Rock Identification with Convolutional Neural Networks: A Case Study

Abstract

Rocks may have different external appearances due to internal factors such as minerals and external factors such as pressure and temperature that are effective in the formation process. Rocks of the same type may be visually different from each other due to reasons in the formation process. It may be difficult to distinguish rocks with the human eye due to differences such as color, gap sizes, and stains. In this study, 3 different rock types (tuff, granite, basalt) that are visually dissimilar were classified with deep learning. As the first step for classification, a data set was prepared for training and test images. Training and test images were created by imaging rock samples with different sizes, colors, and shapes twice with different poses. In the study, images with different features were obtained by applying two different pre-processing methods. These are, instead of reducing the size of the full rock image, dividing the image into square sections of 224x224 pixels and cropping the full image from the edges so that only rocky regions are in the image. After the pre-processing, the rocks were detected by applying GoogleNet and ResNet deep network models. As a result of experimental studies conducted with a total of 504 rock images, approximately 94 % (GoogleNet) and 96 % (ResNet) correct classification rates were obtained. This study has shown that image processing-based automatic rock classification applications will make significant contributions to future studies in the field of geology and mining.