Digital Signal Processing Reference
In-Depth Information
1(
QT Ixy i
∗
[ ,
])
=
fxyi
(, ,)
=
(3)
0
otherwise
Where Q denotes quantized process of HSV color space and
T
is conversion from
RGB to HSV. Given a query image
I
and a database image
I
,
Measure dis-
tance
d
between the normalization color histograms of the two quantized image is
calculated as formula (4).
h
(
I
)
−
h
(
I
)
n
=
d
(
I
,
I
)
=
1
i
2
i
(4)
h
1
2
1
+
h
(
I
)
+
h
(
I
)
i
1
1
i
2
i
Ih
stand for the frequency of color value
i
in
I
and
I
respec-
tively . Traditional distance of color histogram is only according to a single statistical
index of pixel frequency of the same color value and does not consider importance
difference of pixel position and the spatial relationship (scattered or concentrated)of
colors in the image. So it's ability to distinguish images is very limited and two image
with different content may have the same or similar color histogram.
h
(
1
i
I
)
(
2
i
)
and
2.2
Color Contrast Calculation
For an image, not all part of it is important, some parts of the image are more promi-
nent than other parts because they can quickly arouse the attention of the observers,
which reflect the high-level semantic concept of salience object in the image. so in
this paper all pixels are weighed by the color attention measure, and the weighted
histogram which can imply pixel position importance are cumulate.
Itti et al. (1998) have built a computational model of saliency-based spatial attention
derived from a biologically plausible architecture. For a color input image, the model
computes a saliency map from maps for color, luminance, and orientation contrasts at
different scales. But this saliency map is 1/256 size of the original image and the
object in the original image is usually separated into some disjointed parts in the sa-
liency map. These results are disadvantage for subsequent image processing. Vision
attention value is simulated as local contrast of an image region with respect to its
neighborhood at various scales [11]. So the color contrast
Sxy
,)
for a pixel at posi-
tion
(, )
xy
in the image is determined as follows:
1
Sxy D xy D xy D xy
(, )
=
[
(, )
+
(, )
+
(, )]
w
w
w
3
(5)
2
4
8
Dxy dcxymxy
(, )
=
((, ), (, ))
t
t
t
Dxy
is a Euclidend distance between the pixel at position
(, )
(, )
xy
and the aver
age vectors of
tt
xy
,
Since perceptual difference in
CIELab color space is approximately Euclidian, so CIELab color space is
×
neighbor region around
(, )
