Flexible nonlinear contextual classification

We present a framework for flexible nonlinear contextual image classification. The framework integrates classical and recent models for image classification, ranging from a multivariate Gaussian classifier, to MLP neural nets, classification trees and recent regression models based on general additive models, and combines them with a Markov random field for spatial context. The effect of using the different nonlinear discriminant functions is compared with the effect of using an MRF model for spatial context. In general, the use of an MRF model results in larger improvements in classification accuracy than using different nonlinear discriminant functions, but the combination of them can give large improvements.