MuralsNet: Dunhuang Decorative Mural Image Inpainting Based on BiFormer and GLM Fusion

Diverse costume and decorative patterns in Dunhuang murals have significant historical and cultural value. However, natural erosion and material aging have caused cracks, flaking, and missing regions. Existing image restoration methods are still challenged by complex decorative patterns, including insufficient modeling of long-range structural dependencies, incoherent local textures, and weak global consistency. To address these problems, this paper proposes MuralsNet, an EdgeConnect-based inpainting framework adapted to the structural characteristics of Dunhuang decorative murals. The original two-stage strategy of EdgeConnect (G1, G2) is retained to provide explicit edge guidance, while two targeted modifications are introduced into the G2 image completion network. First, BiFormer is embedded into the bottleneck stage of G2 U-Net to capture global contextual information and long-range dependencies through a bi-level routing attention mechanism. Meanwhile, a Global-Local Fusion Module (GLM) is constructed in the decoder stage to facilitate the effective fusion of high-level semantic features and local detail features, thereby improving structural coherence and texture restoration in damaged regions. Experiments show that MuralsNet achieves the highest PSNR, SSIM, and Masked Edge F1, as well as the lowest LPIPS, among the evaluated methods, while LaMa obtains the lowest MAE. Visual comparisons further show that MuralsNet better preserves fine decorative contours and local structural details in several representative examples. Ablation experiments further show that BiFormer and the GLM provide complementary improvements within the proposed framework. These results demonstrate the potential of the proposed task-oriented integration strategy for the digital restoration of complex decorative mural images.