ArtInsight: A detailed dataset for detecting deterioration in easel paintings

Abstract

ArtInsight is an innovative dataset designed to detect dete- rioration in fine art, specifically easel paintings. The dataset includes high-resolution images captured at the University of Granada using a digital camera with a 105 mm lens, ISO 125, F5, and a shutter speed of 1/13, and processed for color cal- ibration. Two types of images are featured: those showing stucco technique interventions and those with Lacune from the loss of the Painting Layer (LPL). The VGG Image Annota- tor was employed for manual damage labeling, with annota- tions exported in JSON format and labeled for stucco and LPL damages. The dataset comprises 14 images with 2909 dis- tinct damage areas, split into training and validation datasets. Developed using Python 3.7 and fine-tuned on a pre-trained Mask-RCNN model, this dataset demonstrates high accuracy rates (98–100 %) in damage detection. ArtInsight aims to fa- cilitate automated damage detection and foster future research in art conservation and restoration.