A Quantitative Representation for the Segmentation of Martian Images

In previous work we have identified the problem of dimensionality associated with learning the data density distributions of textures. In this document we investigate the properties of a family of ways of constructing this space, based upon approaches which have demonstrated merit in the recent literature. We find that for Martian images at least there seem to be advantages to the strategy of encoding textures as 1 bit significant differences between pixel samples in a local region. We suggest that this finding might represent a local optima for these approaches, which maximises spatial context at the expense of grey-level accuracy, and resulting in a representation which is consistent with rank-ordering approaches to patch matching.