OptWake-YOLO: A Lightweight and Efficient Ship Wake Detection Model Based on Optical Remote Sensing Images

Runxi QiuNan Bi^*Chaoyue Yin

• Northeast Electric Power University, Jilin, China

Ship wakes on the sea surface exhibit more distinctive characteristics compared to the vessels themselves, making wake detection more feasible than direct ship detection. However, wake detection faces challenges due to interference from sea surface back-grounds, meteorological conditions, and coastal structures, while practical applications demand lighter models with faster detection speeds. This paper puts forward OptWake-YOLO, a lightweight ship wake detection model. The model incorporates three innovative designs: First, a lightweight multi-branch structure RCEA is adopted in the Backbone, combining an efficient layer aggregation network with the RepConv reparameterization module, which transforms the multi-branch structure during training into a single structure during inference, significantly enhancing feature extraction capabilities. Second, an Adaptive Dynamic Feature Fusion Network (ADFFN) is designed for the Neck, integrating channel attention mechanisms with Dynamic Upsampling technology (Dysample) to efficiently fuse multi-scale features. Finally, a Shared Lightweight Object Detection Head (SLODH) is redesigned, substantially reducing model parameters and computational complexity through parameter sharing and Group Normalization operations. Experiments on the public SWIM dataset demonstrate that, in comparison with the YOLOv11n, the proposed model enhances detection accuracy (mAP50 and mAP50-95) by 1.5% and 2.9%, respectively, whereas it concomitantly decreases parameters by 40.7% and computational load by 25.8%, maintaining high detection speed. Extensive ablation experiments and comparative analyses prove the reliability and superiority of the proposed model in the face of diverse and intricate maritime conditions, providing an effective solution for real-time ship wake detection.