Texture and geometric feature-fusion-based network for Dunhuang mural inpainting

In recent years, deep-learning-based image inpainting methods have emerged as a popular research area. However, the application of such methods to mural inpainting presents several challenges. First, mural images often contain complex textures and rich details. Traditional inpainting methods struggle to preserve texture and detail information effectively and cannot ensure consistency between restored areas and the original mural. Second, the missing regions of murals often contain complex geometric structures and artistic styles, requiring mural inpainting algorithms to understand an image’s global semantics and accurately capture local details. This paper proposes a texture and geometric feature fusion network for Dunhuang mural inpainting consisting of two subnetworks: a Primary Inpainting Network (PIN) and a Refinement Enhancement Network (REN). The PIN extracts geometric features by incorporating fresco line drawings. It utilizes a Mamba-enhanced encoding module and gated convolution in its encoder to capture image texture features effectively, thereby enhancing the clarity of texture details. Then, the Dynamic Multi-scale Semantic Fusion Module (DMSFM) combines global and local information from texture and geometric features, completing the initial inpainting of a damaged mural. The REN specializes in inpainting image details by recovering complex textures, fine edges, and local structures. Randomly selected Narrative, Buddhist, and Caisson murals from a Dunhuang mural painting dataset were used to test the proposed network. Comparative experimental results demonstrate that the proposed method achieves superior mural inpainting outcomes compared with popular existing methods.