Sparse image reconstruction by two phase RBM learning: Application to mine planning

A key problem in mine planning is estimating the locations of underground ore bodies from a set of sparse core samples that span the area to be excavated. Data from each sample location are interpreted by a geologist and rendered as an image depicting the local ore distribution. The goal is to reconstruct these sparse samples into a dense image that can correctly account for the underground structure. From a computer vision perspective, this has the form of a sparse data reconstruction problem, and is often tackled using a stochastic reconstruction approach. However in the present case the nature of the data is such that most conventional approaches fall short. In this paper we introduce a stochastic reconstruction method that uses a Restricted Boltzmann Machine (RBM) architecture to solve the problem in a novel way. Specifically, it incorporates a two-phase learning approach that i) uses dense sample information available from already excavated areas of the mine to build a general appearance model, and then ii) conditions this model to account for the data in the core sample images. Reconstruction is then accomplished by sampling the distribution implicit in the in the RBM after learning. Our results show that this approach offers significant improvements to conventional stochastic reconstruction algorithms as the RBM is better able to learn the distribution underlying the sample data.