Dynamic Mutual Adversarial Learning for Semi-Supervised Semantic Segmentation of Underwater Images with Limited and Noisy Annotations

Swift and accurate semantic segmentation of underwater images is key for precise object recognition in complex underwater environments. Nonetheless, the inherent complexity of these environments and the limited availability of labeled data pose significant challenges to underwater image segmentation. Traditional deep learning methods struggle to cope with limited and noisy annotations. In this paper, we delineate the formulation of a novel semi-supervised paradigm with dynamic mutual adversarial training for the semantic segmentation of underwater images. This paradigm identifies the sources of inaccuracies in pseudo-labeling by analyzing different confidence maps, generated by models with unique prior knowledge. A dynamic reweighing loss function is then employed to orchestrate the mutual instruction of two divergent models. Furthermore, the delineation of confidence map is facilitated via adversarial networks, which involves simultaneous adversarial refinement of the discrimination network and the segmentation model, using the predictions with high-confidence maps as pseudo-labels. Experimental results on public underwater datasets verify that the proposed method can effectively improve semantic segmentation performance under the condition of a small amount of labeled data.