3D Complex Surface Completion of Sculptures Using a Deep Learning Approach

Incomplete or missing data in three-dimensional (3D) models of cultural heritage assets can result in flawed reconstructions, undermining their value for visualization, documentation, digital restoration, and 3D printing. Traditional surface-repair methods often struggle to recover complex morphological features typical of heritage objects, such as ornamentation, wear patterns, or expressive facial traits. Inspired by the success of neural inpainting techniques in 2D image restoration, we propose SR-CurvANN (Surface Reconstruction based on Curvature-Aware Neural Networks), a novel method that integrates AI-driven inpainting with geometric surface reconstruction to repair damaged or incomplete 3D models. The approach leverages curvature-based 2D representations of 3D surfaces to train neural networks capable of inferring plausible geometry in missing regions. Once the inpainted curvature image is generated, a coarse-to-fine deformation process reconstructs the corresponding 3D surface, ensuring geometric coherence. SR-CurvANN is designed to generalize across a wide range of cultural heritage geometries, learning from a large and diverse training set of 3D models. In an extensive evaluation involving 959 damaged models, our method consistently achieves high-fidelity restorations, demonstrating its potential for digital preservation, reconstruction, and interpretation of cultural heritage assets. This work exemplifies how AI can enhance the reliability and expressiveness of digital heritage 3D documentation workflows.