}=w_{t,i}{\mathrm{\β}}_t^{1-e_i}$

Wherein e _i=0 represents x _icorrectly classified, otherwise be then mis-classification,

(3) strong classifier is cascaded into

Wherein ${\mathrm{\α}}_t=log\frac{1}{{\mathrm{\β}}_t}$

Image is detected, constantly the position of adjustment detection window and ratio, look for people's face and by its cutting.

Second step, human face light complicated dynamic behaviour.By the foundation of illumination subspace, carry out tolerance classification to facial image illumination complexity, the recognition of face photo-irradiation treatment entering into the 3rd step according to different illumination complexities is selected.Detailed process is as follows:

(1) set up sample set, this set comprises the facial image under multiple various illumination, the light conditions of image is classified: 1. almost without illumination effect, face in uniform light; 2. a small amount of illumination effect, as a small amount of polarisation, facial contour and edge visible; 3. illumination effect is serious, as face shade blocks scope greatly, and overexposure etc.

(2) Gaussian Blur is carried out to sample set picture:

$g(x,y)=\frac{1}{2{\mathrm{\π\σ}}^2}{e}^{-\frac{(x^2+y^2)}{2{\mathrm{\σ}}^2}}$

Wherein x, y represent the position of coordinate points respectively, and σ represents level and smooth degree, this method selected parameter σ=10.0.Level and smooth facial image, hides its obvious feature, obtains the face matrix of remarkable face comparison of light and shade.Smooth effect as shown in Figure 3.

(3) first step Face datection Haar feature classifiers thought used is used for reference, the sample set of a large amount of Gaussian Blur is trained, obtain many Weak Classifiers, multiple Weak Classifier is coupled together, obtain a strong classifier and light conditions is classified.

1. Weak Classifier is chosen.

Wherein f is feature, and θ is threshold value, and p is the direction of the sign of inequality, and x is one and detects subwindow.

$q_{t,i}=\frac{w_{t,i}}{{\mathrm{\Σ}}_{j=1}^n{w}_{t,j}}$

Wherein w _t,ibe the weight of t sorter i-th feature, q _t,iit is the weight after normalization.

2. in each class, for each feature f, train a Weak Classifier h, calculate the weighting fault rate ε of all images _f,

ε _f＝Σ _iq _i|h(x _i,f,p,θ)-y _i|

Choose and there is minimal error rate ε _fweak Classifier h _i, and adjust the weight of sample

$w_{t+1,i}=w_{t,i}{\mathrm{\β}}_t^{1-e_i}$

Wherein e _i=0 represents x _icorrectly classified, otherwise be then mis-classification,

3. three groups of strong classifiers are cascaded into

Wherein ${\mathrm{\α}}_t=log\frac{1}{{\mathrm{\β}}_t}.$

(4) cutting facial image to be sorted is put into the strong classifier of the 3rd step composition, if the result that ground floor sorter obtains is 0, then enters into lower one deck and judge, if the result obtained is 1, then obtains the classification of this image, enter into next step.

3rd step, recognition of face photo-irradiation treatment:

(1) if to be illumination effect serious for Images Classification, then Systematic selection gradient face.Detailed process is as follows:

1. first ask for the gradient component of image X and Y-direction, gradient conversion is defined as

$\frac{\∂I(x,y)}{\∂x}\≈\frac{\∂G(x,y,\mathrm{\σ})}{\∂x}*I(x,y)$

$\frac{\∂I(x,y)}{\∂y}\≈\frac{\∂G(x,y,\mathrm{\σ})}{\∂y}*I(x,y)$

Wherein wherein σ is standard deviation, and x direction gradient and y direction gradient are calculated as

$\begin{array}{c}{(x-gradient,y-gradient)}^T\\ =\▿\left(G\right(x,y,\mathrm{\σ})*I(x,y))=\left({\▿}_{x}G\right(x,y,\mathrm{\σ}),{\▿}_{x}G{(x,y,\mathrm{\σ})}^T)*I(x,y),\end{array}$

Wherein be expressed as x, the derivative in y direction.

