Improved Ship Detection in Maritime Environments Using Mixed Convolution and Coordinated Attention

Abstract

Accurate and real-time ship detection is essential for maritime monitoring applications. Although YOLO-based detectors provide high efficiency, their performance remains limited in complex maritime environments. This paper proposes an improved detection framework, termed YOLO-Ship, based on the YOLOv5 architecture. The model integrates mixed convolution and coordinated attention mechanisms to enhance feature representation. In addition, focal loss and complete IoU loss are adopted to optimize classification and localization accuracy. Experiments conducted on the Ship7000 dataset demonstrate that the proposed method outperforms YOLOv3 and YOLOv5 under multiple IoU thresholds, achieving an average accuracy improvement from 0.514 to 0.728. The results indicate that YOLO-Ship offers an effective balance between detection precision and computational efficiency for real-time ship recognition.