Physical degradation detection on artwork surface polychromies using deep learning models

摘要

This paper presents the application of a Deep Learning algorithm fora classified accurate detection of different types of physical damages in artworkpolychrome surfaces. The algorithm was trained for the automated detection of threeclasses of typical surface deteriorations: cracks, blisters and detachments (losses). Theimage sets used in this study were previously recorded for the purpose of detailedsurface 3D reconstruction by the means of macro-photogrammetry of a wood painting.These high-resolution images were captured using 2:1 optical macro magnificationwith a generous overlapping, following the 3D reconstruction methodology, andprovided high quality details of the surface features to be classified. Specificactivation maps are used to visually emphasize the detected potential deterioratedareas. The purpose of this work was on one hand to validate a process of reusingphotogrammetry image data sets, used 3D reconstruction, for machine learningfeature detection training and on the other hand to provide a starting point for thedevelopment of an affordable real-time surface damage assessment system.