Neighbour-based MLPs

This paper reviews a Bayesian approach to classifying multi-spectral image data where the pixel labels are assumed to be spatially correlated. A Markov random field (MRF) model is introduced to model localized dependence, so that a label is assumed to be conditional on the labels of the neighbouring pixels only. The multi-layer perceptron model is extended to incorporate the MRF model and it is shown that the posterior distribution of labels given spectral values and neighbouring labels can be maximized by an iterative updating procedure. This updating procedure is in fact an implementation of a (modified) Hopfield network. Finally an example is presented that consists of classifying a remotely sensed image of an agricultural property.