Key points are not available for this paper at this time.
Cracks in heritage timber members pose significant threats to structural safety and conservation. However, conventional inspection methods, whether manual or non-destructive, are often labor-intensive, operator-dependent, and risk causing secondary damage. This paper introduces YOLO-SDD, an enhanced one-stage detector based on YOLOv8n, designed for reliable crack detection in historic timber structures. The model incorporates three key improvements: SPDConv for detail-preserving downsampling, SDLK, which integrates SPPF with deformable large-kernel attention to suppress wood-grain interference and adapt to irregular crack geometry, and C2fDual to enhance cross-scale feature fusion. A dedicated in-situ dataset of 695 images from Weishan, Yunnan, was curated and expanded to 1671 training samples via targeted augmentation. On this dataset, YOLO-SDD achieves 87. 2% precision, 72. 6% recall, 80. 9% mAP50, and 54. 4% mAP50: 0. 95, outperforming the YOLOv8n baseline, which achieved 76. 7% and 50. 5% on the same mAP metrics, respectively. Ablation studies validate the complementary contributions of each module. Additionally, we deploy the model as a lightweight desktop tool using PyQt5 and ONNX, providing visual annotations and detection logs to facilitate rapid, non-destructive screening in conservation practice. The results demonstrate that YOLO-SDD offers a practical and accurate solution for crack detection, supporting the safety assessment and sustainable reuse of heritage timber structures.
Chen et al. (Fri,) studied this question.