As precious cultural heritage, ancient murals suffer from complex damage, and their restoration process faces great challenges, particularly in the reconstruction of delicate textures and intricate structures. To address the issue that existing restoration methods often struggle to restore murals accurately, we propose a three-stage network based on global-local features for mural restoration. Our three-stage restoration process emulates the human eye fixation mechanism, progressing from general browsing to detailed observation, and finally to optimization. This phased approach of observation allows for a more effective restoration of the fine details and overall visual impact of the mural. Experimental results show that our model can restore damaged areas of ancient murals with clearer structures, more realistic texture features, and vibrant colors, outperforming other comparison methods.