A mask-aware adaptive texture and structure fusion network for mural inpainting

Deep learning-based inpainting models have effectively restored natural images; however, their efficacy in mural restoration is constrained, encountering challenges such as texture, structure, and color degradation. This paper presents a mask-aware adaptive texture and structure fusion model designed for mural inpainting. Employing a coupled approach, the model facilitates texture synthesis with structural constraints and texture-guided structural reconstruction, seamlessly integrating image texture and structure to achieve enhanced restoration. The model’s generator, comprising parallel U-Net architectures, extracts texture and structure features during the encoding phase. A Global Context Cooperative Enhancement (GCCE) module captures deeper feature interactions. A Mask-Aware Pixel Reorganization (MPU) module enhances detail restoration during decoding by leveraging complementary texture and structure features. Furthermore, an Adaptive Bidirectional Interactive Fusion Network (ABIF) and a Local-Global Contextual Affinity (LGCA) module work in tandem to integrate and enhance the reconstructed features. In comparison to state-of-the-art methods, our model exhibits superior performance in both qualitative and quantitative evaluations, effectively preserving the textures and structures of ancient murals.