Generation of synthetic labeled datasets for anomaly detection in heritage architecture

The conservation of architectural heritage is a crucial task for preserving the history and cultural value of our cities. This work proposes a methodology that combines advanced techniques of multisensory data acquisition and analysis to generate labeled datasets of anomalous zones in historic buildings. The proposed method includes the use of artificial intelligence techniques, such as the Segment Anything Model (SAM), for image labeling, ensuring high precision through the involvement of conservation experts. Furthermore, due to the limited number of images focusing on specific anomalies, we address this issue by employing the Gaussian Splatting algorithm, which enables the reconstruction of 3D scenes and the generation of synthetic images centered on the anomalies of interest. These synthetic images are automatically labeled, resulting in a high-quality dataset suitable for training neural networks that support preventive conservation strategies in architectural heritage. By tackling the critical challenge of insufficient labeled data, this approach improves the robustness and reliability of machine learning models for anomaly detection in architectural structures.