Multi-branch Underwater Scene Semantic Segmentation by Fusing Depth Information and Enhanced Visual Feature

Underwater image processing plays a crucial role in exploration of marine resources. Nevertheless, the presence of various trace elements and suspended particulates underwater leads to significant absorption and scattering of light, posing serious challenges for underwater scene semantic segmentation. In this work, we adopt the semantic segmentation network DeepLabv3+ as foundational framework, fusing depth information and enhanced visual feature to design a multi-branch underwater image semantic segmentation model. The proposed network incorporates a feature refinement fusion module to effectively utilize object edges, texture information, and depth features. Furthermore, integrating multi-level semantic information maintains a broader range of object and boundary features, thus improve model robustness. Extensive experiments validate the exceptional performance of our model, showcasing excellence in both objective evaluation metrics and subjective segmentation visualizations.