Automated high-resolution asphalt pavement crack segmentation using deep convolutional neural networks with repeated hierarchical feature fusion

Automated collection and detection of asphalt pavement crack conditions are essential for evaluating road conditions and maintenance planning. However, determining crack conditions accurately and efficiently has been challenging due to their small object-to-background information ratio, poor contrast with defect-free regions, and vulnerability to noise, such as stains and shadows. We approach this challenge by developing a series of methods that generate high-resolution asphalt pavement crack segmentation, the HRCrack series, which attends to hairline cracks by repeatedly fusing hierarchical features learned using convolutional neural networks. We validated and compared the models with ten widely used models on six open crack datasets. The results demonstrated that our models outperformed the considered methods on the self-made and six open-source datasets. Our best-performing model, HRCrack48, achieved an optimal dataset scale (ODS) F1 score of 0.948 on the self-made dataset. Our smallest model, HRCrack18s, was the fastest (25 FPS) while still providing competitive performance.