Work

Fast and lightweight automatic lithology recognition based on efficient vision transformer network

Year: 2025

Type: article

Abstract: Traditional lithology classification methods require the expertise of appraisers and a variety of complex measuring instruments, are easily influenced by staff experience, and are excessively labor-intensive. In order to solve the previous problems, some studies have utilized rock images and intelligent algorithms to automatically recognize rocks. However, the models for automatic lithology recognition rely on more powerful devices and cannot be applied to most edge lightweight devices. To address this problem, we deeply extend our previous research work and, based on the previous training framework, propose an automatic lithology identification method (Rock-ConViT) based on an efficient visual converter, and design and implement the training and application flow framework of the method, which requires less than 720MB of device memory and 1.6GB of graphics memory, with an accuracy rate of 94.75%.

Source: Solid Earth Sciences

DOI: 10.1016/j.sesci.2024.100179