Weakly Supervised Semantic Segmentation Based on Image-level Class Labels with Deep Learning: A Survey
Yijiang Wang, Fen Luo, Hongxu Zhang, Zhanqiang Huo · 2022
The training of fully supervised semantic segmentation (FSSS) networks relies on a large amount of data with pixel-level class labels, which limits semantic segmentation's application in practical scenarios. Weakly supervised semantic segmentation (WSSS) based on image-level class labels has become a new research hotspot in the field of image semantic segmentation. In this paper, the WSSS algorithms with image-level class labels are classified and sorted out according to two critical steps of the generic pipeline: seed generation and mask refinement. Different ways of seed generation can be divided into three categories: Multiple-Instance Learning-based methods, Class Activation Maps-based methods, and other generation methods. Different seed mask refinement methods can be divided into affinity-based mask refinement methods, additional supervision-based mask refinement methods, and other mask refinement methods. In addition, this paper analyzes the principles, key ideas, primary contributions, and advantages and disadvantages of various methods. Then, the segmentation performance of different WSSS algorithms on different datasets are compared, and the current state-of-the-art segmentation algorithms are labeled. Finally, the main challenges currently faced in the field and possible future directions have been prospected.