Topology-aware dual-branch network via lightweight Mamba for crack segmentation and quantification

Crack detection plays a crucial role in assessing the technical condition and facilitating the maintenance of pavements. Image segmentation is one of the most promising techniques for crack detection applications. However, crack segmentation is challenging due to complex pavement conditions. Existing methods either overlook the topological continuity of crack structures or exhibit limited capability in extracting semantic information. To address these shortcomings, a dual-branch crack segmentation network is proposed that emphasizes topology awareness and incorporates Mamba. First, a topology-aware module (TAM) based on dynamic snake convolution is proposed to extract topological information, which is used to construct the topology-aware branch. To reduce the high computational complexity of dynamic snake convolution, the TAM integrates horizontal convolution, vertical convolution, and the proposed direction selection module (DSM), which also improves the accuracy. Second, a lightweight vision state space module (LVSSM) is designed to construct the semantic branch, which reduces computational costs based on Mamba while effectively capturing long-distance dependencies. Third, an attention-based feature fusion module (AFFM) is proposed, augmented by a spatial enhancement module (SEM) designed to improve the spatial information within both branches. Features from both branches are dynamically fused layer by layer. Fourth, a segmentation-based crack length quantification method applicable to any width is proposed. This method can be combined with crack segmentation methods to achieve automated tasks such as crack measurement and pavement technical condition inspection. Finally, extensive experiments are conducted on three public datasets. The performance of the proposed model exceeds other state-of-the-art methods.