Key points are not available for this paper at this time.
Ship detection driven by computer vision is the key link of intelligent port management. The main contradiction currently facing is that the computational complexity of the traditional deep learning model is too high to meet the actual deployment requirements; the over-lightweight model often leads to a significant decrease in detection accuracy due to insufficient feature expression ability. To address these issues, this study proposes a lightweight multi-scale ship detection network named LMSD-Net, based on RAPNet. The model effectively combines the efficient feature extraction capabilities of RapidNet with the multi-scale path aggregation abilities of PANet, constructing a RAPNet backbone network that boasts strong computational efficiency and feature representation capabilities. A multi-branch parallel architecture for multi-scale feature fusion is designed to enhance feature learning. An efficient multi-scale attention mechanism is introduced to improve the focusing ability on key features. To address the characteristics of small ship targets, a Lighting-Occlusion Attention Mechanism, combining mixed attention and residual connections, is embedded in the small target detection head to effectively mitigate noise interference and information degradation. Experiments on the SeaShips and SSDD (SAR ship monitoring dataset) demonstrate that LMSD-Net achieves 52.2% and 73.7% mAP@50:95 with 0.5M parameters and 2.3GFLOPs computational cost, respectively. Compared to YOLO11n, which has 2.58M parameters and 6.3GFLOPs computational cost, LMSD-Net increases mAP@50:95 by 2.5% and 1.1%, reduces parameter count by 80.6%, and decreases computational cost by 63.5%, showing excellent edge deployment potential.
Nan et al. (Fri,) studied this question.