CrackInst: A Real-Time Instance Segmentation Method for Underwater Dam Cracks

Abstract

Detecting cracks in underwater dams is crucial for ensuring the quality and safety of the dam. Currently, the mainstream method uses underwater robots equipped with cameras for automatic detection of underwater dam cracks. However, most existing methods for dam crack detection are based on semantic segmentation, which cannot accurately locate and segment individual cracks in images at the same time. To solve this problem, instance segmentation technology was introduced. However, current instance segmentation methods have slow detection speeds and reduced performance in underwater scenarios. Therefore, we propose a new underwater dam crack instance segmentation model. First, we introduce the CSPLiteNet feature extraction network to optimize the backbone network structure and enhance multiscale feature extraction capabilities. Second, we innovatively introduce the LiteC3 module to fuse feature information from different branches, obtaining richer and more global feature representations, thus enhancing the robustness of the model and effectively simplifying computation. Next, we introduce the SFSPP module, redesigning the atrous pyramid module to provide the network with a wider receptive field. Our method has shown excellent performance in underwater dam crack and object detection, with AP50 reaching 51.4% and 55.3%, respectively.