TOPOSegNet: Road Crack Extraction Network Based on Topological Relationship Modeling

Cracks are a typical manifestation of aging transportation infrastructure. Currently, segmentation models based on convolutional neural networks (CNNs) suffer from structural disconnection in recognizing linear cracks, that is, the crack segmentation integrity problem. Through an in-depth study of crack morphology, we propose TOPOSegNet to explore a crack extraction method different from the traditional semantic segmentation paradigm. In this article, our main contribution is to propose a topology decoder (TD) for improving the crack segmentation integrity problem by learning the linear topology of cracks. Meanwhile, an adaptive sampling mechanism based on the crack node density is proposed to solve the sample scale imbalance problem. A dataset called road crack topology dataset (RCTD), which contains crack topological features with large numbers and diversity, is proposed. After experimental validation, TOPOSegNet incorporating crack topology features achieves similar performance to mainstream segmentation models in segmentation performance and improves the integrity of crack segmentation in accurate segmentation results.