<p>The field of underwater image vision focuses on analyzing and understanding underwater images, supporting ocean resource exploration, development, and conservation. Semantic segmentation plays a critical role in scenarios like underwater robot vision and autonomous underwater vehicles. However, due to challenges such as light attenuation, scattering, and marine snow noise, the quality of underwater images decreases, and directly applying natural image segmentation algorithms can lead to reduced performance. To address the performance issues in underwater image segmentation, this study introduces the Underwater High-Resolution Semantic Segmentation Network (UHRS-Net), which is based on HRNetV2. The main components include: (1) Wavelet Dynamic Attention Module Feature Extraction Module: It uses wavelet transform convolution to enhance noise resistance, dynamic gating to suppress irrelevant features, and an Rectangular Self-calibration Attention module to improve performance in complex scenes. To further boost segmentation performance, we introduce two auxiliary components: (2) Multi-Scale Patch Segmentation Head: This head combines global and local features and dynamically selects key information. (3)Sliding Composite Loss Function: This is based on a variable-weight Focal loss and Dice loss, optimizing semantic information at object boundaries. Experiments show that UHRS-Net achieves mIoU and mPA of 67.43 and 81.92%, respectively, on the SUIM dataset, outperforming the baseline by 6.72 and 7.47%. On the DUT-USEG dataset, it achieves a mIoU of 70.09%, an improvement of 3.37%. The model effectively enhances boundary segmentation accuracy and improves performance in complex underwater scenes.</p>