CN101261678A - A method for normalizing face light on feature image with different size - Google Patents

Abstract

The invention discloses a method for unifying face illuminations on pictures with different size characteristics. Firstly, a log-domain total variation model is adopted to decompose an original face picture into small size characteristic pictures and big size characteristic pictures; then an illumination processing is carried out to the big size characteristic pictures which are greatly impacted by illumination changes and a minimal value filtering wave processing with threshold value is carried out to the small size characteristic pictures; finally, the processed pictures with different size characteristics are composed to obtain a face picture with unified illumination. The invention mainly carries out the illumination unification on the big size characteristic pictures which are greatly impacted by illumination changes to avoid the impact on face identification rate caused by changing the small size characteristics with no illumination change. Moreover, the invention does not give up the big size characteristics which are greatly impacted by illumination so as to avoid the identification information lack caused by just adopting the small size characteristics for face identification. The method of the invention can be realized easily without strict alignment to the face pictures and without any training samples, meets various practical application requirements.

Description

Method at the enterprising normalizing face light of feature image with different size

Technical field

The present invention relates to technology such as recognition of face, facial image pre-service, relate in particular to a kind of people's face gray level image illumination normalization method.

Background technology

People's face technology all is widely used at aspects such as public safety system, identity discriminating, man-machine interaction and multimedia recreations, yet the illumination variation problem but is to perplex the one of the main reasons of such technical applicationization for a long time.Illumination variation not only has influence on the visual effect of facial image, and has a strong impact on the recognition of face rate.In the last thirty years, various technology are suggested and solve the human face light variation issue, but wherein great majority do not reach practical requirement as yet or can not satisfy many-sided application requirements.

Based on Lambertian model I (x, y)=R (x, y) L (x, y), wherein I is 2D people's face gray level image, R is the reflecting component image, L is the illumination component image.Because people's face surface reflectivity and illumination variation are irrelevant, so, there are class methods to attempt facial image to be decomposed according to the Lambertian model, only select for use R to carry out recognition of face then.The wherein more famous algorithm that has people such as Terrence Chen to propose based on the LTV model in 2006, promptly with full variation model facial image is decomposed (T.Chen at log-domain, X.S.Zhou, D.Comaniciu and T.S.Huang.Total Variation Modelsfor Variable Lighting Face Recognition.IEEE Transactions on Pattern Analysisand Machine Intelligence, 28 (9): 1519-1524,2006).Yet, in theory, from former figure, extract reflecting component originally as ill-conditioning problem, so, the method that exists only can be decomposed into small scale features image and large-scale characteristics image with a facial image approx at present, carry out recognition of face with the small scale features image then, and abandon the large-scale characteristics image.Consider and still contain in the large scale image in a large number the recognition of face Useful Information that therefore, such technology abandons the large-scale characteristics image will cause the identifying information deficiency, in addition, these class methods can't be improved the visual effect of facial image.In the method that exists, it is original facial image directly to be carried out illumination correct that other class methods are arranged, and produces the facial image under the standard illumination condition, and these class methods can reach the improvement on the visual effect.Yet illumination variation mainly has influence on the low frequency part of facial image, and these class methods are carrying out may having carried out unnecessary change to the constant details composition of illumination when illumination is corrected to whole image, thereby can have influence on the recognition of face rate.In a word, all there is deficiency in above-mentioned two class methods, and these deficiencies can solve by two class technology are carried out combination.

Summary of the invention

The object of the present invention is to provide a kind of method at the enterprising normalizing face light of feature image with different size, this method is practical, applied range, can obviously improve the discrimination of people's face.

It is small scale features image and large-scale characteristics image with the picture breakdown of primitive man's face that this method is at first used the total variation model (logarithmic total variation model is called for short the LTV model below) of log-domain; Then the large-scale characteristics image that is subjected to the illumination variation influence is carried out photo-irradiation treatment, wherein, the present invention is with LOG-DCT technology (W.L.Chen, E.M.Joo and S.Wu.IlluminationCompensation and Normalization for Robust Face Recognition using DiscreteCosine Transform in Logarithm domain.IEEE Transactions on Systems, Manand Cybernetics, Part B, 36 (2): 458～466,2006) be applied to that the large-scale characteristics image is carried out illumination and correct; To of the minimum value filtering of small scale features image with threshold value; The feature image with different size handled of utilization synthesizes the facial image that obtains after the illumination normalization at last.This method has overcome the deficiency of two class photo-irradiation treatment methods in the background introduction, facial image is carried out photo-irradiation treatment can obtain extra high recognition of face rate, has proposed a new framework thinking for people's face photo-irradiation treatment technology simultaneously.

The present invention mainly is achieved through the following technical solutions: the method at the enterprising normalizing face light of feature image with different size comprises the steps:

1. the cutting of aliging is handled to people's face gray level image of input.Promptly to every facial image, three unique points (the pupil center's point of two eyes and the central point of face) of elder generation's this people's face of detection and location, make two eyes of every people's face be horizontal by rotation, use two interpolation algorithm stretching images again, make these three unique points be positioned at the fixed position of image, at last image is cut to unified size.

