ABSTRACT
The optical observation of rock thin section images leads to rock classification, which requires expertise in mineralogy. To overcome the subjectiveness and uncertainty inherent in conventional rock classification practice, this study introduces a deep learning-based classification procedure with a convolutional neural network (CNN). The microscopic images of six igneous rock types were prepared and processed for network model training. Transfer learning was implemented to train the ResNet152 model whose performance was investigated with classification methods at patch and image-levels. The trained model achieved a high accuracy of more than 90%. The results also showed that image-level classification from patches demonstrated the best performance depending on the spatial characteristics of comprising minerals. Based on this approach, it was shown that the CNN model can effectively identify the features in microscopic rock images, leading to reliable rock classification.
