The development of ship intelligence puts forward higher demands on the real-time object detection capability of ship vision perception systems. YOLOv5, the latest achievement of the YOLO (You Only Look Once) series of algorithms, is widely used for object detection at sea with good speed and accuracy. However, in actual sea navigation, it is often accompanied by variable natural conditions and complex activity scenarios, which makes its ability to detect small objects and multi-target classification in complex waters unsatisfactory. Therefore, to improve the target detection capability of YOLOv5 in complex seas, this paper proposes a Multi-Path Aggregation Network (MPANet) structure. MPANet enhances multi-scale localization capability by fusing multi-level feature information in the bottom-up feature transfer process，and enhances higher-order feature semantic information by combining the SimAM attention module and Transformer structure. The experimental results of the custom dataset show that MPANet-YOLOv5 improves AP by 5.4%, recall by 3.3%, AP0.5 by 3.3%, and AP0.5:0.95 by 2.2%，compared with the YOLOv5 model. The results of different sea area tests show that MPANet-YOLOv5 has significantly better detection capability for small objects on the sea surface than YOLOv5.