Abstract: In response to the challenging problem of accurately automating the localization of peeling areas in ancient murals,this paper proposes a lightweight network model based on a multi-scale fusion attention network.Firstly,a multi-scale fusion attention module is introduced to enable the network to learn features at different scales,with a focus on the most critical features,thus improving the accuracy of mural missing area localization.Deep separable convolutions are employed in the proposed multi-scale fusion attention module to make the network model more lightweight.Secondly,a combination of cross-entropy loss and Dice score is used as the loss function,and the Adam optimizer is applied to further enhance the accuracy of mural missing area localization.Additionally,datasets of Dunhuang Mogao Grottoes murals and Yunnan Shiping Luose Temple murals are constructed,and their peeling areas are manually annotated.Experimental results demonstrate that the proposed network model accurately localizes peeling regions in ancient murals.In comparison with existing deep learning methods,this model significantly reduces the number of parameters and exhibits better performance in terms of subjective visual quality,objective evaluation metrics,and generalization capabilities.

Key words: Mural damage, U-Net, Multi-scale, Attention mechanism, Deep learning, Lightweight