DSA-YOLO: dual-stage attention YOLO with deformable large kernel and Swin Transformer for small object detection in aerial building defect detection

Housing constitutes a fundamental component of the urban environment, and the maintenance and management of urban buildings have attracted growing attention from both the public and governmental authorities. With the development of deep learning, automated building defect detection has become more practical and accurate. However, many defects appear as small objects, making them difficult to detect due to their limited size, complex backgrounds, and diverse forms. To address these challenges, this paper proposes a dual-stage attention building defect detection network, named DSA-YOLO. The architectural novelty lies in synergizing a deep semantic enhancement module (DSEM) using large deformable kernels to adaptively capture fine-grained geometric details of micro-defects, alongside a Swin Transformer-based global feature interaction module (GFIM) for robust long-range context modeling. Additionally, a supervised contrastive learning strategy is incorporated to boost the network’s ability to distinguish defects from background noise. Experimental results show that DSA-YOLO achieves mean average precision scores of 0.723 and 0.858 at an intersection-over-union threshold of 0.50 (mAP@50) on the CDD and CUBIT-Det dataset, with corresponding mAP@50–95 scores of 0.483 and 0.567. Notably, it outperforms the state-of-the-art YOLOv13 baseline by 0.016 and 0.021 in mAP@50 on these datasets respectively. Ultimately, this framework provides a highly accurate and deployable solution, demonstrating broad applicability for UAV-based structural health monitoring and diverse remote sensing inspection tasks.