CN101794389B - Illumination pretreatment method of facial image - Google Patents

Abstract

The invention provides an illumination pretreatment method of facial images, comprising the following steps: 1) under m numbered different cut-off scales, respectively decomposing a facial image into a large-scale component u corresponding to illumination factors and a small-scale component v corresponding to the internal characteristics of the face so as to obtain m numbered small-scale components of the facial image; 2) calculating the difference of the small-scale components under adjacent cut-off scales to obtain m-1 numbered scale neighbourhood components of the facial image; and 3) carrying out weighted summation by the minimum small-scale component and m-1 numbered scale neighbourhood components to obtain the illumination pretreatment image of the facial image. The invention can accurately capture and reserve the internal characteristics of faces of different facial images under various illumination conditions so as to effectively strengthen robustness of face recognition to illumination variation.

Description

A kind of illumination pretreatment method of facial image

Technical field

The present invention relates to Flame Image Process, computer vision and mode identification technology, specifically, the present invention relates to a kind of illumination pretreatment method of facial image.

Background technology

Face recognition technology is a kind of biometrics identification technology that utilizes computing machine to carry out identity validation or identification through facial image.It has can carry out contactless IMAQ, but hidden operation, and a series of unique advantages such as the low and interactivity of image capture device cost is strong have a good application prospect.Yet illumination condition complicated and changeable can constitute the performance of recognition of face and has a strong impact on, and this mainly is because in the image imaging process, can produce sidelight, shade, overexposure and multiple unfavorable factor such as under-exposure.In order to reduce the adverse effect that complicated illumination variation produces recognition of face; Realization is to the recognition of face of illumination variation robust, and it is the solution of target that researchers have proposed with unitary of illumination, to the feature extraction of illumination variation robust with to illumination variation modeling etc.In these solutions; There are a lot of schemes to be based on that image processing techniques realizes,, have succinct, characteristics of high efficiency based on the scheme that image processing techniques realizes; And be independent of follow-up face recognition algorithms and carry out in advance, therefore be called as light irradiation preprocess method.Before facial image is discerned; At first utilize image processing techniques that all facial images are carried out pre-service to eliminate complicated and diversified illumination variation in the facial image; And then when utilizing certain recognizer that pretreated facial image is discerned, just can obtain higher, more stable recognition performance.

At present; Light irradiation preprocess method is mainly based on the estimation of the estimation of people's face internal characteristics in the facial image or illumination factor and carry out; In other words; Through the facial image decomposition technique, with the input facial image be decomposed into the corresponding small scale component of people's face internal characteristics and with the corresponding large scale component of illumination factor two parts.Existing facial image decomposition technique comprises: based on the method for discrete Fourier transformation, (can reference: W.Chen based on the method for discrete cosine transform; Et al.; " IlluminationCompensation and Normalization for Robust Face Recognition Using DiscreteCosine Transform in Logarithm Domain; " TSMCB; 2006.), based on the method for Hi-pass filter, based on the method for low-pass filter and (can reference: R.Gonzalez and R.Woods, Digital Image Processing.NJ, USA:Prentice Hall based on the method for BPF.; 1992, pp.91-94.) etc.In addition, the people such as L.Rudin of the U.S. proposed in 1992 based on L ²Norm is as total variation model (the Total Variation-L of fidelity tolerance ², be called for short TV-L ²Perhaps ROF) carries out image denoising.And in 2004, people such as American scholar T .Chan attempted using L ¹Norm replaces L ²Norm is measured as the fidelity in the ROF model, and has studied and transformed back model (Total Variation-L ¹, be called for short TV-L ¹) characteristic, finally find TV-L ¹Model can be applied to the yardstick decomposition of image and select based on the parameter of data-driven.Further, the people such as T.Chen of the U.S. consider TV-L ¹The picture breakdown characteristic of model is with TV-L ¹Model is incorporated into computer vision field and is used for the illumination pretreatment of facial image (can reference: T.Chen, et al., " Total Variation Models for Variable LightingFace Recognition, " TPAMI, 2006).In this light irradiation preprocess method, at first through log-transformation (Logarithmic Transformation is called for short LOG), I is transformed into log-domain with facial image, obtains image f, utilizes TV-L then ¹Model with the facial image f in the log-domain be decomposed into illumination factor corresponding large scale component u and with corresponding small scale component v two parts of people's face internal characteristics; Above-mentioned model is log-domain total variation model (Logarithmic Total Variation is called for short the LTV model).For facial image, at first carry out pre-service through the LTV model, utilize face recognition algorithms that wherein small scale component v is discerned then and can obtain the illumination variation face recognition result of robust comparatively.Yet still there are some defectives in above-mentioned facial image pretreating scheme.Such as: under extensive sample set and non-controlled illumination condition, the eyes in the different facial images, nose, characteristics such as face and profile also not exclusively are distributed in the consistent relatively small scale scope, promptly are not enough to comprised by single small scale component v; Equally; Complicated illumination factor also not exclusively is distributed in the consistent relatively large scale scope; But log-domain total variation model only utilizes a cut-off scales parameter that piece image f is decomposed into u and v two parts; Therefore; When log-domain total variation model carries out illumination pretreatment under extensive sample set and non-controlled illumination condition, thus its final illumination pretreatment as a result v tend to lose some and can't guarantee illumination pretreatment result's stability the useful people's face internal characteristics of recognition of face, thereby cause the unstable properties of final face recognition algorithms.It is thus clear that; The preprocess method that people such as T.Chen propose is not enough to handle the feature difference and complicated illumination variation of different people face, and this has influenced log-domain total variation model to a great extent in the validity of extensive sample set, non-controlled illumination condition being carried out illumination pretreatment down on the face database of collection.

