Multi-Scale Gated Attention Transformer-Based Defect Detection for Historical Buildings: A Case Study of the Kaifeng Wall (Ming and Qing Dynasties, China)

Historical buildings, subjected to long-term weathering, pollution, and human activities, often develop defects such as cracks and mortar loss. Accurate detection is vital for cultural heritage preservation, but manual inspections are inefficient and subjective. Although CNNs have made progress in image segmentation, their limited receptive fields restrict the capture of global crack features. This study takes the Kaifeng Ming-Qing City Wall as a case, constructing a high-precision defect dataset with phased augmentation and strict data partitioning to prevent data leakage. In this study, we present a Multi-Scale Gated Attention Transformer (MSGAT), formulated not as a fundamentally new architecture but as a purposefully engineered integration of complementary mechanisms — including depthwise separable convolutions, Efficient Additive Attention, and CBAM — within a cascaded Swin Transformer framework. Experimental results show that MSGAT achieves a MIoU of 66.96%, Mean Dice of 71.11%, Precision of 66.93%, Recall of 58.82%, and F1-score of 61.54% on the test set, significantly outperforming traditional CNN-based segmentation models. Moreover, MSGAT exhibits superior crack edge continuity and more accurate localization of mortar loss areas, demonstrating strong robustness against environmental interference and validating its effectiveness in historical building defect detection.