Abstract: This paper introduces a line-drawing-guided Transformer-enhanced U-Net (LDG-TEUN) for the digital restoration of Dunhuang murals. A cross-attention module that integrates axial attention with two-dimensional positional encoding is embedded in the encoder to capture global structures and long-range dependencies, thereby alleviating the structural loss caused by large-scale damage. A dual-domain partial convolution (DPConv) unit is then designed to jointly model spatial- and frequency-domain features, enhancing the reconstruction of complex textures and fine edges while addressing challenges in detail recovery. Finally, a composite loss function is formulated to enforce structural consistency, texture fidelity, and color distribution simultaneously, which improves overall restoration quality and, in particular, enables more authentic color reconstruction. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches in both structural coherence and color restoration, confirming its effectiveness and practical value for the digital conservation of Dunhuang murals.

Key words: image inpainting, Transformer, U-Net, line-drawing-guided, Dunhuang murals