Comparison of model initialization methods in machine learning for thin-section rock image classification

Le résumé fourni par la source

Microscopic rock image analysis aids geotechnical and geological studies, often with computational methods. The growing availability of image data has led to the widespread adoption of automation in image analysis. However, the lack of large, publicly available datasets has hindered the development of dedicated machine learning models for geological applications. This study explores the use of transfer learning techniques to overcome this limitation by leveraging pre-trained machine learning models for rock type classification based on thin-section images. The research compares models trained from scratch with those utilizing pre-trained architectures to assess whether models trained on non-geological data can effectively support rock classification. The experiments were conducted using a dataset comprising 11901 microscopic images representing 40 rock types. The study evaluates different model initialization methods to assess their performance in geological applications. The results indicate that transfer learning enhances classification accuracy compared to models trained from scratch, demonstrating the potential in geoscientific research. This work provides insights into the practical application of machine learning in rock classification and serves as a reference and survey for specialists in geotechnics and geology seeking to integrate artificial intelligence into their research.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.