A pavement crack segmentation method based on deformable convolution and enhanced perceive network

Cracks in pavement roads are critical indicators of their structural and operational conditions. Timely identification and remediation of these cracks can extend the infrastructure’s lifespan, reduce vehicle fuel consumption, and improve safety and comfort. Traditional manual inspection of road images is costly and time-consuming. Non-data-driven automated methods have relatively low accuracy, and while data-driven approaches have shown improvement, we found that previous data-driven methods focused too much on global information and neglected details. Additionally, related datasets generally suffer from an imbalance between crack areas and background areas.To address the aforementioned issues, this paper introduces a novel pavement crack segmentation method employing deformable convolution and the Enhanced Perceive Network (EPerNet). Our approach combines deformable convolution with layer normalization and a feed-forward layer to produce both high-level and low-level feature maps. Additionally, given the emphasis on detail in crack segmentation, we designed a backbone with fewer layers and mixed the output feature maps from different output layers of the backbone. The feature maps from different heads were finally fused in a weighted manner. To address the imbalance between crack areas and background areas, we incorporate the Online Hard Example Mining strategy during training. Experimental results demonstrate that our proposed network delivers superior performance, offering an effective solution for precise pavement crack segmentation.