In order to solve the deployment difficulties caused by low detection accuracy, large number of common model parameters and slow calculation speed of deep learning target detection models deployed on offshore unmanned platforms such as buoys. Based on the YOLOv8n model, this paper proposes an improved lightweight ship detection model YOLO-FE. Firstly, the BottleBlock of C2f module is replaced by FasterN et Block in the FasterNet module to reduce the complexity of the model and improve the calculation speed of the model. Secondly, the EMA attention mechanism is introduced to further save the computational cost, and the global context information can be better obtained by using the EMA attention mechanism. Finally, the experimental results show that on the self-made ship data set, compared with the original model YOLOv8 $n$, YOLO-FE has the P, mAP0.5-0.95 increased by 2.1%and 0.2^%respectively, and the model parameters and GFLOPs reduced by 25.8%and 19.7%respectively. Compared with the common lightweight version of YOLO model, the YOLO-FE model is more lightweight and requires less computing resources while maintaining high accuracy for ship identification.