Pulse Coupled Neural Networks:an exploration of parameterisation methods

Abstract

Pulse coupled neural networks (PCNNs) comprise a family of biologically motivated models originally developed to replicate the phase-synchronised pulsing behaviour observed amongst collections of neurons in the mammalian visual cortex. They have been applied to a number of applications within the image processing field: most notably image smoothing and segmentation. PCNNs are complex dynamical models with a number of adjustable parameters of reciprocal influence. As a result their behaviour is difficult to accurately predict, control or analyse. This paper follows the development and analysis of a number of parameterisation methods for the PCNN aimed at making it a more powerful and reliable image segmentation model. Experimental results are used to examine the strengths of each of these methods relative to one another in both qualitative and quantitative terms. An energy function formalism for a sub-class of the PCNN family is then proposed and analysed and a Bayesian interpretation is offered.