Unmanned surface vehicles (USVs) have been extensively used in various dangerous maritime tasks. Vision-based sea surface object detection algorithms can improve the environment perception abilities of USVs. In recent years, the object detection algorithms based on neural networks have greatly enhanced the accuracy and speed of object detection. However, the balance between speed and accuracy is a difficulty in the application of object detection algorithms for USVs. Most of the existing object detection algorithms have limited performance when they are applied in the object detection technology for USVs. Therefore, a sea surface object detection algorithm based on You Only Look Once v4 (YOLO v4) was proposed. Reverse Depthwise Separable Convolution (RDSC) was developed and applied to the backbone network and feature fusion network of YOLO v4. The number of weights of the improved YOLO v4 is reduced by more than 40% compared with the original number. A large number of ablation experiments were conducted on the improved YOLO v4 in the sea ship dataset SeaShips and a buoy dataset SeaBuoys. The experimental results showed that the detection speed of the improved YOLO v4 increased by more than 20%, and mAP increased by 1.78% and 0.95%, respectively, in the two datasets. The improved YOLO v4 effectively improved the speed and accuracy in the sea surface object detection task. The improved YOLO v4 algorithm fused with RDSC has a smaller network size and better real-time performance. It can be easily applied in the hardware platforms with weak computing power and has shown great application potential in the sea surface object detection.

Countries around the world have paid increasing attention to the issue of marine security, and sea target detection is a key task to ensure marine safety. Therefore, it is of great significance to propose an efficient and accurate sea-surface target detection algorithm. The anchor-setting method of the traditional YOLO v3 only uses the degree of overlap between the anchor and the ground-truth box as the standard. As a result, the information of some feature maps cannot be used, and the required accuracy of target detection is hard to achieve in a complex sea environment. Therefore, two new anchor-setting methods for the visual detection of sea targets were proposed in this paper: the average method and the select-all method. In addition, cross PANet, a feature fusion structure for cross-feature maps was developed and was used to obtain a better baseline cross YOLO v3, where different anchor-setting methods were combined with a focal loss for experimental comparison in the datasets of sea buoys and existing sea ships, SeaBuoys and SeaShips, respectively. The results showed that the method proposed in this paper could significantly improve the accuracy of YOLO v3 in detecting sea-surface targets, and the highest value of mAP in the two datasets is 98.37% and 90.58%, respectively.