Summary of the invention

The purpose of this invention is to provide a kind of be suitable for the handling feature difference of different people face and the light irradiation preprocess method of the illumination variation of complicacy.

For realizing the foregoing invention purpose, the invention provides a kind of illumination pretreatment method of facial image, comprise the following steps:

1) under m different cut-off scales, with facial image be decomposed into respectively with illumination factor corresponding large scale component u and with the corresponding small scale component of people's face internal characteristics v, thereby obtain m small scale component of said facial image; M is at least 2 integer;

2) difference of the small scale component under the adjacent cut-off scales of calculating obtains m-1 yardstick neighborhood component of said facial image;

3) carry out weighted sum through small scale component and m-1 yardstick neighborhood component, obtain the illumination pretreatment image of said facial image minimum.

Wherein, in the said step 1), utilize log-domain total variation model that said facial image is decomposed, obtain said m small scale component.

Wherein, said step 3) also comprises: the small scale component of said minimum and less yardstick neighborhood component are set bigger weight, and bigger yardstick neighborhood component is set less weight.

Wherein, said step 3) also comprises: the weight that obtains the small scale component and m-1 the yardstick neighborhood component of said minimum through the study based on training set.

Wherein, said step 3) also comprises: the differentiation power that is based on each yardstick neighborhood component on the training set is confirmed said weight; Be based on the order of on the training set each yardstick neighborhood component being carried out feature selecting and confirm said weight; The differentiation entropy that perhaps is based on each yardstick neighborhood component on the test set is confirmed said weight.

Compared with prior art, the present invention has following technique effect:

The present invention can more accurately catch and keep people's face internal characteristics of the facial image of the different people under the various illumination conditions, and then strengthens the robustness of recognition of face to illumination variation effectively.

Description of drawings

Below, specify embodiments of the invention in conjunction with accompanying drawing, wherein:

Fig. 1 is based on the schematic flow sheet of the facial image identification of one embodiment of the invention;

Fig. 2 is the light irradiation preprocess method schematic flow sheet of one embodiment of the invention based on yardstick weighting log-domain total variation model.

Embodiment

The invention provides a kind of illumination pretreatment method of facial image, it is a pre-treatment step of facial image identification, and Fig. 1 shows the schematic flow sheet based on the facial image identification of one embodiment of the invention.

The theoretical foundation that facial image illumination preconditioning technique relates to comprises reflection-illumination model (reflectance-illumination model).Based on this model, the imaging process of a width of cloth facial image I can be represented reflecting attribute (people's face internal characteristics) R and the product of illumination L for people's face:

I＝RL (1)

In general, illumination pretreatment is exactly reflecting attribute (the people's face internal characteristics) R that will from facial image, extract wherein, so just can eliminate the influence of illumination L to the facial image recognition performance.

According to one embodiment of present invention, adopted log-domain total variation model that facial image is carried out multiple dimensioned decomposition.At first; Because shown in the formula (1) is the property taken advantage of model; Therefore the imaging process of facial image of need changing commanders is additive model (adopting logarithm operation to realize as this step 1) by the property taken advantage of model conversion, so that itself and log-domain total variation Model Matching, the additive model after the conversion is following:

logI＝logR+logL (2)

Follow-up for ease description is expressed as formula (2) again:

f＝u+v (3)

Wherein, f=logI is the facial image in the log-domain, and u=logL is illumination factor in the log-domain (large scale component), and v=logR is the people's face internal characteristics (small scale component) in the log-domain.

Based on the above-mentioned theory basis, describe each step of this embodiment in detail below in conjunction with accompanying drawing 2.

Step 1: input facial image.

Step 2: utilize log-domain total variation model that facial image is carried out multiple dimensioned decomposition, obtain the combination of a plurality of large scale components and small scale component.

