Abstract: Aiming at the poor detection of small-scale surface targets on unmanned vessels under complex sea conditions and the sample imbalance in the unmanned vessel dataset, an improved surface target detection method for unmanned vessels with YOLOv5 is proposed. Firstly, the backbone network in YOLOv5 is replaced by the GhostNet. Secondly, the global spatial adaptive module is added to the feature extraction network to limit the effect on the spatial dimension of the detection map. Finally, in order to reduce the impact of sample imbalances, trainable noise parameters are added to the loss function based on the multi-classification cross-entropy loss function. The calculation shows that the mAP value has reached 86.1%, which is 6.9% higher than the original YOLOv5.The parameter values have been reduced by 43.4%, and the detection speed reaches 69.44 frames/s, meeting the needs of embedded devices such as unmanned vessels. The improved YOLOv5 method has better detection effect in detecting water targets in unmanned vessel viewpoint.