<![CDATA[Ship object detection in aerial imagery remains a critical challenge due to complex marine backgrounds, varying object scales, and occlusion, which often lead to unstable model performance. This research proposes integrating the Efficient Channel Attention (ECA) module into the YOLO12-L architecture to enhance feature selectivity and prediction robustness. The model was trained for 500 epochs on the Ship Detection from Aerial Images dataset, comprising 621 images and 1,951 annotated ship instances, and performance was evaluated across five distinct random seeds to ensure statistical reliability. Quantitative results demonstrate that the proposed YOLO12-L + ECA model achieved a median Average Precision (mAP@50) of 71.32% and a Precision of 92.5%, outperforming the baseline YOLO12-L model. To evaluate statistical validity, a Paired Bootstrap Median Test with 100 resamples confirmed a statistically significant improvement in median performance (Δ = +1.01%, p = 0.02). Furthermore, the standard deviation of mAP@50 decreased from 1.1% in the baseline to 0.3% in the ECA model, representing a 72.7% reduction in performance variance. Computational efficiency analysis revealed that the ECA module introduced negligible overhead, adding merely 5 parameters (totaling 26,389,880) and keeping FLOPs constant at 89.4, while maintaining a high inference speed of 10.7 FPS (a marginal 2.5% reduction). These findings confirm that ECA effectively suppresses background noise, stabilizes detection outputs, and provides statistically significant improvements without compromising architectural efficiency. The proposed architecture offers a lightweight and reliable solution for automated maritime monitoring systems, particularly in challenging visual environments.]]>
