Abstract
This research proposes a deep learning workflow for the detection and segmentation of overpaint in paintings by integrating advanced imaging techniques with expert validation. 3D digital optical microscope images, raking light imaging, Ultraviolet-induced Visible Fluorescence Photography (UIVFP), Macro X-ray Fluorescence (MA-XRF), and 3D surface data are collected to train the deep learning model. This multimodal approach ensures that a variety of overpaint characteristics, including subtle surface textures and layered features, are accurately represented. The deep learning model employs Vision Transformer (ViT)-based architecture to detect overpainted areas. Iterative refinement cycles, guided by quantitative performance metrics and expert feedback, ensure continuous improvement in predictive accuracy. Applying the trained model to new case studies will assess its generalisation beyond the training dataset. Workshops for conservation professionals will evaluate the model's usability. Beyond the professional sphere, targeted outreach initiatives will communicate the project's findings to wider audiences, highlighting how AI and digital imaging contribute to cultural heritage preservation. By combining advanced imaging, deep learning, and expert knowledge, this project fosters scalable, data-driven approaches to artwork analysis.
Researcher(s)
Research team(s)
Project type(s)