Image Enhancement Technique using Adaptive

Multiscale Retinex for Face Recognition Systems

Khairul Anuar Ishak

, Salina Abdul Samad

M. A. Hannan

and Maizura Mohd Sani

Dept. of Electrical, Electronics and Systems Engineering

Faculty of Engineering and Built Environment, University Kebangsaan Malaysia

43600, UKM Bangi, Selangor, Malaysia

Institute of Microengineering and Nanoelectronics, University Kebangsaan Malaysia

43600, UKM Bangi, Selangor, Malaysia

Abstract. Various illumination effects in an image are one of the states of

difficulty that should be solved in order to get a satisfactory result in face

recognition task. The inhomogeneous intensities of the image has led to many

plans and algorithms to devastate the cause and next to eliminate the

illumination. The focus of this paper is to enhance the image by reducing

illumination effects; employing a preprocessing step i.e. adaptive multiscale

retinex as the illumination correction method before accomplishing the

recognition task. The performance of this method is evaluated using the Yale

database and has lower equal error rate compared with single scale retinex and

conventional multiscale retinex.

1 Introduction

In face recognition, usually there are some inconsistencies between the real scenes

and the training set images. One of them is illumination variations such as shadow,

blur, dark and noise occurring in the images. Sometimes this can cause degradation in

the algorithm to recognize the face image. In this paper, we want to reduce the

unwanted effects in face images by applying adaptive multiscale retinex as a

preprocessing step. Multiscale retinex was initially used to provide stability in color

images; however it is also competent to be used in gray scale images.

Lightness and color uniformity refer to wide range of intensity and spectral

illumination variations [1]. Multiscale retinex is formed from the retinex theory by

Edwin Land [2]. Land proposed the idea of retinex as a model of lightness to measure

the lightness response in an image.

However Land did not apply the model to image enhancement algorithm, but this

is done by Jobson where they define the properties of the surround/center retinex

function [3]. The characteristic they describe is single scale retinex when they

performed logarithmic after the surround function. They also apply ‘canonical’ gain

offset to the retinex output to clip certain parts of the highest and lowest signal

excursion. However, single scale retinex can either provide dynamic range

Ishak K., Abdul Samad S., Hannan M. and Mohd Sani M. (2009).

Image Enhancement Technique using Adaptive Multiscale Retinex for Face Recognition Systems.

In Proceedings of the 5th International Workshop on Artiﬁcial Neural Networks and Intelligent Information Processing, pages 43-49

DOI: 10.5220/0002262600430049

 SciTePress

compression on small scale, or tonal rendition for large scale image. This limitation

expands single scale retinex to a more balanced method that is multiscale retinex.

After that, another characteristic of multiscale retinex were found where other than

dynamic range compression, multiscale retinex purposes are to replicate tone in an

image in order to reduce its dependencies in lighting conditions and improved spatial

resolution[4]. So we use this characteristic in this paper to lessen illumination effects

in order to obtain controlled lighting condition in a face image. We modify the present

multiscale retinex by including histogram shifting and adaptive histogram

equalization to the original algorithm as to have a more uniform face image contrast

than the original method. The detail of this proposed method will be discussed in

section 3. Before that, section 2 will cover the original theory of multiscale retinex.

Section 4 will describe the experimental results and lastly is the conclusion in Section

2 The Original Multiscale Retinex

The original multiscale retinex essentially measure the intensity of an image and

estimate the illumination from the proportion of the local image mean intensity value.

By applying Gaussian filter, the image is smoothed at different weight and size in

order to find the mean of the image. To obtain the retinex output, the filtered image is

divided using the illuminated image (input). Then, logarithmic function is done to

compress dynamic range of images with large variations in pixel value [5] before the

image is reconstructed again using additive function.

The original multiscale retinex algorithm is obtained from single scale retinex [4]

as in (1):

R(x,y)=logI(x,y)-log[F(x,y)*I(x,y)] (1)

where I(x,y) is input image, R(x,y) is retinex output, F(x,y) is the Gaussian surround

function. Symbol * denotes convolution. Gaussian surround function is given by:



x,y



=K.e

-(x

)/c

(2)

where c is Gaussian shaped surrounding space constant. The value of cis related to

visual angle in the direct observation which is determined through experiment. K is

selected such that:





x,y



dxdy=1 (3)

Until this stage the single scale retinex would only provide tone reproduction and

dynamic range compression at certain scale in an image. The image would have only

one of the important characteristics. Thus, to overcome this limitation, superposition

of different scale at certain weight would solve this problem as shown in (4), where

Nis number of scale, where R

is different scale of single scale retinex. ω

is the

weight of each single scale retinex with equal value.

