Learning to grade deterioration for ancient murals

Abstract

Ancient murals recording histories and myths are suffering different types of deteriorations, which makes these valuable cultural heritage unrecoverable. For preventive protection of ancient murals, the researchers need to rate the grade of the mural deteriorations. However, the present grading is mainly rely on the experience of the cultural relics’ protection staff. In this paper, we propose to learn a mural deterioration grading network (MDGN) for generating robust and accurate mural deterioration grading. MDGN has similar architecture with AlexNet but with less network parameters. We built two small mural deterioration grading datasets on briquettes and real grottoes. We use image patches to generate enough training samples for MDGN. Dense pixel-level grading can be obtained by the means of sliding window. Extensive experiments on simulated data and real murals verify the effectiveness of the proposed ancient mural deterioration grading method. Our method achieves superior grading results on both quantitative and qualitative comparisons with baseline.