YOLO-CPCL: Compact Multi-Class Oriented Ship Detection with Adaptive Feature Fusion and Aspect-Ratio-Coupled Angle Supervision.

Multi-class oriented ship detection in optical remote sensing images remains challenging in densely berthed and nearshore scenes. Elongated hulls, arbitrary headings, background clutter, and similar vessel appearances can weaken feature aggregation and reduce the accuracy of rotated-box localization. This study proposes YOLO-CPCL, a compact oriented detector developed from YOLOv8n-OBB. In the final YOLO-CPCL architecture, a Ship-Oriented Slender Adaptive Fusion module (SOSA-Fuse) replaces all four C2f fusion units in the Neck. It combines learned content-adaptive sampling with a C2f-style split-and-concatenation pathway to improve feature aggregation for elongated and arbitrarily oriented ships. An Aspect-Ratio-Coupled Angle Supervision Loss (ARCAS-Loss) is further introduced by applying a bounded logarithmic aspect-ratio weight to a periodic cosine angle term. This formulation strengthens angle supervision for slender targets while limiting the influence of extreme samples. On the four-class HRSC2016 task, YOLO-CPCL improves precision, recall, mAP@50, mAP@75, and mAP@50-95 by 3.54, 7.39, 5.48, 8.60, and 7.23 percentage points, respectively. The parameter count is reduced from 3.08 M to 2.92 M, corresponding to a decrease of 5.19%, while the computational cost is reduced from 8.3 to 7.6 GFLOPs, a decrease of 8.43%. Additional evaluations on the Level-2 24-class setting of ShipRSImageNet and a custom DOTA-v1.0 protocol with multi-class training and ship-class reporting show positive aggregate gains. These results demonstrate that the proposed method improves ship recall and high-IoU oriented localization while reducing the parameter count and GFLOPs.