Improved SegNeXt Algorithm for Semantic Segmentation of Underwater Targets in Optical Images

Underwater semantic segmentation is degraded by wavelength‐dependent attenuation, scattering, nonuniform illumination, and low target‐background contrast. This study improves SegNeXt by integrating efficient squeeze‐and‐excitation (eSE) channel recalibration into MSCAN and replacing the Hamburger decoder head with HGStem to preserve small‐object and boundary cues. Under a unified 300‐epoch protocol on SUIM, the proposed model achieves 84.35% aAcc, 63.05% mAcc, 37.35% mIoU, and 36.84 FPS, improving mIoU by 4.25 percentage points over SegNeXt‐tiny. On TrashCan‐Instance, it improves aAcc, mAcc, and mIoU by 0.48, 3.50, and 3.88 percentage points, respectively. Controlled ablations, complementary evaluation on two underwater datasets, qualitative segmentation, feature heatmaps, and channel‐attention distributions indicate that eSE improves feature selectivity and HGStem contributes to spatial‐detail recovery. The resulting accuracy‐efficiency balance supports real‐time underwater perception.