Noise-Agnostic Self-Adaptive Evolvable Hardware for Real Time Video Filtering Applications

Abstract

Evolvable hardware is a hot topic in digital electronics. As more complex digital circuits are required for specific applications, the design of said circuits becomes more and more complicated. A solution to this problem is making circuits capable themselves of changing depending on the conditions under which they have to work. In order to do this, an optimization algorithm must find the optimal solution to a particular problem. Evolvable hardware is an approach to this strategy which uses an evolutionary algorithm as the optimization algorithm.

An example of an application of evolvable hardware is a digital signal processor for removing a noise signal from an image. This is a complex task that can have variable requirements, which may not be known when the hardware is designed. Therefore, the system must allow being reconfigured when the working conditions change, which implies the usage of reconfigurable hardware. Furthermore, the complexity of the problem makes it complicated to design the processor functionality, especially if it is intended to do this in an automated manner, which leads to replacing the systematic design of said functionality with an optimization algorithm. Thus, evolvable hardware is a good option for such an application.

One of the problems of evolvable systems is that they need to be trained. This is often done by supplying a training set of data which models the conditions in which the system will operate. However, the generation of this set of data is often done offline, thus reducing the system autonomy.

In this work, a solution for this problem is provided, implemented, and analyzed, together with the obtained results. It will be shown that the system will be capable of evolving without such training reference and, additionally, not being aware of what noise type and levels are present in the image.