Bayesian Patch-Based Inpainting for Art Restoration: Uncertainty-Aware Recovery of Damaged Artworks

Loan Gia

Keywords: art restoration, Bayesian inpainting, uncertainty quantification, cultural heritage, diffusion models, style preservation, digital conservation

TL;DR: Bayesian patch-based inpainting for artworks combining diffusion models with variational inference to achieve uncertainty-aware restoration while preserving artistic style (28.7 dB PSNR, 0.85 SCS).

Abstract: This paper presents a Bayesian patch-based inpainting framework for art restoration that combines diffusion models with probabilistic inference to achieve uncertainty-aware reconstruction of damaged artworks. Our method introduces patch-wise variational autoencoders for localized uncertainty quantification and iterative refinement of missing regions. Experiments on cultural heritage datasets demonstrate superior accuracy (28.7 dB PSNR) and style preservation (0.85 SCS) compared to existing methods, while providing interpretable confidence estimates for conservators. The technical innovation lies in our hybrid architecture that maintains artistic fidelity while quantifying reconstruction ambiguity.

Submission Number: 8