Semi-supervised oil spill detection of SAR images based on pseudo-labelling

Semantic segmentation is an emerging topic in the processing of high-resolution Synthetic Aperture Radar (SAR) images of oil spill. Due to the complexity of information in SAR images of oil spill, the efficacy of most semantic segmentation methods for images of oil spill hinges on the quantity of labels; however, annotating images demands considerable time and manual effort. To address this issue, we propose a semi-supervised semantic segmentation approach based on pseudolabeling. This method amalgamates multiple functions to optimize the model. Specifically, we introduce DeepLab v3 as the backbone network, combined with a dual-student framework, to enhance the model's resilience to image noise through consistency and stability constraints. Moreover, we employ per-pixel contrastive learning to explore the global semantic relationships between sample pixels. By using a truncated Gaussian function to weight pseudo-labels, we reduce the entropy in pseudo-labels. Experimental results on public datasets demonstrate that our method can fully comprehend the latent information in unlabeled images and reduce the number of labeled images required while ensuring accuracy.