Semi-supervised evidential fusion for robust ship segmentation

Pixel-level ship image segmentation remains challenging due to the high cost of manual annotation and the complex maritime environment. To reduce reliance on labeled data, this study proposes SEEviF-Net, a semi-supervised framework based on Dempster-Shafer Evidence Theory. The model follows a sequential processing flow: first, it extracts hierarchical semantic features from input images via the modified DeepLabV3+ backbone, capturing both global context and local details. Next, a parallel network structure operates on the extracted features : the probability computation module uses Softmax to generate pixel-wise class probability distributions, while the confidence evaluation module based on evidential neural networks outputs mass functions that explicitly quantify epistemic uncertainty. Subsequently, Dempster’s rule fuses these two complementary evidence sources, resolving conflicts between probabilistic and evidential information to produce robust segmentation results. SEEviF-Net achieves an mIoU of 63.22% on the HRSID dataset, and 67.69% on the Small ShipInsSeg dataset under the setting of only 20% labeled data. Compared to the state-of-the-art semi-supervised baseline SemiSegSAR, these results represent improvements of 1.54% and 0.12% in average, respectively. These findings highlight the effectiveness of the proposed SEEviF-Net in remote sensing ship segmentation under limited annotation scenarios.