Automated Segmentation and Length Estimation of Tiny Building Sealant Cracks Using Deep Learning

Crack formation in building sealants allows moisture ingress, accelerates degradation, and reduces building envelope durability. Traditional visual inspections are slow and subjective, underscoring the need for automated, quantitative methods. This study presents an end-to-end deep learning framework for segmenting fine adhesive cracks and converting pixel-level representations into real-world lengths using marker-based calibration. Six segmentation architectures: U-Net, Attention U-Net, Residual Attention U-Net, DeepLabv3+, YOLOv11-seg, and Samurai, were evaluated on a custom image dataset collected from outdoor field-aged sealant specimens. Attention U-Net achieved the best performance (Dice = 0.867, IoU = 0.764), preserving narrow crack geometry. A YOLO-based marker detector (precision = 0.996, mAP@0.5 = 0.787) enabled accurate pixel-to-millimeter calibration. Crack length estimation via skeletonization produced consistent results across varying conditions. The framework provides a scalable, data-driven solution for automated sealant inspection, supporting predictive maintenance, cost reduction, and enhanced durability of building envelopes.