Content-Based Image Retrieval with Cellular Automata - Cellular Automata in Image Processing and Geometry - page 148

Colour histogram analysis can be effectively combined with CA models for im-
age retrieval [2, 18]. The chosen colour model has an influence on the results of the
colour analysis. The well-known standard RGB model simply takes the red, green
and blue colour components in a stored image, and any pixel then has a colour con-
structed as a combination of the red, green and blue values in a cube. In image
analysis, colour models such as HSI (hue, saturation and intensity) is preferred, as
it is nearer to the human visual colour perception system. Note that conversions
between different colour models can be applied, so that the best colour model for
a given application can be used. The L*a*b* colour model, for example, has the
property that perception of colour differences are uniform. Here, the L* represents
luminance, the a* refers to a range describing the relative green-red aspect of the
image, and b* refers to the relative blue-yellow aspect of the image. The reader
may consult any standard text book on image analysis (such as [26]) for an in-depth
discussion of colour models.
A common problem in image analysis is that images that are similar in colour, but
with small spatial scene variations, may produce histograms that differ by a large
amount. One of the contributing reasons for this phenomenon is that histograms are
constructed globally from the whole image. CA approaches counter this effect, as
CA implementations stress the local properties in an image without having to re-
sort to the computational and space overhead of a large number of local histograms.
For example, Konstantinidis et al. [18] proposed a pre-classification of the image
database, based on the use of CA and the L*a*b* colour space. Then, a histogram
based on the hue in the HSV (hue, saturation, value) colour model is used for com-
parison purposes.
Konstantinidis et al. apply two CA with a Moore neighbourhood, one on the a*
component and one on the b* component, for each image in the database. The up-
date rule replaces the current pixel value of a cell with the sum of the pixels in the
neighbourhood for five evolutions, or until the maximum absolute values in the a*
and b* range is reached (that is, ). As the CA evolves through its

time steps, the number of pixels that change in each evolution is recorded. These
are entered into the bins of the histogram, so that the final histogram contains the
number of changing pixels for each time step. Hence, the image database has two
histogram values for each image. In the retrieval step, the histogram of the query im-
age ( H Q ) is compared to those in the image database ( H C ) using the Bhattacharyya
distance

|− 128 | or | 127 |

∑ i H Q (

.
The hundred best matches from the database is then considered as having been
preclassified as possible matches. Finally, an HSV histogram colour comparison is
done between the query image and the preclassified subset of the database, to find
the final best match(es).
Although not directly related to image retrieval, it is interesting to note that cel-
lular automata are also useful for colourisation [19] and colour reduction [23].

− ln i ) × H C ( i )