A New Transformer Based On Thin Section Rock Images with Self-Attention Module

Abstract

The rocks, as the main component of the earth's crust, form the basis for intra-earth events such as mineralization, determination of oil-natural gas reservoir rock, seismicity and tectonism. The main difference between rocks is their mineral content and conditions under which they form. The mineral types that make up rocks, with their different mineral compositions, textures and structures, can be determined by thin section studies on a microscope or chemical analysis methods. Identifying minerals and naming rocks is a long and tiring process. It requires experience and knowledge. Increasing classification performance in convolutional neural networks in geology, as in other fields of science, has made this approach possible for rock classification. This new visual transformer-based approach proposed for rock types classification constitutes an important alternative to convolutional neural network-based approaches. The proposed approach makes the information in the representation space more specific with the self-attention mechanism and proposes a more robust structure for the complex classification problem. The performance of Proposed ViT-base 16 was applied to the sedimentary, metamorphic and igneous rock data set and a high success of 97.38% accuracy was achieved. ©2024 IEEE.