Blind Restoration of Masked Dongba Scripts Images via Hybrid CNN-Transformer Network

The restoration of degraded Dongba scripts, an ancient pictographic language of significant cultural value, faces unique challenges due to sparse structures and irregular damage patterns. Blind image inpainting, which repairs damaged regions without prior mask guidance, is critical for preserving such historical documents. Existing methods often fail to harmonize local texture recovery and global structural consistency, particularly under varying damage conditions. This study proposes a hybrid CNN-Transformer framework enhanced by a gated attention mechanism. The model integrates convolutional neural networks for local feature extraction and self-attention mechanisms for global dependency modeling, achieving precise reconstruction of fragmented strokes and textures. Evaluations on the Dongba text dataset (DB1404) demonstrate superior performance over state-of-the-art methods in both pixel accuracy and perceptual consistency. Qualitative results confirm the ability of our framework to preserve intricate details while maintaining structural coherence, offering a practical solution for safeguarding culturally invaluable manuscripts.