Image semantic segmentation t is widely used in aquatic product measurement, aquatic biological cell segmentation, and aquatic biological classifications. However, underwater image segmentation has low accuracy and poor robustness because of turbid underwater environments and insufficient light. Therefore, this paper proposes an Underwater Image Semantic Segmentation Network (UISS-Net) for underwater scenes. Firstly, the backbone network uses an auxiliary feature extraction network to improve the extraction of semantic features for the backbone network. Secondly, the channel attention mechanism enhances the vital attention information during feature fusion. Then, multi-stage feature input up-sampling is used to recover better semantic features in the network during up-sampling. Finally, the cross-entropy loss function and dice loss function are used to focus on the boundary semantic information of the target. The experimental results show that the network effectively improves the boundary of the target object after segmentation, avoids aliasing with other classes of pixels, improves the segmentation accuracy of the target boundary, and retains more feature information. The mIoU and mPA of UISS-Net in the semantic Segmentation of Underwater IMagery (SUIM) dataset achieve 72.09% and 80.37%, respectively, 9.68% and 7.63% higher than the baseline model. In the Deep Fish dataset, UISS-Net achieved 95.05% mIoU, 12.3% higher than the baseline model.