Hierarchical Ship Classification Employing a Multi-Level Framework Based on YOLOv11: Enhancing Accuracy in Similar Class Differentiation

This work proposes a hierarchical approach for ship classification in optical images, aiming to improve the separation between classes with high visual similarity. The strategy employs two stages based on the YOLOv11 architecture, with the first stage using a generalist detector that groups similar classes to reduce classification complexity and the second stage applying a specialized classifier to distinguish between subcategories. Experiments conducted with the InaTechShips dataset showed that the hierarchical framework increased the mAP from 96.3% to 98.6% and improved classification accuracy for the similar classes from 83.46% to 91.03%.