where the x
r
y
r
are the vector descriptors of the image
region ROI
r
, W
r
are the weights associated to each
region and K is a normalizing factor.
The recognition score is computed by inspecting
the nearest items in the feature space. Let be d(n),
d(n) ∈ [0, 1], the normalized distance between the
query image and the n-th nearest item in the database
and s(n) the similarity defined as s(n)=(1− d(n)).
The recognition score obtained by class c (i.e. by a
specific museum painting) is defined as:
S(c)=
n
c
w(n
c
)s(n
c
) (3)
where n
c
is the ranking in the nearest neighbor list of
class c objects, and w(n) is a tunable weight function.
If only the nearest item in the database is considered,
the weight function is defined as w(n)=δ
1n
.
Once the recognition score is computed, the visual
recognition engine returns to the PDA the sorted list
of pointing identifiers and scores of the most relevant
hypothesis. If the score of the first hypothesis is above
a given confidence threshold the presentation corre-
sponding to the guessed painting is shown, otherwise
the client notifies the rejection to the user. The confi-
dence threshold may be tuned on the client side.
2.2 Preprocessing
The shots of museum paintings acquired by the palm-
top camera are of poor quality and characterized by
a low contrast due to limited dynamic range of the
sensor (a common problem in low cost cameras with
CMOS sensors) and poor light condition (in Figure
5(a) the graylevel intensity histogram of a sample im-
age is depicted). Moreover the paintings themselves
lack saturated colors, making the color information,
in most cases, unreliable making preprocessing stage
necessary. In order to normalize and increase the dy-
namic range of the pictures, color and intensity equal-
ization algorithms are employed.
Among the many color equalization algorithm de-
veloped do far, the two most widely used are: Gray
World (GW) (Buchsbaum, 1980) and White Patch
(WP) (Funt and Cardei, 1994). These two models
are considered alternatives to each other in methods
of color correction.
Both models try to emulate two human visual adap-
tation mechanisms: lightness constancy and color
constancy. The Gray World approach is typical of
the lightness constancy adaptation because it modifies
the dynamic range of the histogram, assuming that
the average world is gray i.e. assumes that the av-
erage of the surface reflectance over the entire scene
is gray. Alternatively, the White Patch approach is
typical of the color constancy adaptation, searching
for the lightest patch to be used as a white reference
Figure 4: Image processing flow: original image, normal-
ized image with the description grid superimposed, detected
overexposed areas (seed regions are green and the detected
blooming area red), features regions weights used in the
comparison.
similar to how the human visual system does. The
human vision system mechanism is also highly non
linear, since it can be global and local at the same
time. Among the models that compute local color
adaptation using spatial relation and image content we
can consider Land’s Retinex theory (Land, 1977). A
recent approach, called Autmatic Color Equalization
(ACE), merges the Retinex model and the GW model,
performing simultaneously global and local filtering
(Rizzi et al., 2002).
Even if local adaptive methods give the best results,
they are computational demanding for real time ap-
plications, therefore we moved to a simple and effi-
cient approach based on a variant of the GW algo-
rithm. A contrast stretching transformation was con-
sidered: the image is normalized using a piecewise
linear function whose control points are determined
by inspecting the original histogram and computing
the expected gray point as in the GW method. The
normalization method enhances the contrast of a color
image by adjusting the pixels color to span as much
as possible the entire range of available colors. The
histogram tails are cut locating the histogram bound-
aries: 0.1 percent in the black range and 0.5 percent in
VISAPP 2006 - IMAGE UNDERSTANDING
140