Evolution of efficient real-time non-linear image filters for FPGAs

Non-linear image filter, noise removal, Cartesian genetic programming,
evolutionary design, digital circuit

Image processing represents a research field in which high-quality solutions have
been obtained using various soft computing techniques. Evolutionary algorithms
constitute a class of stochastic search methods that are applicable in both
optimization and design tasks. In the area of circuit design Cartesian Genetic
Programming has often been utilized in combination with an algorithm of
Evolutionary Strategy. Digital image filters represent a specific class of
circuits whose design can be performed by means of this approach. Switching
filters are advanced non-linear filtering techniques in which the main idea is to
detect and filter the noise pixels while keeping the uncorrupted pixels unchanged
in order to increase the quality of the resulting image. The aim of this article
is to present a robust design technique based on Cartesian Genetic Programming
for the automatic synthesis of switching image filters intended for real-time
processing applications. The robustness of the proposed evolutionary approach is
evaluated using four design problems including the removal of salt and pepper
noise, random shot noise, impulse burst noise and impulse burst noise combined
with random shot noise. An extensive evaluation is performed in order to compare
the properties of the evolved switching filters with the best conventional
solutions. The evaluation has shown that the evolved switching filters exhibit
a very good trade off between the quality of filtering and the implementation
cost in field programmable gate arrays.