Pavement Crack Detection Based on an Improved YOLOv11 Network with CrackFormerV2 and Adaptive Threshold Focal Loss

Pavement cracking is a common form of distress in asphalt pavements, and crack detection is essential for pavement maintenance. To improve detection accuracy in complex backgrounds, this study develops an enhanced YOLOv11-based detection algorithm. First, we construct CACrack2025, a pavement crack dataset that provides samples for model training and evaluation. Second, we design CrackFormerV2, a feature extraction module that combines cross-scale feature interaction with contextual information aggregation to improve the representation of slender, low-contrast cracks. To address the imbalance between crack targets and background regions, we incorporate Adaptive Threshold Focal Loss (ATFL), which reweights easy and hard classification samples according to their confidence. Experiments on CACrack2025 show that the proposed method achieves an mAP@50 of 86.0%, 4.6 percentage points higher than the baseline YOLOv11n model. Under the current dataset and evaluation protocol, the method maintains higher detection accuracy than the evaluated baselines. These results support the effectiveness of the proposed components for pavement crack detection.