24
one can formulate two different classes of limitations that are: a) the one related to the
software architecture and to the algorithms for manipulating pixels and afterwards
extracting meaningful data; and b) hardware limitations and inflexibility of design. A much
better solution for image processing is the use of a parallel processing architecture.
Multiple Instructions - Multiple Data
streams (MIMD) is the most common architecture for
parallel processing; most modern computers fall into this category [51, 52]. In this model
of processing every processing unit may access to different memory and data streams. For
increasing the performance for reaching a sufficient speedup, dedicated hardware for each
algorithm has been suggested. Hence, designers implement each application on different
platform, and this redundancy in hardware increases the price and complexity of the
system [53, 54].
A direct mapping of algorithms on the hardware is often viewed as the best way of
processing [53, 55, 56]. The only issue that should be considered in this case is the low
flexibility of the system concerning design time. Therefore, the only drawbacks of this
approach are: a) the complexity of mapping many image processing operators/functions
onto the hardware, and b) the limitation of hardware resources.
Hence, we are looking for a reconfigurable model/concept/architecture that can change
(or be reconfigured) into different functions and thereby significantly saving hardware
resources. Another important factor of this model is parallel processing of pixels.
