BG-UISNet: A Novel Boundary-Guided Network for Robust and Precise Underwater Image Segmentation

The exploration of marine environments is a key focus in marine biology, ecological conservation, and underwater robotics, driving the demand for precise underwater image segmentation to advance related scientific research. Yet, the inherent complexity of underwater settings, marked by light attenuation, turbidity, and the camouflage traits of marine organisms, severely hampers segmentation performance, posing challenges to meet the practical needs of marine applications. To address this, we propose BG-UISNet, a novel boundary-guided underwater image segmentation network (BG-UISNet) built upon recent advances in powerful, large-scale pretrained models. This network incorporates a boundary-aware module (BAM) to sharpen edge detection, a boundary-guided feature enhancement module (BFEM) to refine edge-related features, and a multiscale context aggregation module (MCAM) to capture global contextual details. Experiments show that the BG-UISNet outperforms existing methods across four classic underwater datasets. Ablation studies on the MAS3K and RMAS datasets further reveal average performance gains of 3.25%, 1.65%, 3.60%, and 1.00% in mean intersection over union (mIoU), ${\mathrm{S}}_{\alpha }$ , ${\mathrm{F}}_{\beta }^{w}$ , and ${\mathrm {mE}}_{\phi }$ , respectively. Moreover, the BG-UISNet demonstrates remarkable generalization ability. It also performs well on two general object segmentation tasks: camouflaged object detection (COD) and salient object detection (SOD). These results confirm its ability to effectively fuse low-level edge details with high-level global semantics. By overcoming the shortcomings of traditional convolutional neural networks (CNNs) and Transformer models in underwater contexts, the BG-UISNet offers a robust segmentation tool for marine vision tasks, paving new technical avenues for future underwater environmental studies.