2. with the LTV model facial image is decomposed, the facial image I to after every cutting alignment carries out log-transformation to it

f(x，y)＝logI(x，y) (1)，

Find the solution following variation model:

$u=\arg \underset{u}{\min }\∫|\▿u|+\mathrm{\λ}{\left|\right|f-u\left|\right|}_{L^1}\mathrm{dx}---\left(2\right)$

And

v＝f-u (3)，

Obtain the large-scale characteristics image S and the small scale features image ρ of people's face this moment:

S＝exp(u)，ρ＝exp(v) (4)。

3. to the minimum value filtering of small scale features image ρ with threshold value.Carrying out this step processing mainly is in order to eliminate people's face white point spot that LTV model decomposable process produces, to improve visual effect.Way is with 3 * 3 filter window ρ to be carried out minimum value filtering, if the gray-scale value of current window central point, is then got the gray-scale value that the minimum gradation value of current window replaces center pixel greater than assign thresholds.Remember that filtered small scale features image is ρ '.

4. remove some low frequency discrete cosine transforms (DCT) coefficient of large-scale characteristics image at log-domain.Promptly u is carried out DCT, note DCT coefficient be C (α, β), α=0,1 ..., M, β=0,1 ..., N, wherein M, N are length and wide (pixel) of facial image, then with n around the frequency domain initial point ²Individual DCT coefficient is set to 0:

C(α，β)＝0，α＝0，1，…，n，β＝0，1，…，n (5)，

Obtain by anti-dct transform

, carry out the large-scale characteristics image S after the illumination normalization is obtained in exponential transform at last _Norm:

$S_{\mathrm{norm}}(x,y)=\exp \hat{u}(x,y)---\left(6\right).$

Being called for short step below 4. is LOG-DCT.

5. the facial image after the synthetic illumination normalization:

I _norm(x，y)＝ρ′(x，y)S _norm(x，y) (7)。

The present invention compared with prior art has following advantage and beneficial effect:

1, the present invention proposes carries out human face light normalization method at feature image with different size and does not abandon the people who the is subjected to illumination effect scale feature of being bold, having avoided the bashful scale feature of only personnel selection to carry out recognition of face and caused the problem of identifying information deficiency, experiment shows that this method can obtain the recognition of face rate higher than LTV model.In addition, keep large-scale characteristics, guarantee to carry out the improvement that facial image after the photo-irradiation treatment has really reached visual effect.

2, the inventive method mainly is being subjected to carry out the illumination normalization on the large-scale characteristics image of illumination effect, avoids the constant small scale features of illumination is changed and influences the recognition of face rate.

3, algorithm of the present invention is easy to realize, does not need facial image is carried out the strictness alignment, also without any need for training sample, meets various application request.

Description of drawings

Fig. 1 is the operational flowchart of the inventive method.

Fig. 2 is facial image shape alignment normalization constitutional diagram of the present invention: a left side is protoplast's face figure, the right people's face figure that is cut to 100 * 100 pixel sizes for alignment.

Fig. 3 is LTV model decomposition result figure of the present invention: Zuo Weiyuan figure, in be the large-scale characteristics image, the right side is the small scale features image.

Fig. 4 is the minimum filtering of band threshold value of the present invention figure as a result: a left side is the bashful scale feature figure of protoplast, and the right side is corresponding filtering result.

Fig. 5 uses the LOG-DCT algorithm to the people result that the scale feature image carries out the illumination normalization that is bold for the present invention: a left side is the protoplast scale feature figure that is bold, and the right side is corresponding illumination normalization result.

Fig. 6 is an algorithm flow exemplary plot of the present invention.

Fig. 7 is for carrying out the result that human face light is handled with the inventive method: the former figure of first behavior, second behavior illumination normalization of the present invention is figure as a result.

Fig. 8 is that distinct methods is at the ROC of CMU face database curve synoptic diagram.

Embodiment

The present invention is described in further detail below in conjunction with embodiment and accompanying drawing, but embodiments of the present invention are not limited thereto.

Fig. 1 shows operating process of the present invention, and as seen from Figure 1, this comprises the steps: in the method for the enterprising normalizing face light of feature image with different size

(1) cutting of aliging is handled to people's face gray level image I of input.Promptly to every facial image, choose three unique points (the pupil center's point of two eyes and the central point of face) of this people's face by hand, make two eyes of every people's face be horizontal by rotation, use two interpolation algorithm stretching images again, make three unique points be positioned at the fixed position of image, at last image is cut to 100 * 100 sizes.(see figure 2).

(2) by finding the solution the LTV model facial image I decomposition is the large-scale characteristics image S of people's face and small scale features image ρ (see figure 3).

(3) to the minimum value filtering of small scale features image ρ with threshold value, the filtering result is ρ '.(see figure 4).

