Rock Fragmentation Revealed by a Computer Vision Algorithm

Abstract

Features of rock deformation and destruction can be traced at different scale levels. Detecting peculiarities of rock destruction under intense deformation is essential for understanding the patterns of rock mass evolution. Here we propose a method of segmentation of images of petrographic thin sections and detection of intact areas and grains to identify microstructural properties of rocks. The segmentation method is based on the combination of a special technique of microstructural analysis (STMA) developed at IGEM RAS and the Richer convolutional features (RCF) multilayer neural network. Estimating the error of determining the size of segments due to a false detection of lineaments (STMA algorithm) and inaccuracy of edge detection (RCF algorithm) was performed basing on the Monte Carlo simulation. The method was used to segment 234 thin sections of rocks making up the central part of Primorsky fault of the Baikal Rift Zone and representing different types of tectonites. Analysis of segmented images showed that at scales from 10^–5 to 10^–2 m, in 44% of cases, the rock structure obeys a lognormal distribution of the areas of intact segments, and in 3% of cases, a power distribution. The Weibull distribution does not describe the statistics of the areas of intact segments. The result indicates that fragmentation of rocks is not a scale invariant process.