Diff-BAM: a generalized adaptive diffusion model for cultural heritage image inpainting

Abstract Cultural heritage, as a precious carrier of history and culture, embodies profound artistic and historical value. However, over time, cultural heritage items are vulnerable to varying degrees of damage due to improper preservation. Existing image-inpainting methods have limitations in detail recovery and inpainting efficiency, making it difficult to meet the demands for high precision and efficiency. This article proposes an image-inpainting method for cultural heritage based on adaptive denoising and multi-scale filtering (Diff-BAM). The model improves inpainting effectiveness and efficiency by dynamically adjusting the reverse denoising process and introduces an edge-aware strategy to address the issue of rough edges in inpainting the image. Additionally, the model incorporates a multi-scale filtering mechanism within the U-Net architecture to further enhance the inpainting details. Due to the lack of publicly available Thang-ka datasets, the article uses a self-built dataset and a landscape painting dataset for experimental validation. Experimental results show that Diff-BAM outperforms the mainstream methods in terms of inpainting details and efficiency. This study demonstrates the potential application of image inpainting technologies in the field of digital culture for the preservation of cultural heritage, providing an efficient and precise solution for artifact inpainting.