assignment of pixel s to colors and the grouping of optimal colors in the image also
facilitates this. The main objective is to make nearby colors map into a common bin
so that they a re all treated as perceptually similar for color matching and retrieval.
The loss in detail of the color information will not affect the matching process dras-
tically. The number of colors may be increased if finer matching is required depend-
ing on the application dom ain.
Many approaches exist for color quantization that include vector quantization,
clustering [21,47] and neural networks [9]. The entire RGB color space can be repre-
sented using a smaller set of color categories that are perceptual to humans. A pos-
sible theoretical justification for this is provided by Corridoni et al. [47], who present
a system supporting image retrieval by high-level chromatic contents, which are
meant as sensations that color accordances generate on the observer. Images are ar-
chived by describing the spatial arrangement of regions with homogeneous chro-
matic attributes. They choose a color space which replaces the original 3D color
space by a discrete set of repres entative colors. A final set of 180 reference colors
is used in their work.
Cluster-based approach has the advantage that, if we apply it to all or at least
some representative images in the database, the clustering process will take into ac-
count the color distribution of images over the entire database. This minimizes the
likelihood of color bins in which none or very few pixels fall, thus resulting in a very
efficient color quantization. A color-look-up table can be used to store the reduced
space and then colors of original images may be mapp ed into this reduced color
space thereby speeding up the overall process. Such an approach has been followed
by Kankanahalli et al. [21], wher ein they use a set of 27 discrete colors to map the
entire color space, such that all the colors in the application are covered perceptually.
An unsupervised learning algorithm based on clustering is used to deduce from the
representative sample of images the table of colors.
Syeda-Mahmood [43] proposes a color region-based retrieval system in which per-
ceptual color categories are used. A color-look-up table using 220 different colors is
made use of to coarsely describe the colors of regions and later used to index the im-
ages. This perceived color of a region specifies the color category corresponding to
the dominant color of the region.
According to the MPEG-7 visual feature descriptor language, users can perform
tasks on images to define objects, including color patches or textures, and get, in re-
turn, example images from which one can select regions of interest. Colors in a re-
gion are clustered into small number of representative colors and along with their
percentages, spatial coherency, and variance form the descriptor [42].
Following [21,43,47], and as per the recent MPEG-7 standar dization directives for
visual descriptors [42], we too use a color quantization in RGB space using 25 per-
ceptual color categories to segment images based on dominant colors. This is justi-
fied since these 25 colors are sufficient to clearly distinguish all the colored regions in
our chosen domain of image database. From the segmented image we find the en-
closing minimum bounding rectangle (MBR) of the region, its location, number of
regions in the image, etc., and all these are stored in a metafile for further use in
the construction of an image index using a hash data structure.
B.G. Prasad et al. / Computer Vision and Image Understanding 94 (2004) 193–233 201