Image inpainting approach for Dunhuang murals based on dynamic radius guidance and Improved Fast Marching Method architecture

Limited sample data due to the difficulty in digital acquisition of Dunhuang murals renders existing deep learning-based inpainting approaches prone to overfitting, causing color deviation and texture distortion in inpainted images. To address these challenges, this paper proposes DR-IFMM, an inpainting approach based on a dynamic radius strategy and an improved Fast Marching Method (IFMM) architecture. DR-IFMM addresses these challenges by decomposing the inpainting process into several key stages. First, it adaptively computes two optimal radii based on the pixel density of damaged regions within the neighborhood of the target pixel, dynamically optimizing the inpainting of irregular and large-area defects. Second, the weight calculation rules of the FMM algorithm are refined to improve the accuracy of boundary and texture line inpainting. Finally, an image recomposition strategy integrates global structure and local details, yielding coherent textures and fine details. Experimental results on the Dunhuang mural dataset demonstrate that DR-IFMM outperforms competing approaches in terms of SSIM, PSNR, and LPIPS, effectively recovering the original appearance of damaged murals. This validates the practical value of DR-IFMM in Dunhuang mural inpainting, contributing to the digital preservation and inheritance of cultural heritage.