To overcome the visual degradation of optical images caused by the complexity of underwater environments and address the challenge of deploying existing high-precision models on resource-constrained devices， this study aims to enhance the optical perception and autonomous operational capabilities of intelligent fishery equipment such as autonomous underwater vehicles. This paper proposes a lightweight backbone-based underwater dual-task network， UED-Transformer （Underwater Efficient Dual-task Transformer）. The network integrates monocular depth estimation and semantic segmentation tasks， utilizes two single-task teacher models to generate pseudo-labels， and trains the UED-Transformer using the generated underwater image dataset. Experimental results on the public underwater multi-label dataset MIMIR-UW show that the model achieves a mean Intersection over Union （mIoU） of 79.3% and a pixel accuracy （PA） of 93.7% in the semantic segmentation task， while the root mean square error （RMSE） for the monocular depth estimation task is reduced to 4.975. The research indicates that this lightweight multi-task network exhibits robust multi-task performance while maintaining low computational costs， which will effectively improve the intelligent scene perception and understanding capabilities of fishery equipment in complex underwater environments. This study provides an efficient and lightweight vision algorithm system for intelligent fishery devices， reducing the dependence on high-computational hardware and promoting the practical application of artificial intelligence in the intelligentization of fishery equipment.