DBDCNet: Dual-Branch Bottleneck with Deformable Convolution for Real-Time Crack Segmentation

In recent years, real-time semantic segmentation has been developed rapidly and successfully applied in various fields.In this paper,a new real-time semantic segmentation network was proposed to be applied into crack detection of concrete buildings under complex background, which is based on dual branch feature extraction module, deformable convolution and nested multi-scale attention decoder. With such encoder-decoder structure, our network parameters are only 1.05M.We conduct experiments on self-made and open crack datasets, and the network achieves excellent performance.Compared with real-time semantic segmentation networks such as ENet, CGNet, DABNet, and LMFFNet, we have significantly improved prediction accuracy while maintaining an advanced level of inference speed.