To solve the problem of the large number of parameters and computation of ship target detection algorithm, as well as the difficulties of ship detection caused by the influence of the nearshore complex backgrounds and the mutual occlusion of ships in inland river environments, this paper makes improvements based on YOLOv7-tiny and proposes a lightweight algorithm MED-YOLO for ship target detection. Firstly, the MobileNetV3 network is used as the backbone feature extraction network, which greatly reduces the calculation cost of the model. Secondly, EMA attention module was introduced into the neck network, and EMA-ELAN module was constructed to enhance the multi-dimensional perception and multi-scale feature extraction capability of the network. Then, Dyhead, which combines scale perception, spatial perception, and task perception, is selected as the detection head of the improved model to obtain stronger feature expression ability. Finally, WIoU with dynamic non-monotonic focusing mechanism is used as the bounding box loss function to improve the model's ability to cope with ship occlusion and improve the detection performance. The experimental results show that compared with YOLOv7-tiny, MED-YOLO has 39.8% fewer parameters and 55.0% less computation, and its precision and mAP@0.5 have increased by 1.4% and 1.0% respectively, reaching 98.3% and 98.9%, which not only achieves lightweight, but also has better detection performance. It meets the deployment requirements in the environment with limited computing resources, and has certain practical engineering significance.