Semantics-Enhanced Refiner in Skip Connection for Crack Segmentation

Automatic crack detection has important engineering significance for pavement surfaces and civil structures. At present, feature enhancement methods in UNet and its most variants are scale-wise, and each scale has an independent skip connection, so they do not pay attention to both spatial details and global semantics at the same time. Therefore, we propose a semantics-enhanced refiner (SER) that can simultaneously introduce low-level detail information and high-level semantic information in each skip connection, achieving feature enhancement for each scale. As a plug-and-play module, the SER is of multi-input single-output, and can be embedded into any encoder-decoder structure. In addition, we construct a lightweight and real-time U-shaped crack segmentation network that insert a SER at each scale of the encoder and decoder. Finally, by comparing proposed model with six established segmentation algorithms on two public crack datasets, DeepCrack and MSCI, our model achieves higher segmentation accuracy with extremely low parameters and FLOPs.