A dual-branch selection method with pseudo-label based self-training for semi-supervised smoke image segmentation

Abstract

In recent years, there has been significant progress in fully supervised learning for smoke image segmentation. However, these methods rely heavily on manually annotated labels, which is difficult for the field of smoke segmentation that lacks public datasets. In this work, we address this issue by employing a semi-supervised semantic segmentation approach that iteratively assigns pseudo-labels to unlabeled data, reducing the need for smoke-labeled data. However, traditional self-training methods tend to prioritize reliable pseudo-labels and overlook the potential of unreliable pseudo-labels. At the same time, there may be false reliable pseudo-labels that introduce false supervisory signals, leading to the confirmation bias problem. To alleviate these problems, we propose a novel method called the Dual-Branch Pseudo-label Selection (DBPS). DBPS focuses on two aspects: selecting more reliable pseudo-labels and exploiting the potential of unreliable pseudo-labels. Intuitively, one branch identifies potentially incorrect regions within reliable pseudo-labels, while the other identifies potentially correct regions within unreliable pseudo-labels. This dual-branch approach ensures a more comprehensive and balanced pseudo-label selection process. In addition, we propose a multi-scale decoder representation head that leverages pixel-level contrastive learning to learn discriminative feature representations. This allows us to enable the gathering of similar samples and the scattering of dissimilar samples. The experimental results on the synthetic smoke dataset demonstrate the effectiveness of our proposed method in mitigating the noise of pseudo-labels and achieving outstanding performance, particularly in challenging scenarios with a limited number of labeled data.