Retinex Algorithm Based on Mean Shift Applied in Face Recognition

2016 International Conference on Wireless Communication and Network Engineering (WCNE 2016) ISBN: 978-1-60595-403-5

Retinex Algorithm Based on Mean-Shift Applied in Face Recognition

Yang LI

and Zhu-ming CAO

1_{Department of Information Engineering, Tianjin Modern Vocational Technology College,}

Tianjin, China 2

School of Mechanical Engineering, Beijing Polytechnic, Beijing, China

keywords: Retinex theory, Illumination pretreatment, Mean-Shift smoothing filtering, Face recognition.

Abstract. In order to improve the face recognition rate with complex illumination conditions changing, a Retinex algorithm based on Mean-Shift smoothing filtering was proposed and applied to illumination pretreatment of face recognition. The original image was enhanced by this method with local nonlinear contrast enhancement, illumination was evaluated by Mean-Shift smoothing filtering instead of Gaussian filter in traditional Retinex and the problem that halo formation couldn't be removed by Retinex algorithm was solved. The result of experiment in CAS_PEAL face library shows that, the algorithm has better illumination pretreatment effect than tradition algorithms, and also can improve the face recognition rate.

Introduction

At present, face recognition is played an important role in many fields, Related algorithm research in the past few years increased significantly, But face recognition is still faced with many unsolved problems. Light is one of the most important factors affecting the quality of face image, Caused by illumination changes in the appearance of the greater than the appearance of the differences between individuals, So in the face recognition under varying illumination, recognition efficiency will drop sharply. In face recognition, face images from different illumination change correctly recognize the face is still a very challenging problem[1-2]. In order to solve the influence of illumination in face recognition, the domestic and foreign researchers have proposed many illumination pretreatment methods. At present, on the whole, these methods can be divided into three categories: The method based on image processing technology[3-4]. This image preprocessing method is a kind of early, the classic methods of histogram equalization and histogram specification and Gamma correction, etc. The advantage of this kind of method is simple, easy to implement, the disadvantage is that when the illumination change is larger for image preprocessing effect is not obvious, can't achieve very good effect on the performance. The method based on illumination invariant features[5-6]. Method that is based on illumination invariant feature in the feature extraction stage use is not sensitive to illumination change characteristics, and this feature is used to extract feature vector of face image recognition, use these same features can greatly reduce the influence of illumination. But the existence of image detail be enhanced at the same time, will produce noise will bring images at the same time reduced the overall visual effect of faults.

Business image is based on light processing method of the model such as light cone. This method is to meet lambert reflection characteristics model. In this model, face images in the texture and shape information is inherent attribute, and external properties is light. Light processing method based on the model is to use the facial texture and shape intrinsic properties, such as to eliminate external properties such as light as much as possible. But the method has the shortcoming of large amount of calculation.

illumination pretreatment algorithm of smoothing filter. The algorithm adopts nonlinear global contrast enhancement in the initial stage, depending on the degree of light and shade image, adjust the brightness of the whole image,; then using Mean - Shift of smoothing effect to remove existing in the traditional Retinex method of "halo", so as to remove the light's influence on the original face image and improve the recognition rate of face recognition.

The Limitations of Traditional Retinex Algorithm

Retinex theory by Edwin Land, put forward in the early 1970 s, a color invariant perception of color theory. Retinex theory pointed out that the irradiation light directly determines the pixels in an image that can reach the dynamic range, and the reflected light is decided to the intrinsic properties of an image. The essence of the Retinex theory is reflection properties of the image from the image. Due to the illuminate of objects in the scene brightness, corresponding to the low frequency part of image, the reflection of the objects in the scene brightness, corresponding to the high frequency part of image, so the traditional gaussian filter Retinex algorithm commonly used to estimate the brightness of the image, thus to eliminate the influence of the irradiation light for object was supposed to look. Mathematical description of Retinex algorithm are as follows:





2 2 2 ( / )

( , ) log ( , ) log ( , ) * ( , )

( , )

( , ) 1 x y c

R x y I x y F x y I x y

F x y K e

F x y dxdy

 

  

 _{ }



 _





A large number of experiments show that the Retinex algorithm can get better effect in light treatment, but there are removal of light at the same time will produce the phenomenon of "halo". This is because the Retinex algorithm in the beginning of design, caused by the assumption of Retinex algorithm itself is the premise of the illumination change in space is relatively smooth, but in the actual face image is not like hypothesis, when the difference in the brightness of the two parts of the image, in the light change big edge flare phenomenon, if serious, will cause the change of face image contrast areas of partial loss of detail.

