57
In
this template
𝜆
is a scalar coefficient as a ratio between contrast enhancement and
smoothness level. Mapping any PDEs with parametric
coefficient is possible; therefore
depending
on the situation, a main controller can change these coefficients to adapt the
CNN result. By this way we can pre-calculate templates for a wide range of problems and
use them dynamically in a real-time situation.
There is a hardware based solution also for template calculation.
Theoretically genetic
algorithm is a time consuming optimization technique in any domain of science. There are
two issues that make GA slow. The first issue is the nature of the selection process. This
means that for optimizing a solution we have to perform the objective function to a huge
amount of chromosomes until it converges to the global minimum. Another problem comes
from the performance of the system. In our case (i.e. Finding template for CNN) for
evaluating a quality
of chromosomes as a solution, the system needs to decode the
chromosomes and apply it on the CNN and check the results
according to the objective
function. In [107] D. Balya
et al
did analyse several papers to study techniques of template
calculation based on genetic algorithms and proposed an analogic implementation of the
genetic algorithm based template calculation on FPGA. In genetic
algorithms the fitness
function is a soft computing module and the rest of the system could be implement on any
hybrid or homogenous machine. Figure 5-13 shows the schematic of ideal system for
calculating the CNN templates.
Figure
5-13: Integration of CNN on FPGA with PowerPC
for speeding up of genetic
algorithm
UM-CNN
On FPGA
PowerPC
Fitness Function
DDR
Memory for storing
Chromosomes
GUI and consol for
monitoring results
