Computational framework for the evaluation of the composition and degradation state of metal heritage assets by deep learning

The accurate assessment of the material constitution and degradation in newly discovered archaeological artefacts is paramount for the decisions surrounding a thorough treatment of the object during the restoration and conservation stages. The laboratories possess competent experts and complex devices to perform this analysis properly. Nevertheless, a timely hint of an artificial intelligence assistant regarding the chemical composition and corrosion compound localization of a metal asset could save additional time and resources. The present paper proposes such a computational framework based on deep learning techniques that, on the base of its automatic determination of the chemical concentration of the predominant metal from a microscope image, can subsequently independently also recognize and delineate the corrosion spots of the products specific to that metal. The experiments have been performed on iron and copper heritage items from the Oltenia Museum, Romania. The results suggest that, even with an economic training information in terms of microscope images and annotations, the artificial intelligence framework can provide on-site support for an early examination of metal heritage assets.(c) 2023 Consiglio Nazionale delle Ricerche (CNR). Published by Elsevier Masson SAS. All rights reserved.

Automated summary

The accurate assessment of material composition and degradation in newly discovered archaeological artefacts is crucial for decision-making in the restoration and conservation stages. This study proposes a computational framework based on deep learning techniques that can automatically determine the chemical concentration of the predominant metal from microscope images and identify corrosion spots specific to that metal. The results suggest that the artificial intelligence framework can provide on-site support for early examination of metal heritage assets, even with limited training data.