Deeplab-based Road Crack Segmentation with Attention and Dynamic Convolution

The automatic segmentation of road cracks is pivotal for detecting engineering road defects. This paper proposes a crack segmentation method based on the DeepLabV3+ model, which integrates the attention mechanism and dynamic snake convolution. The attention module enhances both channel and spatial attention. The dynamic snake convolution is employed to precisely capture features of linear structures, with adaptability to focus on elongated and tortuous local structures. The module’s design ensures effective preservation of crucial information across different global morphologies, and significantly improves the extraction of detailed features, leading to more accurate segmentation outcomes. The proposed method demonstrates state-of-the-art results on two benchmark datasets, DeepCrack and Crack500. On DeepCrack, it achieves an accuracy, recall, F 1 , and MIoU scores of 94.1%, 93.6%, 89.4%, and 89.4%, respectively. On Crack500, the recall, F 1 , and MIoU scores are 89.4%, 83.9%, and 82.9%, respectively.