It was mentioned in [15] that one of the major problems
of K-SOMs, in general, is in selecting and tuning annealing
schemes, i.e. changing of learning rates, both for winner
neurons and their neighbours. This must be done empirically
in the absence of a firm theoretical basis. With respect
to the previously proposed hardware implementations of
K-SOM quantizers, an integer arithmetic has mainly been
chosen for system realization in order to keep the utilization
of the internal hardware resources to a minimum. As a
result, the annealing schemes were inevitably limited to
rely on a linearly decaying function. A codebook consisting
of a group of neural cells which is an intermediate result
from K-SOM quantizers was also represented by an integer
basis. This could significantly affect the quality of the final
image since the processes of rounding and truncating have
been accumulated iteratively. Unfortunately, this problem
has not been experimentally studied in depth. In addition, if
it is evident that the data representation of a codebook has
an impact on the compressed image quality, an alternative
suitable data representation of a codebook is worth
pursuing.