Based on Mean Shift Retinex Algorithm of Smoothing Filter

Nonlinear Face Image and Local Contrast Enhancement

Before the removal of light weight, first of all to adjust image gray level to increase the image contrast and to enhance the overall brightness of the image that is darker. This paper uses a local nonlinear contrast enhancement method, a calculation method is as follows:

In the first place in a local field of 3 x 3 pixel values, calculation of minimum and maximum

pixel value Smax _pixels Smin _{and its average pixel values} Savg_. Then using non-linear contrast enhancement of formula (2):

( , ) [ ( , ) ][ ]

no line avg n

N _ S x y  S x y S S   

(2)

The Sn _{for local regularization results, can be represented as:}

min

max min

( , )

n

S x y S

S

S S 

 

 

(3)

Mean Shift Smooth Filtering

1 1 ( ) ( )( ) ( ) ( ) ( ) n

H i i i

i n

H i i

i

G x x w x x x M x

G x x w x      





as the kernel function, H says bandwidth matrix, in the application, generally take

bandwidth matrix for the unit matrix form, namely, Hh I2 . So Mean - Shift can be represented as: , , 1 ( ) ( ) n

k d i

h K d i

c x x

f x K

nh  h

 



( ) _{k d} ( )

K x C k x

, k x( ) K x( ) profile function. It fh K, ( )x _{gradient function to}

differential s, namely:

2

2

2 _{, 1}

( )

, 2 ₁ 2

1

n x x_i x g_i

h c_{k d n x x i}

i

f_{h K} x _d g x

h

nh _{i n x x} i g h i  _{ }    _{ }      _{ }_ _ _{ }   _{  } _{  }   _  _{ }_{  }  _    _{ } _  _    _{ }    _{  }     

  ₍₆₎

( ) '( )

g x  k x _{, according to the definition of the kernel function,} G x( )ck d, g x( 2 )

is known, is the type of the second type (5) in the modified forms, namely the Mean - Shift vector, remember

tomk G, ( )x _.

2 1 , 2 1 ( ) n i i i k G n i i x x x g h

m x x

x x g h    _           _       





x _{kernel function, as a representative of the initial point ( )}G x _,  _{said as the allowable error,}

Mean - Shift algorithm can repeat the following steps, until convergence.

① calculate the Mean Shift vector.

② translation of x and mh G, ( )x _.

③ to determine whether satisfy mh G, ( )x x 

if meet the conditions, the convergence; if not satisfied, continue to (1) Cycle.

The Experimental Results and Analysis

In order to verify the validity of the algorithm in this paper, respectively on purify face image lighting effects and the general algorithm in this paper to the promotion of human face recognition rate comparing with the traditional illumination pretreatment algorithm.

Remove the Lighting Effects of Qualitative Analysis

(a) the original face image (b) histogram equalization

effect

(c) Gamma transform effect

(d) effect of SSR (e) effect of MSR (f) the effect of algorithm in this paper

Figure 1. Preprocessing algorithm to remove all kinds of light lighting effects.

As shown in figure 1 (a) original image is affected by the strong light, from the point of the effect of various preprocessing algorithm, histogram equalization (HE) is only the image grayscale have adjusted their homogenization, slant image as a whole is too bright, and at the same time in the removal of light can produce noise. Gamma correction, though up to a certain extent in addition to the light, but to the left side of the face image is darker, there are still some facial features is the shadow. Single scale Retinex (SSR) and multiscale Retinex (MSR) can be enhanced the face image affected by light whole, but local details of the face image processing is not enough ideal, some local produce "halo" phenomenon. Can see from figure 1 (f), the proposed algorithm not only can make the affected by the light of the dark parts of the face image area has been effectively enhanced, highlight the hidden in the dark area, the details of the original image at the same time the light areas remain unchanged, and little "halo" in the image, the image illumination pretreatment effect significantly better than the other four kinds of algorithms.

Face Recognition Experiments

In order to verify the effectiveness of the algorithm is applied to face recognition, in CAS_PEAL face recognition experiments, in CAS_PEAL face library choose Lighting (Lighting) subset of 30 people, each 9 pieces, a total of 270 images in the experiment, each of them 9 images, are standard posture, neutral expression, only each image of the light is different. The experimental results are as follows:

Table 2. Different pretreatment methods on CAS_PEAL libraries face recognition rate.

algorithm No pretreatment

HE algorithm

Gamma correction

SSR algorithm

MSR algorithm

In this paper, algorithm Recognition

rate (%)

50.32 61.21 63.38 70.19 73.83 80.31

From the experimental results, this algorithm after recognition rate is significantly higher than without illumination pretreatment, and higher than that of traditional illumination pretreatment algorithm, demonstrate the algorithm in the pretreatment stage, can effectively remove the illumination of face image, the effect of improving the recognition rate of face recognition.

Conclusion

the quality of face image preprocessing algorithm is superior to the traditional illumination, and the ability to improve under the condition of complex is the human face recognition rate.

References

[1]Zhou S, Aggarwal G, Chellapa R. Appearance characterization of linear lambertian object, Generalized photometric stereo and illumination- Invariant face recognition[J]. IEEE Trans on PAMI, 2007, 29(2): 230-245.

[2]Xie X D, Lam K M. Face recognition under varying illumination based on a 2D face shape model[J]. Pattern Recognition, 2005,38(2): 221-230.

[3]Aricit D, Ikbas S, Altunbasak Y. A histogram modification framework and its application for image contrast enhancement[J]. IEEE Trans on Image Processing, 2009, 18(9):1921-1935.

[4]Savvides M, Kumar V. Illumination normalization using logarithm transforms for face authentication[C]// Proc IAPR AVBPA, 2003:549-556.

[5]Tan X, Triggs B. Enhanced local texture feature sets for face recongnition under difficult lighting conditions[J]. IEEE Trans on Image Processing, 2010, 19(6): 1635-1650.