Archival analog drawings for semantic segmentation of Roman Architectural Heritage using Deep Learning

Abstract. The present study aims to investigate whether the graphic code embedded in analogue architectural drawings—characterised by standardised textures and conventions—provides sufficient semantic information to support a simple, robust, and reproducible deep-learning-based segmentation approach, even under conditions of limited annotated data. The research focuses on a corpus of historical drawings preserved in the Archivio dei Disegni e Fototeca of the Dipartimento di Storia, Disegno e Restauro dell’Architettura at Sapienza Università di Roma, in which materials and construction techniques of Roman architectural heritage are represented through encoded graphic patterns and conventions. Starting from a limited set of 19 manually annotated drawings, a reproducible pipeline based on a U-Net architecture with a ResNet-34 backbone is developed, combining tiling strategies, data augmentation, and high-resolution inference. The results show high Overall Accuracy and Weighted IoU values, confirming the model’s ability to interpret the implicit graphic language of the drawings, even in conditions of strong class imbalance and limited data availability. Inference on unseen drawings demonstrates an acceptable degree of generalisation, opening new possibilities for the automatic semantic digitisation of historical graphic archives. The study highlights the potential of analogue architectural drawings as a structured source of knowledge for artificial intelligence applications in the documentation, analysis, and conservation of the built heritage.