2. being calculated as of eigenface:

$G=arctan\left(\frac{I_{y-gradient}}{I_{x-gradient}}\right).G\∈\[0.2\mathrm{\π})$

(2) if Images Classification is for being subject to a small amount of illumination effect, then Systematic selection illumination pretreatment chain processes illumination effect, and it is by being integrated by the light irradiation preprocess methods such as Gamma correction, contrast equalization, obtains an illumination pretreatment chain.Detailed process is as follows:

1. I is used ^γ(γ > 0) or log (I) (γ=0) replaces gradation of image I, as γ ∈ [0,1] and γ is a User Defined parameter.Index is γ and the power law of γ ∈ [0,0.5] scope is well compromise.The present invention selects γ=0.2 to arrange by default.

2. in order to reach illumination robustness, this method can ensure again the overall contrast of image in the overall situation within using contrast equalization to readjust gradation of image a to scope expected after Gamma correction simultaneously.Process is as follows:

$I(x,y)\←\frac{I(x,y)}{{\left(mean\right(\left|I\right(x,y\left){|}^a\right))}^{1/a}}$

$I(x,y)\←\frac{I(x,y)}{{\left(mean\right(min{(\mathrm{\τ},|I(x,y)\left|\right)}^a\left)\right)}^{1/a}}$

Wherein, a is a powerful compression index, can reduce the impact brought compared with high-gray level value individually; τ is a threshold value, is used for intercepting comparatively high-gray level value in equalization; Mean () is the averaging operation operator to whole image.The present invention gets α=0.1 herein and τ=10 are worth by default.

3. last, the strategy that this method applies a kind of Nonlinear Mapping has carried out the suppression to image intensity value maximum value, i.e. I (x, y) ← τ tanh (I (x, y)/τ), wherein tanh () is hyperbolic tangent function, and the span of I is limited in (-τ, τ).

(3) if almost without illumination effect, native system selects histogram equalization method to process.Step is as follows:

1. the probability density function of computed image, becomes 0 ~ 1 by tonal range from 0 ~ 255, by its normalization.Use P _rr () represents the probability density function of original image, P _s(S) represent the probability density function that equalization is later, r, s represent the gray-scale value before and after histogram equalization respectively, and belong to [0,1].Can be obtained by the knowledge of theory of probability wherein T ^-1s inverse transform function that () is T (r), and have $s=T\[r\]={\∫}_0^r{P}_r\left(w\right)dw$

2. Cumulative Distribution Function is calculated.The probability occurred for certain gray-level pixels is:

P _r(r _k)＝n _k/N

Wherein P _r(r _k) be the probability that K gray-level pixels of original image occurs, r _kbe a kth gray level, i.e. current color range, scope is [0,1], and N is total number of image pixels, n _kfor r _kpixel quantity.

3. gray-level histogram equalization formula is:

$s=T\left(r_k\right)=\underset{j=0}{\overset{k}{\Σ}}{n}_j/N=\underset{j=0}{\overset{k}{\Σ}}{P}_r\left(r_k\right)$

Wherein, T (rk) represents the transfer function of a kth gray level of original image.∑ Pr (rk) represents the gray level probability of occurrence cumulative addition of 0th ~ k.Because s is normalized numerical value (s ∈ [0,1]), the color value of 0 ~ 255 is converted to, needs to be multiplied by 255 again, namely

S＝∑Pr(rk)*255

In sum, the present invention, by merging the human face light disposal route of multiple existing excellent, can carry out adaptive photo-irradiation treatment normalization to containing facial image.And by judging different light conditions, selection difference targetedly photo-irradiation treatment method carries out unitary of illumination process to containing facial image, makes photo-irradiation treatment have very strong robustness and anti-noise ability.

