Diversity in multiple classifier ensembles based on binary feature quantisation with application to face recognition

In this paper we present two methods to create multiple classifier systems based on an initial transformation of the original features to the binary domain and subsequent decompositions (quantisation). Both methods are generally applicable although in this work they are applied to grey-scale pixel values of facial images which form the original feature domain. We further investigate the issue of diversity within the generated ensembles of classifiers which emerges as an important concept in classifier fusion and propose a formal definition based on statistically independent classifiers using the kappa statistic to quantitatively assess it. Results show that our methods outperform a number of alternative algorithms applied on the same dataset, while our analysis indicates that diversity among the classifiers in a combination scheme is not sufficient to guarantee performance improvements. Rather, some type of trade off seems to be necessary between participant classifiers' accuracy and ensemble diversity in order to achieve maximum recognition gains.