Attentive Sparsity and Bidirectional Fusion for Efficient Crack Instance Segmentation in UAV-based Bridge Inspection

• Dynamic Sparse Convolution adaptively focuses on fine crack features • Weighted Bi-directional FPN enhances multi-scale fusion for thin cracks • YOLOv11-DSC achieves 89.2% mAP@0.5 at 68 FPS on UAV bridge dataset • Comprehensive 2D-to-3D projection validation with 1.97pixels reprojection error Automated crack segmentation from unmanned aerial vehicle (UAV) imagery is critical for efficient bridge inspection but remains challenging due to the fine-scale, low-contrast nature of cracks and the complex, noisy environments. To address this, YOLOv11-DSC is proposed, which is a novel real-time instance segmentation network that introduces two key architectural innovations to the YOLOv11 framework for domain-specific optimization. Dynamic Sparse Convolution (DSC) modules adaptively sparsify feature computations to focus network capacity on high-spatial-frequency crack regions, while the Weighted Bidirectional Feature Pyramid Network (WB-FPN) employs learnable channel-wise attention to dynamically fuse multi-scale features, collectively enhancing sensitivity to fine cracks and improving the continuity of elongated crack structures. Trained and evaluated on a challenging self-collected UAV bridge crack dataset featuring shadows, water stains, and handwriting interference, YOLOv11-DSC achieves 89.2% mAP@0.5, surpassing the baseline YOLOv11n-seg by 43.8%, while maintaining a high inference speed of 68 FPS. Comprehensive ablation studies validate the contribution of each component, and comparisons with state-of-the-art real-time segmenters demonstrate the superior accuracy and robustness of the proposed model. The proposed framework provides a reliable, efficient AI solution for automated visual inspection, establishing a high-quality 2D segmentation foundation essential for subsequent 3D geometric quantification pipelines.