Automatic image analysis applied to the recognition of quartz surface microtextures using neural network

Abstract

Microtextures imprinted on the surface of quartz grains provide in-depth information on the environmental conditions and sedimentary processes that affected the study sediments. Microtextural analyses are therefore widely used in the provenance studies of sediments. In order to minimize the subjectivity of microtextural recognition, we propose a new software, called DeepGrain (source codes are available at https://github.com/deepgrains/deepgrain), for the automatic identification of microtextures on the surface of quartz grains using the DeepLabV3 model with applied improving techniques. The approach provides an accuracy of 99 % of the area of the tested grains and 63 % of the mechanical features on the surfaces of the tested grains. The inference of a single SEM image of quartz grain took an average of 3.10 sec, leading to a significant reduction in the analysis time of a single grain.