At present, the visual detection method based on deep learning is more and more widely used in the field of surface ship target detection. Based on the investigation and analysis of the structure and characteristics of the mainstream one-step and two-step models, in order to further improve the detection performance of ship targets, the mainstream YOLOv3 feature extraction network DarkNet50 is used to obtain image features, and FPN is added. The network structure fuses features to extract the deep and shallow semantic information in the network, and adds an attention mechanism module to further optimize the network performance. The improved Attention-YOLOv3 model is applied to the detection of ships on the sea surface, using the collected coco format ship target data set for training, and using the images in the actual scene for testing. The experimental results show that the improved Attention-YOLOv3 network has better performance and higher recognition accuracy than the original detection network model.