Claims (5)

1. an adaptive recognition of face photo-irradiation treatment method, is characterized in that step is as follows:

The first step, the process of training Haar feature classifiers, containing the image of face, realizes Face detection, is cut out people face part;

Second step, sets up illumination subspace, for three kinds of possible light conditions, organizes multiple sample set through Gaussian Blur process respectively, sample set is trained to the strong classifier obtaining several Weak Classifier cascades;

3rd step, puts in sorter by the facial image through first step process, obtains the illumination complexity of face;

4th step, according to the illumination complexity of the face that the 3rd step obtains, if illumination effect is serious, gradient face is then selected to carry out photo-irradiation treatment, if a small amount of illumination effect, then select illumination pretreatment chain to carry out photo-irradiation treatment, if almost without illumination effect, then select histogram equalization to carry out photo-irradiation treatment.

2. adaptive recognition of face photo-irradiation treatment method according to claim 1, is characterized in that setting up illumination subspace and to carry out the concrete steps of illumination complexity judgement as follows:

2.1 set up sample set, and this set comprises the facial image under multiple various illumination, classify: 1. almost without illumination effect, i.e. face in uniform light according to the light conditions of criterion by sample image; 2. a small amount of illumination effect, i.e. a small amount of polarisation, facial contour and edge visible; 3. serious illumination effect, namely face shade blocks scope greatly, overexposure;

2.2 pairs of sample set pictures carry out Gaussian Blur:

Wherein, x, y represent the position of coordinate points respectively, and σ represents level and smooth degree, use the level and smooth facial image of Gaussian Blur, hide its facial face feature, only obtain the image of face's comparison of light and shade;

2.3 employings are trained the sample set through Gaussian Blur based on the Adaboost method of Haar feature, obtain many Weak Classifiers, are coupled together by multiple Weak Classifier, obtain a strong classifier.

3. adaptive recognition of face photo-irradiation treatment method according to claim 2, is characterized in that: the concrete grammar described in step 2.3 is as follows:

2.3.1 initializes weights:

Wherein f is feature, and θ is threshold value, and p is the direction of the sign of inequality, and x is one and detects subwindow;

Wherein, w _t,ibe the weight of t sorter i-th feature, q _t,iit is the weight after normalization;

2.3.2 for the light image f of the multiple angle of each class, train a Weak Classifier h, calculate the weighting fault rate ε of all images _f,

ε _f＝Σ _iq _i|h(x _i,f,p,θ)-y _i|

Choose and there is minimal error rate ε _fweak Classifier h _i, and adjust the weight of sample

Wherein e _i=0 represents x _icorrectly classified, otherwise be then mis-classification,

Wherein ε _tbe the weighting fault rate of t sorter, for Dynamic gene;

2.3.3 the Weak Classifier obtained in 2.3.2 is cascaded into three groups of strong classifiers

Wherein

4. adaptive recognition of face photo-irradiation treatment method according to claim 1 and 2, it is characterized in that: the strong classifier described in step 3, cutting facial image to be sorted being put into second step composition, if the result that ground floor sorter obtains is 1, then to belong to illumination effect serious for image; If the result that ground floor sorter obtains is 0, then enters into lower one deck and judge, if the result that now sorter obtains is 1, then to belong to illumination effect moderate for image; Otherwise judge that it is almost do not have illumination effect.

5. self-adaptation recognition of face photo-irradiation treatment method according to claim 1, is characterized in that the concrete grammar of described step 4 is as follows:

When being categorized as normal illumination, the image array dark pixel namely after Gaussian Blur accounts for 5% ~ 30%, then select histogram equalization to process picture;

When being categorized as illumination effect slightly, the image array dark pixel namely after Gaussian Blur accounts for 30% ~ 70%, then select illumination pretreatment chain to process picture;

When being categorized as serious illumination effect, the image array dark pixel namely after Gaussian Blur accounts for 70% ~ 100%, then select gradient face to process picture.