Ship target detection is the most critical step in the field of intelligent port. a theory YOLOv5 about multi-category and multi-target ship detection under complex port is proposed. Firstly, the feature layer of low resolution is removed and the multiple features are integrated into the lower resolution layer, which reduce the number of parameters, then the GAM is introduced into the network. Secondly, we propose a new loss function SCYLLA Efficient Intersection over Union (SEIoU) to accelerate the convergence speed. Finally, the NMS is prone to miss detection, the Soft-NMS algorithm is introduced to further screen the overlapping target box. The model improves the accuracy of ship target by 6.8%, which makes it superior to other target detection models and suitable for complex ports.