CrackSegMamba: A Lightweight Mamba Model for Crack Segmentation

Crack localization and segmentation are essential for infrastructure maintenance and safety assessments, enabling timely repairs and preventing structural failures. Despite advancements in deep learning, crack segmentation remains challenging due to the need for real-time performance and computational efficiency. Existing methods often rely on large, resource-intensive models, limiting their practical deployment. We introduce CrackSegMamba, a novel model featuring Channel-wise Parallel Mamba (CPM) Modules, which achieves state-of-the-art performance with fewer than 0.23 million parameters and just 0.7 GFLOPs. CrackSegMamba reduces computational cost by 40-fold and parameter count by nearly 100-fold compared to existing models, while maintaining comparable accuracy. These features make CrackSegMamba ideal for real-time applications. Additionally, we present Crack20000, an annotated dataset of 20,000 concrete crack images to support further research and validation. Evaluations on the Crack500 [1] and Crack20000 datasets demonstrate that CrackSegMamba delivers comparable accuracy to leading methods, with significantly reduced computational requirements. Project page is available at: https://sites.google.com/view/cracksegmamba.