An Adaptive Crack Detection Network Based on Global-to-local Spatial Aggregation

Road cracks are a common form of damage in road maintenance. The complexity of crack shapes, textures, lighting conditions, and interference factors make accurate detection of cracks still a challenging task. To address these issues, we propose a novel adaptive crack detection network, which achieves precise crack recognition through an adaptive mechanism and multi-level feature fusion. The network is divided into multiple branches along the channel dimension, employing different convolution operations and assigning importance to dynamically select and emphasize key features. By utilizing cascaded large kernel convolution, the network captures rich multi-scale contextual information in deep layers. Additionally, the dual-stream design effectively integrates features from different hierarchical levels. We evaluated the proposed method on three benchmark crack detection datasets: DeepCrack, CFD, and Crack500. The method achieved F1 scores of 0.880, 0.669, and 0.746, and MIoU values of 0.887, 0.745, and 0.781, respectively, outperforming existing approaches.