<![CDATA[Ship surveillance plays a crucial role in enhancing defense systems in coastal areas. An automatic vessel detection system is necessary to accurately identify vessels and their categories, typically utilizing a reliable computer vision system. The nano version of YOLO11 has emerged as one of the object detection methods that officially provides lightweight computing, but still has limitations in extracting complex features. Contextual Transformer (CoT) efficiently utilizes long-range relationships, thereby enhancing feature discrimination performance. This study proposes a vessel detection system by modifying the YOLO11 architecture using the Contextual Transformer block. This work introduces YOLO11-Pico, a lighter version of nano, with channel size adjustments at certain stages for further efficiency. The proposed CoT block applies fewer multiplication mapping operations, which are capable of representing global features to obtain richer contextual information. The SeaShips dataset is used as the source of data for model training and evaluation. Experimental results demonstrate that the proposed model YOLO11-pico-CoT achieves superior performance compared to prominent lightweight YOLO architectures, including the YOLO11n baseline, YOLOv5n, YOLOv10n, and the latest YOLOv12n. The integration of CoT contributes positively to improving the accuracy of ship category and location predictions, achieving 0.964 mAP50 and 0.714 mAP50:95. Additionally, efficiency evaluations show that the proposed module is computationally lighter and has fewer parameters, specifically 1,711,250 parameters while operating at 3.97 FPS, giving it an advantage in terms of capabilities over the comparison methods.]]>
Copyrights © 2025
