Cross teacher pseudo supervision: Enhancing semi-supervised crack segmentation with consistency learning

Abstract

In recent years, many crack segmentation techniques based on supervised learning have been widely employed in civil infrastructure maintenance. However, the accuracy of these supervised learning algorithms is heavily impacted by the quantity and quality of labels, while it is time-consuming and labor-intensive to manually annotate the crack locations. To address this issue, a new semi-supervised algorithm is proposed to leverage both labeled and unlabeled data through a cross-teacher-pseudo-supervision framework and cross-augmentation strategy. The proposed method employs two pairs of teacher–student models to mutually supervise each other using pseudo-labels generated from their respective teacher models. To boost the performance of the proposed algorithm, input, feature, and network perturbances are applied during training. In addition, there is no additional computation and storage overhead in the test phase. In comparative experiments, the proposed method achieves superior performance over that of several existing algorithms on four public crack datasets. Specifically, on the four datasets, the proposed method outperforms the supervised-only baseline by 0.92%, 1.29%, 4.14%, and 5.38% respectively, in mean intersection over union under the labeled ratio of 5%, and by 0.76%, 1.06%, 1.79% and 1.35% under the ratio of 10%. Additionally, detailed ablation experiments further confirm the efficiency of the proposed method.