Based on the expression of people's face imaging model in log-domain in the formula (3),,, calculate large scale component u through the minimization problem in the solution formula (4) according to log-domain total variation model:

$\hat{u}=\arg \underset{u}{\min \∫}|\▿u|+\mathrm{\λ}{\left|\right|I-u\left|\right|}_{L^1}---\left(4\right)$

In case obtain the estimation of large scale component u,, can draw small scale component v=f-u according to formula (3).Cut-off scales when wherein parameter lambda can be controlled image f is decomposed into large scale component u and small scale component v (can reference: T.Chen; Et al.; " Total Variation Models for Variable LightingFace Recognition, " TPAMI, 2006).As shown in Figure 2, in the present embodiment, (in a preferred embodiment, can set m=13,13 λ are respectively λ to have adopted m different parameter lambda ₁=1.2, λ ₂=1.1 ..., λ ₁₃=0).Based on this m different parameter lambda, utilize log-domain total variation model that the input facial image is decomposed, obtain small scale component v _i, i=1,2 ..., m.When model parameter satisfies λ ₁＞λ ₂＞...,＞λ _m, according to the character of log-domain total variation model, then the small scale component satisfies $v_1\⋐v_2\⋐,...,\⋐v_m.$ Wherein, $v_i\⋐v_{i+1}$ Be expression v _I+1In comprised v _iIn everyone face internal characteristics, and v _I+1In also comprised and compared v _iIn bigger other people's face internal characteristics of everyone face internal characteristics yardstick (i=1 here, 2 ..., m-1).People's face internal characteristics had both comprised human faces such as eyes, nose, face, also comprised the facial contour characteristic.For same organ, under different scale, also can reflect different people's face internal characteristicses, such as in Fig. 2, v ₁In the face region show as a camber line, Here it is people's face internal characteristics, and v than small scale _mIn the face region then can show the profile of lip, this is a people's face internal characteristics than large scale.And, v _mIn the people's face internal characteristics that comprises than large scale, also must comprise v ₁In people's face internal characteristics of being comprised than small scale.It should be noted that above description to people's face internal characteristics and corresponding yardstick thereof only is an exemplary in nature, and non exhaustive.

Step 3: obtain yardstick neighborhood component.

Two adjacent small scale components (promptly the component difference v) corresponding to people's face internal characteristics can be defined as yardstick neighborhood component, and minimum small scale component v ₁Can regard the yardstick neighborhood component between this small scale component and 0 component as, therefore, definable yardstick neighborhood component is s _i, i=1,2 ..., m:

$s_i=\left\{\begin{array}{cc}{v}_1& i=1\\ v_i-v_{i-1},& i=\mathrm{2,3},...,m\end{array}\right.---\left(5\right)$

Usually, all contain one or more people's face internal characteristicses in each yardstick neighborhood component.Such as under a cut-off scales; The small scale component contains people's face internal characteristics A, B, C; And under adjacent with it bigger cut-off scales, the small scale component only contains people's face internal characteristics A, B, and two small scale components subtract each other can obtain people's face internal characteristics C comparatively accurately.Therefore, calculate all yardstick neighborhood components after, can obtain each required individual face internal characteristics of recognition of face comparatively accurately.

Step 4:, obtain final illumination pretreatment result based on yardstick weighting log-domain total variation model through weighted sum to all yardstick neighborhood components.Illumination pretreatment I ' computing formula as a result is following:

$I^{\′}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i{s}_i---\left(6\right)$

Adopted the fixed weight strategy in the present embodiment; Consider that the useful people's face internal characteristics majority of recognition of face is distributed in the less yardstick neighborhood component, and minority is distributed in the bigger yardstick neighborhood component; Therefore; Adopt bigger weight for less yardstick neighborhood component, and adopt less wooden fork heavy, help improving the accuracy rate of identification like this for bigger yardstick neighborhood component.In a preferred embodiment, the heavy α of wooden fork _iBy formula calculate (7):

${\mathrm{\α}}_i=\frac{1}{M}(m-i+1),i=\mathrm{1,2},...,m---\left(7\right)$

Wherein, $M=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}i.$ Illumination pretreatment result in formula this moment (6) can be expressed as:

$I^{\′}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_i{s}_i$

$=\frac{1}{M}(\mathrm{<figure-callout\; id="1"\; label="m\; s"\; filenames="H2009102442710E00011.png,H2009102442710E00021.png"\; state="\{\{state\}\}">m</figure-callout>}{s}_{}{}_1+(m-1)s_2+...+s_m)$

$=\frac{1}{M}({\mathrm{<figure-callout\; id="1"\; label="s"\; filenames="H2009102442710E00011.png,H2009102442710E00021.png"\; state="\{\{state\}\}">s</figure-callout>}}_1+({\mathrm{<figure-callout\; id="1"\; label="s"\; filenames="H2009102442710E00011.png,H2009102442710E00021.png"\; state="\{\{state\}\}">s</figure-callout>}}_1+s_2)+...+({\mathrm{<figure-callout\; id="1"\; label="s"\; filenames="H2009102442710E00011.png,H2009102442710E00021.png"\; state="\{\{state\}\}">s</figure-callout>}}_1+s_2+...+s_m))---\left(8\right)$

$=\frac{1}{M}(v_1+v_2+...+v_m)$

$=\frac{1}{M}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{v}_i$

Step 5: output facial image illumination pre-service is I ' as a result.