MSRi

= ω

n=1

(4)

3 The Adaptive Multiscale Retinex

After applying the original multiscale retinex, we found that the image was too

dark. This meant that the image brightness and contrast needed to be altered. Thus we

modified the algorithm by applying a recombination with the original image.

According to [6], a method need to be applied to restore the information in different

regions to smoothen the global contras in the image according to

which region is

darker or brighter. The information here is, different intensity in different regions in

the original picture. For this reason, recombination is needed to restore the

information as in (5):

MSRi =

ω

n=1

+ω

original

·log(original) .

(5)

After recombination with weighted original picture, adjustment is made on the

histogram by performing a constant shift which helps improve the entire global

brightness of the image. To shift the histogram is a simple task, where in the range of

0-255 the image pixels should be. In order to allocate the pixels in the range, we set

the initial maximum pixel (MinVal) as 0, and the minimum pixel (MaxVal) as 255.

Then we evaluate the entire image pixels one by one and update the new value

(NewVal) using (6). Prior to that, every pixel value (PixVal) has to be tested whether

it is higher than the MaxVal or lower than the MinVal. If the value is higher, then the

MaxVal will be the PixVal value and the similarly, MinVal if the value of PixVal is

lower than MinVal. These values are needed to find the new value which is compute

from

NewVal  

Pixval-MinVal

MaxVal-MinVal

255.

(6)

Next, we execute a local image enhancement technique that divides the image into

rectangular blocks. Usually how many blocks should be used is determine through

experiments. First, obtain the cumulative density function of the small region

histogram. Then the centre pixel of the region is equalized (histogram equalization)

and moved to the adjacent pixel in the rectangular region. This process is called

adaptive histogram equalization (AHE) [7].

Overall, the adaptive multiscale retinex algorithm is shown as in figure 1.

Fig. 1. The adaptive multiscale retinex flowchart.

4 Experimental Results

We evaluate the performance of the preprocessing methods using Eigenface [8] as the

feature extractor while Euclidean distance is used for the matching purpose. We also

implement a fusion of Principal Component Analysis (PCA) and Linear Discriminant

Analysis (LDA) [9] for data reduction. The performance is evaluated using well

known benchmarking measures for biometrics system i.e. Equal Error Rate (EER)

[10]. To compute EER, two components must be determined. The first is false

acceptance rate (FAR), when the impostor is falsely regarded as the client. Another

one is false rejection rate (FRR), when the client is falsely regarded as the imposter.

Here the client is the authorize person in the face recognition system. The EER is the

cross-over value where FAR and FRR coincide.

The dataset we use to evaluate the algorithms is Yale [11], which contains 165

grayscale images of 15 individuals. There are 11 images per subject, one per different

facial expression or configuration. The 11 images show various extreme illuminations

and pose criteria. We randomly selected 5 images from each subject to be the training

sample and 5 images of each subject as the testing sample. Our experimental face

condition is cropped face images. The size of all images is standardized to 50x60.

The experiments are done using grayscale images as the inputs. Three methods are

compared, i.e., single scale retinex, multiscale scale retinex and the adaptive

multiscale retinex. Table 1 shows the EER for these methods where we can see that

the EER for the adaptive multiscale retinex is the lowest compared to the original

multiscale retinex and single scale retinex.

Table 1. EER comparison of three methods.

Method EER (%)

Single Scale Retinex 11.87

Multiscale Retinex 10.33

Adaptive Multiscale Retinex 10.27

Histogram

shifting

Original

multiscale retinex

Recombination

with logarithmic

function of

original image

Adaptive

Histogram

Equalization

The EER results correlate with the output images of the three methods. To

illustrate, we choose 4 faces from the database which have been illuminated with

different lighting conditions: cast shadow, attached shadow, specular reflection and

diffuse reflection [12]. The lighting conditions are shown in figure 2. Figure 3

illustrates the outputs for all three methods. For face number 1, only the adaptive

multiscale retinex is able to eliminate the diffuse reflection. For face number 2, there

are specular reflection and strong cast and attached shadow, where the adaptive

algorithm is capable in removing the specular reflection. The condition in face

number 3 is the same as number 1. Face number 4 contains cast shadow, specular and

diffuse reflection. The adaptive multiscale retinex is able to remove all the lighting

distraction on the face image, except the cast shadow.

