CMMS-Net: Data-Efficient Semi-Supervised Crack Segmentation With Multi-Scale Dataset Memory

Abstract

Automated pavement crack detection is vital for infrastructure maintenance. We propose CMMS-Net, a data-efficient semi-supervised framework that achieves engineering-grade segmentation accuracy with only 50% pixel-level annotations. Core to our approach is the Multi-Scale Dataset Memory Unit (MSCMU), which leverages dataset-level crack prototypes to preserve topological continuity and suppress background noise under limited supervision. Validated across 12 public datasets, the proposed algorithm has achieved state-of-the-art performance and even outperforms SOTA algorithms on some datasets. We demonstrate that 50% annotations represent a practical optimum for safety-critical inspection. This work establishes a reproducible benchmark for resource-constrained infrastructure monitoring in intelligent transportation systems.