融合多尺度特征与空间重构的敦煌壁画修复

Dunhuang Mural Inpainting Combining Multi-scale features and Spatial Reconstruction

摘要:

针对敦煌壁画修复过程中存在空间特征冗余和特征交互不充分的问题，提出一种融合多尺度特征与空间重构的敦煌壁画修复模型。该方法首先利用空间重构单元，通过特征分离与重构操作来优化特征学习能力。其次，设计了一种结合多尺度融合与门控残差连接的聚合多尺度上下文模块，用于更新缺失区域特征。最后，在结构纹理特征融合网络中引入3D排列的全局注意力机制，从而强化纹理和结构特征之间的交互。在敦煌壁画数据集上的实验结果表明，所提方法能够有效修复破损的敦煌壁画，修复后的壁画具有较好的结构及细节信息，且在大面积破损修复及复杂纹理修复方面表现出较强的性能。

Abstract:

To address the issues of spatial feature redundancy and insufficient feature interaction in the inpainting process of Dunhuang murals, a Dunhuang mural inpainting model that Combining Multi-scale Features and Spatial Reconstruction is proposed. This method initially employs a spatial reconstruction unit to enhance feature learning capacity through feature separation and reconstruction operations. Subsequently, an aggregated multiscale context module, which combines multi-scale fusion and gated residual connections, is designed to update the features in missing regions. Finally, a 3D- configured global attention mechanism is integrated into the structure-texture feature fusion network to reinforce the synergy and interaction between texture and structural features. Experimental results on the Dunhuang mural dataset demonstrate that the proposed method effectively inpaints damaged Dunhuang murals. The inpainted murals exhibit well-preserved structural integrity and detailed features, demonstrating superior performance in inpainting large-area damage and reconstructing complex textures.