hide
Free keywords:
-
Abstract:
Color is an important feature for searching large databases of images, since it is invariant
with respect to camera position, object orientation and size, and partial occlusion. There
are currently many color-based image indexing systems (e.g. Flickner et al., 1995), which
all basically work by building color histograms in RGB space. Our goal was to construct
a color-indexing system based on the known properties of the human color vision system.
Our images were chosen from a large commercially available (COREL) image database
consisting of 60.000 digitized photographs. For each image, we buildt a color histogram
by converting the RGB triplets for each pixel into color-opponent coordinates. These luminance,
red-green, and yellow-blue coordinates correspond to the color directions found
in human color vision (Krauskopf, Williams Heeley, 1982). Luminance was averaged
for each color value, and the resulting color circle was split into 127 segments. The categories
were constructed so that the number of hue categories increased with increasing
saturation. Six different rings were used for saturation, with the radius doubling as saturation
increases. Thus, there is little discrimination of hues for unsaturated colors, whereas
there are 64 different hues at the most saturated level. Two different histograms were
built, one using the frequencies with which the different colors occurred, and another one
that used the average luminance level of each color segment.
We used a query-by-example stragey for searching. Several different distance measures
were evaluated by asking human observers to make similarity judgments. For most images,
this search, based on color only, results in images that are perceptually and often semantically
similar to the target image.
