A ship detection method based on improved YOLOv8 models and ensemble learning

To solve the challenges of reduced accuracy and robustness in maritime ship detection using a single detector, particularly in complex environments with fog interference, an ensemble learning method based on two improved YOLOv8 models is proposed. The YOLOv8 architecture is enhanced by incorporating the Coordinate Attention mechanism to improve multi-scale detection performance, and replacing the CIOU loss function with WIoU to enhance accuracy. Two independently trained YOLOv8 models are employed as base detectors, each optimized for ship detection in either foggy or clear conditions. A coordinate-weighted algorithm merges outputs from the two detectors, using ensemble learning to enhance robustness in foggy and clear conditions. A dataset of 5, 546 images, divided into foggy and clear subsets, was created and expanded through data augmentation. Experimental results demonstrate detection accuracies of 85.9% and 92.1% with recall rates of 93.0% and 95.1% for the proposed method.