Mitigating model coupling in semi-supervised segmentation via deep non-consistent mean teacher and fully collaborative learning

Abstract

Semi-supervised learning methods based on teacher-student frameworks have achieved remarkable success in image segmentation. However, popular teacher-student models are prone to early subnet coupling, which limits segmentation performance. Moreover, most existing approaches rely on strong-weak perturbation schemes for consistency learning, overlooking peer-level supervision between different perturbations and failing to fully exploit the potential information from unlabeled data. To address the above issues, we propose ComMatch, a novel semi-supervised image segmentation method built upon deep non-consistency and fully collaborative learning. Specifically, a deeply non-consistent mean-teacher structure is designed, which expands the learning space by constructing deep inconsistencies at both the data and network levels within a multi-stream learning framework and can effectively alleviate the problem of early subnet coupling. Meanwhile, to maximize the latent information from unlabeled data, a fully collaborative learning strategy is proposed, which explores the necessity of peer-level loss under deep inconsistency perturbations and further combines cross-level and peer-level losses to deeply mine the latent knowledge from unlabeled data. Experimental results show that the proposed ComMatch method surpasses the current state-of-the-art methods, achieving segmentation accuracies of 78.68% and 77.89% (1/16) respectively in the mIoU metric on the PASCAL VOC and Cityscapes datasets. Code is available at https://github.com/Minchongdan/ComMatch.