YOLOv11-MFFM-A-OBB: A Multimodal Feature Fusion and Oriented Bounding Box Optimization Framework for Ship Wake Detection

To address the detection challenges due to the complex flow patterns of ship wakes, strong background interference and limited sample sizes, this paper proposes an improved YOLOv11-MFFM-A-OBB detection framework based on the YOLOv11-OBB baseline model. A multimodal feature fusion module (MFFM) and an adaptive oriented bounding box optimization module (A-OBB) are designed. The MFFM module integrates optical texture and physical field features to enhance the expression of wake bubble clusters and turbulent flow characteristics. The A-OBB module uses Kullback-Leibler divergence (KLD) loss to solve the discontinuity problem in the regression of tilted wake boundaries. Finally, the efficient fine-tuning on the 583-image self-built wake_583 dataset is achieved by using a staged transfer learning strategy and a pre-trained model based on the 12 000-image SWIM open-source wake dataset. On the wake_583 dataset, the proposed model improves average detection accuracy mAP@0.5:0.95 from 61.3% to 68.2% and the angle estimation accuracy AngleAcc@5° by 17.5%, compared to the baseline model YOLOv11-OBB.