Illumination pretreatment as a result I ' can with multiple recognizer commonly used coupling, comprising: based on the arest neighbors sorting technique of correlativity or various distances such as Euclidean distance or mahalanobis distance (can reference: R.Duda, et al. pattern classification; Second edition, China Machine Press, 2003; Pp.146-150.), the Eigenfaces method (can reference: M.Turk and A.Pentland; " EigenEaces for Recognition, " JCN, 1991.), the Fisherfaces method (can reference: P.Belhumeur; Et al.; " Eigenfaces vs.Fisherfaces:Recognition Using Class Specific Linear Projection, " TPAMI, 1997.) etc.

According to the foregoing description; YaleB+Extended YaleB and CAS-PEAL lighting face database are carried out illumination pretreatment, adopt the experiment of discerning as the enterprising pedestrian's face of face database of recognizer after the process illumination pretreatment based on the arest neighbors sorting technique of correlativity then.The light irradiation preprocess method of present embodiment has been obtained 86.80% and 14.40% discrimination respectively, is higher than under equal test condition the discrimination 48.76% and 5.13% that the log-domain total variation model of single yardstick is obtained on these two databases far away.

In addition, discover, in the foregoing description; Even original image is not done log-transformation and directly original image is decomposed, perhaps be similar to transformation of variable, such as index variation etc. with other conversion; Carry out picture breakdown then, also can obtain good picture breakdown effect.

Also be pointed out that; Though the picture breakdown algorithm that the foregoing description adopts is based on the algorithm of log-domain total variation model; But those skilled in the art also can be according to design of the present invention; Employing realizes the picture breakdown algorithm based on methods such as discrete Fourier transformation, discrete cosine transform, Hi-pass filter, low-pass filter or BPF.s, and then realizes illumination pretreatment.In these picture breakdown algorithms, through setting the different cut-off frequency, can obtain the small scale component under the different cut-off scales, and then draw yardstick field component and final pre-service result.And in these picture breakdown algorithms, can not will the imaging model of input picture be additive model from the property taken advantage of model conversion.

Foregoing only is for describing the present invention's listed examples; And the scope of unrestricted patent protection of the present invention; Technical scheme and improvement thereof that all do not break away from the modification that aim of the present invention carries out or are equal to replacement all should not got rid of outside the protection domain of claim of the present invention.

Claims (8)

1. an illumination pretreatment method of facial image comprises the following steps:

1) under m different cut-off scales, with facial image be decomposed into respectively with illumination factor corresponding large scale component u and with the corresponding small scale component of people's face internal characteristics v, thereby obtain m small scale component of said facial image; M is at least 2 integer;

2) difference of the small scale component under the adjacent cut-off scales of calculating obtains m-1 yardstick neighborhood component of said facial image;

3) carry out weighted sum through small scale component and m-1 yardstick neighborhood component, obtain the illumination pretreatment image of said facial image minimum.

2. illumination pretreatment method of facial image according to claim 1; It is characterized in that; In the said step 1); Utilize log-domain total variation model, discrete Fourier transformation, discrete cosine transform, Hi-pass filter, low-pass filter or BPF. that said facial image is decomposed, obtain said m small scale component.

3. illumination pretreatment method of facial image according to claim 1; It is characterized in that; Said step 3) also comprises: the small scale component of minimum and less yardstick neighborhood component are all set bigger weight, bigger yardstick neighborhood component is set less weight.

4. illumination pretreatment method of facial image according to claim 3 is characterized in that, in the said step 3), is s for yardstick neighborhood component _i, weight corresponding with it

Wherein, i=1,2 ..., m,

5. illumination pretreatment method of facial image according to claim 1 is characterized in that, said step 3) also comprises: the weight that obtains the small scale component and m-1 the yardstick neighborhood component of said minimum through the study based on training set.

6. illumination pretreatment method of facial image according to claim 5 is characterized in that, said step 3) also comprises: the differentiation power that is based on each yardstick neighborhood component on the training set is confirmed said weight.

7. illumination pretreatment method of facial image according to claim 5 is characterized in that, said step 3) also comprises: be based on the order of on the training set each yardstick neighborhood component being carried out feature selecting and confirm said weight.

8. illumination pretreatment method of facial image according to claim 5 is characterized in that, said step 3) also comprises: the differentiation entropy that is based on each yardstick neighborhood component on the test set is confirmed said weight.

Images