YOLO-Underwater-Tiny: High-Efficiency Object Detection in Underwater Robots

Huilin Ge, Zhiyu Zhu, Biao Wang, Zhiwen Qiu

Abstract: This paper proposes the YOLOv7-underwater-tiny model for real-time and lightweight underwater target detection, focusing on applications in underwater robotics. Validated on a real-world underwater biological dataset, the model maintains YOLOv7's high detection efficiency and stability while significantly reducing inference speed and parameters, meeting real-time requirements on edge devices. Contributions include introducing Mobile One's re-parameterization for scaling, enhancing performance, and using Ghost modules to reduce computational costs without compromising accuracy. Experiments demonstrate the model's superiority, achieving efficient underwater target detection and recognition, suitable for real-time underwater robotic applications.