Rock tomography image retrieval: a deep metric learning approach with N-pair loss

Abstract

Content-based image retrieval (CBIR) has become an essential tool in several industries, allowing images to be searched efficiently based on their properties. In the context of digital rock studies, a CBIR system assists, for example, in applications in the oil and gas industry, where the identification and categorization of different rock formations directly impact resource management and extraction. In this paper, we present a Deep Metric Learning (DML) approach using the N-pair Loss function in conjunction with an architecture composed of notable backbones from the literature. The experiments were conducted on two datasets, one consisting of carbonates and the other of sandstones. The result of the model trained with the DenseNet201 backbone reached \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{98.87 \pm 1.17\%}$$\end{document} of F1-Score and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{98.83 \pm 1.22\%}$$\end{document} of mAP in the carbonate set. Furthermore, in the sandstone dataset, the model achieved F1-Score and mAP above 99%. This shows the potential of using N-pair Loss to improve the precision of image retrieval in materials industry applications, supporting better decision-making in industrial processes.