Existing inpainting methods for Dunhuang murals suffer from issues such as insufficient utilization of local spatial information and poor flexibility and adaptability in dealing with complex and irregular defects. To address these issues, a Dunhuang mural inpainting method that integrates local spatial enhancement with dynamic filtering is proposed. This method takes a locally efficient visual state space model that integrates axial shift convolutions and dynamic spatial attention as the encoder, aiming to enhance the model's capability in modeling the dependencies among adjacent pixels and achieve dynamic focusing on key defective regions. Meanwhile, a multi-feature adaptive selective fusion module is constructed to adaptively screen and efficiently fuse multi-scale features. Additionally, dynamic filtering Fourier convolution is introduced to dynamically generate frequency-domain weight matrices, thereby enhancing the realism of restored textures. Experimental results demonstrate that, compared with existing mainstream inpainting methods, the proposed approach achieves significant improvements in both subjective visual effects and objective evaluation metrics under scenarios involving irregular masks, powder masks, and real damaged murals.