Basic Competitive Neural Networks as Adaptive Mechanisms for Non-Stationary Colour Quantisation

Abstract

In this paper we consider the application of two basic Competitive Neural Networks (CNN) to the adaptive computation of colour representatives on image sequences that show non-stationary distributions of pixel colours. The tested algorithms are the Simple Competitive Learning (SCL) algorithm and the Frequency-Sensitive Competitive Learning (FSCL) algorithm. Both, SCL and FCSL are the simplest adaptive methods based, respectively, on minimising the distortion and on the search for a uniform quantisation. The aim of this paper is to study several computational properties of these methods when applied to non-stationary clustering as adaptive vector quantisation algorithms. Non-stationary colour quantisation is, therefore, representative of the more general class of non-station-ary clustering problems. We expect our results to be meaningful for other algorithms that involve either the minimisation of the distortion or the search for uniform quantisers. We study experimentally the effect of the size of the image sample employed in the one-pass adaptation, their robustness to initial conditions, and the effect of local versus global scheduling of the learning rate.