A Face Recognition under Varying Illumination - Advances in Image and Graphics Technologies - page 122

 wxy (, ) = ʱ (, ) xy ⇅ ʲ (, ) xy



'

ʱ (, ) xy = g ( (, ), ) ˄ xy h (5)




ʲ (, ) xy =∇ g ( Ixy S (, ), )



In equation (3), L represents illumination and w represents mask. The w consists of
two masks which can regulate its weight at each point by calculating the light discon-
tinuity. So the

Lx + may contain more light in once iteration. In addition, the
( t +1)th L may involve less light than in ( t )th with iteration counting. Consequently, R
may include more light shadow for less light in L , which can be expressed by (2). In
conclude, reflectance will become worse with the iteration increasing.

'( t 1)

(, )

2.2 Face Processing

There are two steps in the processing: first, the retinex should be applied to obtain
invariant; and second, an eye detection is used to align and segment human face in the
reflectance. For the first one, the detail has been expressed in [7]; and in the second
case, there are always useless backgrounds in the image. Besides, the face may tilt
when people is under an uncontrolled environment. Therefore, many researchers will
manually align and segment the images at first. For this reason, we are trying to designs
an auto face processing method in this research paper. We use algorithms [8,9] to locate
eyes positions and use a line to connect the two eyes. Then, we select the center of the
line as reference point and extract an effective square with side length of 1.8 l in the
reflectance ( l is the distance between two eyes. Moreover, the distances of reference
point to left and right side are 0.9 l and 0.9 l . 0.5 l and 1.3 l are the distances to top and
bottom side). The processed face images are illustrated in Fig.1.
As showed in Fig.1 (a), the original image is affected by the illumination and the
background, besides, the face slightly tilt to left. Fig.1 (b) to Fig.1 (d) has the some
results acquired by the method presented above. Obviously, there are not only an ef-
fective field obtained by aligning and segment, but also less illumination left. After
that, the Fig.1 (d) will be used to recognize.

2.3 The Effect of Illumination Element to Face Recognition

According to section 2.1, the quality of invariant is related to iteration t . When t is
small, the shadow is removed effectively in the invariant. When t is lager, the reverse
would be true. So we suppose that a satisfying recognition accuracy should be obtained
with a small t . However, a series of experiments suggest that this opinion is not true.
We select 10 different people with 10 illuminations for each one. Then, the images are
divided into two sets, one and half are used to train and others are used to test. All
images are processed with measure in section 2.2. In the experiment, t increases with 10
and the classifier is the NN (Nearest Neighbor).