Three-stage progressive framework for Dongba ancient texts inpainting

Dongba ancient books have a thousand-year history, however, long-term natural erosion has caused varying degrees of damage, and the dual textual-visual nature of Dongba script makes inpainting more challenging than other script-based images. Existing methods work for slightly damaged images but fail for large-scale missing scenarios. To address this, we propose a novel three-stage progressive framework for severely damaged Dongba text image inpainting. Specifically, TsP first uses Edge Reconstruction Conv-Transformer to roughly inpaint character edges for contours, then employs Complete-character Matching Multi-Attention to retrieve corresponding complete-characters from a Dongba dictionary. Finally, the repaired edges and complete-characters serve as priors to guide Dual-Branch ResNet and Prior-Driven Fourier Convolution for fine inpainting. Extensive experiments on DB1404 dataset confirm the superiority of TsP in PSNR, SSIM, LPIPS, and FID across irregular mask rates from 1% to 50%, significantly outperforming state-of-the-art methods in all evaluation metrics.