A semi-supervised crack segmentation framework based on uncertainty perception and dual-student mean teacher

Abstract Fully supervised crack segmentation techniques can significantly enhance the accuracy of civil structure health monitoring; however, they heavily depend on a large volume of annotated data. To mitigate this reliance, we propose a semi-supervised crack segmentation framework based on the uncertainty-aware dual-student mean teacher (U-DSMT). The proposed method utilizes cooperative training of the DSMT to leverage discrepancies between sub-networks for multi-view learning, thereby facilitating the acquisition of richer and more comprehensive feature representations while alleviating the coupling issue between the teacher and student models. Additionally, to address the trade-off between sub-network discrepancies and the quality of pseudo-labels, we introduce dynamic uncertainty patch adjustment technology. Furthermore, we propose the CrackUnet model, which preserves crack details through hierarchical feature fusion, thereby enhancing the accuracy of semi-supervised crack segmentation. Findings from the experiments show that the U-DSMT approach achieves leading segmentation performance across two publicly available datasets, exhibiting robust segmentation accuracy even with a small number of annotated samples.