(4) use LOG-DCT people's scalogram of being bold is looked like to carry out the illumination normalization, its Chinese style is got n=13 in (5), and result is S _Norm(see figure 5).

(5) with ρ ' and S _NormFacial image after the synthetic illumination normalization.

Step (2)～step (5) has shown the illumination normalization flow process (see Fig. 6 and Fig. 7) of the inventive method to concrete certain facial image.

The present invention describes effect of the present invention by the recognition of face experiment: the recognition of face experiment is carried out on Yale B face database of expanding and CMU PIE face database respectively.The masterplate coupling is adopted in recognition methods, and used sorter is a nearest neighbor classifier, and the similarity of facial image is described and selected the cosine correlativity for use.The Yale B face database of expansion amounts to 38 people, and everyone has 65 full faces under the different illumination conditions, and all images is divided into 5 subclass according to the intensity of variation of illumination.In the experiment everyone is only chosen 1 picture under the illumination condition of front as masterplate, (image of Set2～Set5) is as test with the 2nd～5 subclass.The CMU face database amounts to 68 people, and everyone has 21 full faces under the illumination condition not of the same race, in the experiment everyone is also only chosen 1 front illumination condition hypograph as masterplate, and all the other 20 images are as test.Carry out corresponding discrimination of recognition of face such as table 1 with the image after the various photo-irradiation treatment methods processing.Wherein LTV just is based on the illumination recovery algorithms of LTV model, promptly decomposes the small scale features that obtains with the LTV model and discerns; LOG-DCT then directly carries out the human face light normalization to former figure with the LOG-DCT algorithm.

Table 1 face recognition result relatively

Consider the face database at CMU, method of the present invention reaches identical discrimination with the LTV model method, further, provides the ROC curve (Fig. 8) of two kinds of methods on the CMU face database here for comparing.As can be seen, under the same error receptance, all than the height of LTV method, this shows that method of the present invention is better than the performance of LTV model method aspect recognition of face to the discrimination of the inventive method from the ROC curve.

From experimental result as can be seen, adopt the present invention to carry out the illumination normalization, in each illumination type subclass of YaleB face database expanded and CMU face database, all can improve discrimination significantly.Can draw from recognition result, better than the effect of carrying out the illumination normalization on former figure in the enterprising normalizing face light effect of large-scale characteristics image, in addition, the people is bold and exists the recognition of face Useful Information in the scale feature image, should not be dropped.

Claims (3)

1. In the method for the enterprising normalizing face light of feature image with different size, it is characterized in that 1, its step comprises:

   1) cutting of aliging is handled to people's face gray level image of input, is divided into for three steps to carry out:

   (1) to every facial image, three unique points of this people's face of first detection and location: the pupil center's point of two eyes and the central point of face;

   (2) make two eyes of every people's face be horizontal by rotation;

   (3) the two interpolation algorithm stretching images of utilization make these three unique points be positioned at the fixed position of image, at last image are cut to unified size;

   2) with the LTV model facial image is decomposed: the facial image I to after every cutting alignment, carry out log-transformation to it

   f(x，y)＝logI(x，y) (1)，

   Find the solution following variation model:

   $u=\arg \underset{u}{\min }\∫|\▿u|+\mathrm{\λ}{\left|\right|f-u\left|\right|}_{L^1}\mathrm{dx}---\left(2\right)$

   And

   v＝f-u (3)，

   Obtain the large-scale characteristics image S and the small scale features image ρ of people's face this moment:

   S＝exp(u)，ρ＝exp(v) (4)；

   3) to the minimum value filtering of small scale features image ρ with threshold value;

   4) remove some low frequency discrete cosine transforms (DCT) coefficient of large-scale characteristics image at log-domain: promptly u is carried out DCT, note DCT coefficient be C (α, β), α=0,1,, M, β=0,1,, N, wherein M, N are the length of facial image and wide, then with n around the frequency domain initial point ²Individual DCT coefficient is set to 0,

   C(α，β)＝0，α＝0，1，…，n，β＝0，1，…，n (5)，

   Obtain by anti-dct transform

   

   , carry out the large-scale characteristics image S after the illumination normalization is obtained in exponential transform at last _Norm:

   $S_{\mathrm{morm}}(x,y)=\exp \hat{u}(x,y)---\left(6\right);$

   5) facial image after the synthetic illumination normalization:

   I _norm(x，y)＝ρ′(x，y)S _norm(x，y) (7)。
2. 2, the method at the enterprising normalizing face light of feature image with different size according to claim 1 is characterized in that the unified size of (3) in the described step 1) in the step is specially 100 * 100 pixel sizes.
3. 3, the method at the enterprising normalizing face light of feature image with different size according to claim 1, it is characterized in that, described step 3) adopts is that 3 * 3 filter window carries out minimum value filtering to ρ, if promptly the gray-scale value of current window central point is greater than assign thresholds, the minimum gradation value of then getting current window replaces the gray-scale value of center pixel.

Images