Fig. 2. Different lighting conditions of a face.

Fig. 3. The image outputs using three methods for person 1, 2, 3, and 4.

Figure 4(a), (b), (c) and (d) show the output histograms for face images 1, 2, 3, and

4 before and after applying the adaptive multiscale retinex. For all the histograms, the

upper figure indicates the histogram before and the lower figure is the one after. All

the histograms show more balanced tone representation in gray scale values. The

histograms centred at the middle show ordinary conditions with peaks and gradually

tapering off on the left and right sides of the histogram. This proves that the method is

able to reduce broad tonal range from the original face image.

Cast shadow

Specular

reflection

Attached

shadow

Diffuse

reflection

Single Scale

Retinex

Ori

inal

Multiscale

Retinex

Adaptive Multiscale

Retinex

Before (a) After

Before (b) After

Before (c) After

Before (d) After

Fig. 4. (a), (b), (c), (d). The histogram before pre-processes (the original image) and after pre-

processed with adaptive multiscale retinex for person 1, 2, 3 and 4.

5 Conclusions

In this paper, an adaptive multiscale retinex algorithm is presented. The purpose is to

remove illumination appearances. This is achieved by modifying the multiscale

retinex algorithm with adaptive histogram equalization and histogram shifting. The

performance of this method is tested using the Yale dataset and shown in terms of

EER rate and output comparisons with single scale retinex and conventional

multiscale retinex.

References

1. Rahman, Z.: Properties of a Center/Surround Retinex Part One: Signal Processing Design.

NASA Contractor Report#198194 (1995).

2. Land, E.: An alternative technique for the computation of the designator in the retinex

theory of color vision. Proceedings of the national Academy of Sciences of the USA,

Vol.83.3078-3080 (1986).

3. J.Jobson,D., Rahman,Z., A.Woodell, G.: A Multiscale Retinex for Bridging the Gap

between Color Images and the Human Observation of Scenes. IEEE Transaction on Image

Processing, Vol. 6(3).965-976 (1997).

4. J.Jobson,D., Rahman,Z., A.Woodell, G.: Properties and Performance of a Center/Surround

Retinex. IEEE Transaction on Image Processing, Vol.6(7).451-462 (1997).

5. Gonzalez,R.C., Woods,R.E.: Digital Image Processing. Addison-Wesley (1992).

6. Herscovitz,M., Yadid-Pecht,O.: A Modified Multi scale Retinex with an improved global

impression of brightness for wide dynamic range pictures. Journal of Machine Vision and

Application,Vol.15.220-228 (2004).

7. Zhiming,W.,Jianhua,T.: A Fast Implementation of Adaptive Histogram Equalization.

Proceedings of The 8th International Conference on Signal Processing (2006).

8. Turk, M. and A.P.Pentland.: Eigenfaces for recognition. J.Cognitive Neurosci, Vol.13.71-

86 (1991).

9. Ishak,K.A., Abdul Samad,S.A., and Hussain,A.: A Face Detection and Recognition for

Intelligent Vehicles. Information Technology Journal,Vol. 5(3).507-515 (2006).

10. Wayman, J.L.: Error rate equations for the general biometric system. IEEE Robotic &

Automation, Vol.6 (3). 35-48 (1999).

11. Georghiades, A.S., Belhumeur, P.N. and Kriegman, D.J.: From Few to Many: Illumination

Cone Models for Face Recognition under Variable Lighting and Pose IEEE Transaction

Pattern Analysis Machine Intelligence,Vol.23.643-660 (2001).

12. Nishiyama,M., Kozakaya,T., and Yamaguchi,O.:Illumination Normalization using Quotient

Image-based Techniques. Recent Advances in Face Recognition. Austria: I-Tech Education

and Publishing (2008).