Face Recognition Using Contourlet Transform and Multidirectional Illumination from a Computer Screen - Advanced Concepts for Intelligent Vision Systems - page 333

convex Lambertian surfaces can be approximated by a nine dimensional linear
subspace. According to Lee et al.'s interpretation [13], there exist nine universal
virtual lighting conditions such that the images under these illuminations are
sucient to approximate its illumination cone. Lee et al. [13] showed that a
linear subspace can be constructed from nine physical lighting conditions that
provides a good representation for illumination invariant face recognition. With
nine physical lighting directions, the need for 3D face construction and albedo
required by [9][2] can be avoided. However, some of the light source directions
suggested in [13] are at angles greater than 100 degrees. Distant light sources at
such angles are dicult to achieve in practical situations due to space limitations.
Another diculty with point light sources is that they must be of significantly
high intensity. Schechner et al. [18] showed that images under multiplexed illu-
mination of a collection of point light sources can solve this problem by offering
better signal to noise ratio. The results of Lee et al. [13] suggest that the super-
position of images under different point source lighting or images with a strong
ambient component are more effective for face recognition. These findings natu-
rally hint towards studying face recognition under extended light sources which
is the focus of our research. In this paper, we try to answer the question: Is
it possible to construct a subspace representation of the face for illumination
invariant face recognition using extended light sources? Besides, minimizing the
need for space, the proposed face recognition algorithm is designed with the fol-
lowing practical constraints. (1) Use of desktop/oce equipment and no custom
hardware. (2) Minimization of the number of training images. (3) Minimization
of representation/memory requirements.
Unlike distant point light source, extended light source implies that it will
not essentially form a constant vector towards all points on the face. Thus stan-
dard photometric stereo techniques cannot be used in this case and neither can
the illumination cone be estimated. However, on the bright side, extended light
sources can be placed close to the face alleviating the need for large space and
high brightness. In our setup, illumination is varied by scanning a horizontal and
then a vertical white stripe (with black background) on the computer screen in
front of the subject. Fig. 1 shows an illustration of our approach. The Contourlet
coecients [8] of the images at different scales and orientations are projected
separately to PCA subspaces and then stacked to form a feature vector. These
features are projected once again to a linear subspace and used for classification.
Our setup was initially proposed in [17] where we used 47 images. In this
paper, we drop the number of images to 23 because adjacent images had quite
similar illumination in [17]. In [17], we constructed two global space-time rep-
resentations using multiple images per face and sliding windows to match the
two representations to the database separately. In this paper, a single image per
face is used to construct its spatial representation and multiple representations
are used to construct a subspace for training a classifier. Hence, recognition is
performed using a single image. The database has been increased from 10 to 106
subjects and comparison with other techniques is performed on the extended
Yale B and CMU-PIE databases.