Detection of Damaged Areas in Ancient Murals Based on Multi-Scale Dual Channel Attention U-Net

To address the issue that existing methods for detecting damage in ancient murals struggle to accurately identify damaged areas, this paper proposes a detection model based on a multi-scale dual channel attention U-Net. Firstly, a multi-scale dual channel module is designed to replace the encoder-decoder of the original U-Net, creating a network capable of detecting damaged areas of different scales in complex backgrounds. Secondly, an attention guidance module is designed to effectively merge the features of higher and lower layers. This enables the network to emphasize the prominent features of the damaged areas in murals. Additionally, this paper proposes a multi-stage loss summation strategy to enhance the performance of the network model. Experimental results on the DHMural and YNMural datasets show that this model surpasses other comparative methods in both subjective visual perception and objective evaluation metrics.