DHLNet: A Dynamic Hierarchical Lightweight Network for Enhanced Ship Detection in Remote Sensing Images

Abstract

Ship detection in remote sensing images is essential for maritime surveillance and environmental monitoring. Traditional methods often struggle to accurately identify ships in complex scenes or when targets are small, and recent deep learning approaches, while promising, still face tradeoffs between detection accuracy, inference speed, and computational complexity. To overcome these limitations, we propose dynamic hybrid convolutional network (DHLNet), a novel detection model comprising three specialized modules. DHLNet includes a dynamic hybrid block module that adaptively selects convolutional kernels for multiscale feature extraction, a faster hierarchical attention fusion block that integrates local details with global context through a multilevel attention mechanism, and a lightweight quality estimation BN head that balances spatial, channel, and scale features for efficient decoding. These innovations collectively enhance feature representation and improve detection performance without significantly increasing the computational cost. Extensive experiments on a self-collected ship dataset and public benchmarks (DOTA-Ship and VisDrone2019) validate the effectiveness of DHLNet. The proposed model outperforms state-of-the-art detectors (e.g., YOLOv8, YOLO-KAN, Mamba) in both mAP50 and F1-score metrics. For example, on our dataset, DHLNet achieves an mAP50 of 91.4%, which is 2.7% higher than that of YOLO-KAN. These results demonstrate that DHLNet effectively handles complex backgrounds and small targets, offering significant improvements in detection accuracy and efficiency for remote sensing-based ship detection.