Geometrically consistent energy-derivative attention CNN for semantic segmentation of multicategory structural damage

Abstract

Engineering structural damage often exhibits diverse and complex features across multiple scales within small-scale regions of interest (ROI), complicating post-earthquake assessments. This paper proposes an interpretable deep learning (DL) framework for semantic segmentation of multicategory damage. Energy-derivative attention modules are integrated into convolutional neural networks (CNNs) to enhance feature extraction of small-scale ROI. Geometrically consistent and focal-informed (GCF) loss function emphasizes the regions and boundaries of small-scale ROI, incorporating geometrical constraints of split line length, curvature, and area. Mosaic data augmentation method further mitigates feature imbalance. The proposed method outperforms the baseline with an mIoU increase from 80.67 % to 88.88 %. IoU for concrete spalling reaches 89.16 %, and for bar buckling improves to 82.96 %. The synergy of geometrical consistency, energy-derivative attention, and mosaic augmentation method significantly enhances CNN performance for multicategory damage. Finally, the framework is deployed in graphical user interface (GUI) software, enabling structural assessment of post-earthquake buildings.