Review of deep learning-based semi-supervised semantic segmentation
Semantic segmentation finds extensive applications across various real-worlddomains,yet its training necessitates a substantial number of pixel-level annotated images,incurring high costs. Semi-supervised semantic segmentation,which leverages a small set of labeled images alongside a large pool of unlabeled ones,aligns better with real-world scenarios,garnering widespread attention globally. This paper summarizes and analyzes the recent advancements in deep learning-based semi-supervised semantic segmentation,categorizing and discussing existing methodologies. Firstly,it outlines prevalent benchmark datasets,experimental protocols,and evaluation metrics in semi-supervised semantic segmentation. Secondly,it categorizes deep learning-based semi-supervised semantic segmentation algorithms into four paradigms:adversarial learning,multi-network architecture,multi-stage architecture,and single-stage end-to-end architecture. Then,it conducts equitable comparative experiments across various representative methodologies on mainstream benchmark datasets. Finally,it discusses the challenges confronting semi-supervised semantic segmentation tasks and potential avenues for future research,encompassing aspects like fundamental model structures,latent capabilities of single-stage end-to-end methods,long-tail distribution issues of labeled data,and integration with advanced large-scale models.
万方数据
维普
