RYOLO-LWMD-Lite: A Lightweight Rotating Ship Target Detection Model for Optical Remote Sensing Images
Zhaohui Li, Sheng Qi, Haohao Yang, Haolin Li, Hongyu Jia · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2025
Combining optical remote sensing images for ship monitoring is a practical approach for maritime surveillance. However, existing research lacks sufficient detection accuracy and fails to consider computational resource constraints in ship detection processing. This paper proposes a novel lightweight rotating ship target detection model. First, we enhance the detection accuracy by expanding the YOLOv8n-obb model with Large Selective Kernel (LSK) attention mechanism, Weight-Fusion Multi-Branch Auxiliary FPN (WFMAFPN), and Dynamic Task-Aligned Detection Head (DTAH). Specifically, the LSK attention mechanism dynamically adjusts the receptive field, effectively capturing multi-scale features. The WFMAFPN improves the capacity of feature fusion by the multi-directional paths and adaptive weight assignment to individual feature maps. The DTAH further enhances detection performance by improving task interaction between classification and localization. Second, we reduce the computational resource consumption of our model. This technique is developed by pruning based on layer adaptive magnitude on the enhanced architecture and designing the DTAH module with shared parameters. Considering the above improvement, we name our model RYOLO-LWMD-Lite. Finally, we constructed a large-scale dataset for rotating ships, named AShipClass9, with diverse ship categories to evaluate our model. Experimental results indicate that the RYOLO-LWMD-Lite model achieves higher detection accuracy while maintaining a lower parameter count. Specifically, the model's parameter count is approximately 2/3 that of YOLOv8n-obb, and the test accuracy on AShipClass9 reaches 48.2% (in terms of AP50), a 6% improvement over the baseline. In addition, experiments conducted on the DOTA1.5 dataset validate the generalization capability of the proposed model.The source code is available athttps://github.com/QSuser/RYOLO-LWMD.