Pattern recognition at different scales: a statistical perspective

In this paper we borrow concepts from Information Theory and Statistical Mechanics to perform a pattern recognition procedure on a set of x-ray hazelnut images. We identify two relevant statistical scales, whose ratio affects the performance of a machine learning algorithm based on statistical observables, and discuss the dependence of such scales on the image resolution. Finally, by averaging the performance of a Support Vector Machines algorithm over a set of training samples, we numerically verify the predicted onset of an optimal scale of resolution, at which the pattern recognition is favoured.