YOLO-FAST for Maritime Surveillance: A Lightweight Approach to Ship Detection

Abstract

Ship detection in maritime waters plays a crucial role in ensuring transportation safety, monitoring traffic, and facilitating faster responses during naval accidents. Traditional methods are less effective because they depend heavily on human resources. To address this issue, automated detection systems are required to identify ships and their categories more accurately. Deep learning, particularly Convolutional Neural Networks (CNN), has achieved significant success in object detection, with YOLO (You Only Look Once) widely recognized as a fast and reliable real-time method. YOLO-FAST is one of the latest improvements, offering a good balance between model size and accuracy. For this research, the SeaShips dataset is used for training since it features various ship types and realistic sea conditions. Experimental results show that the proposed model achieves 94.6% mAP50 and 66.9% mAP50:95 while maintaining a lightweight structure with only 2,306,818 parameters and a computational cost of 6.4 GFLOPs. These findings highlight a strong balance between detection accuracy and computational efficiency, making YOLO-FAST suitable for deployment on low-cost devices to support scalable, real-time ship monitoring in maritime environments.