Multi-stage Progressive Reasoning for Dunhuang Murals Inpainting

Dunhuang murals suffer from fading, breakage, surface brittleness, and extensive peeling affected by prolonged environmental erosion. The recurrent networks achieved promising performance when repairing severe damages to the mural. However, the recurrent reasoning occurs in the single receptive field network, which lacks awareness of different fields of murals. In this paper, we innovatively designed the multi-stage progressive reasoning (MPR) network containing global and local receptive fields. This network can recursively infer the structural and textural features from the hole boundaries and progressively tighten the constraints on the hole center. Moreover, To adaptively fuse feature information at various scales of murals, a multi-scale feature aggregation module (MFA) is designed to empower the capability to select the significant features of murals. The execution of the model is similar to the process of the mural restorer (i.e., inpainting the structure of a damaged mural globally first and then progressively adding the local texture details based on this). Qualitative and quantitative experiments show that our method has excellent and comprehensive performance that exceeds existing image inpainting methods.