In the process of repairing damaged mural images, it is often overlooked that these images have unique painting techniques. To address this issue, we propose a method based on recurrent feature inference and global-local attention for restoring damaged mural images. In order to enhance the ability of the network model to capture the structural and textural features specific to mural images, we designed a coherent semantic attention inference module at the feature inference stage of the network model. Additionally, we replaced the original attention module with a global-local attention module in the feature inference stage to strengthen the connection between missing and background regions and generate high-quality feature maps based on the morphological characteristics of the mural. Experimental results show that our method can significantly improve the semantic disorder and pseudo-shadow phenomenon in mural images repaired by other algorithms, and generate more detailed information for damaged mural images, thus making the results of mural image restoration more natural and accurate.