MO-YOLO: Enhanced YOLOv11 Lightweight Ship Detection Based on Multi-Scale Attention Mechanism
Dongjie Ju, Chenxin Zhao, Yunsheng Fan · 2025
We propose an enhanced algorithm, Mo- Yolo, based on the YOLOv11 model for ship object detection, with the aim of improving both detection accuracy and computational efficiency. A key innovation in our approach is the integration of the DynamicConv module, which replaces the standard downsampling layer. Dynamic convolution allows the model to adaptively adjust convolutional kernel parameters in response to input variations, thereby improving its generalization and robustness. Additionally, we introduce a Multi-Scale Dilated Transformer for Visual Recognition (MSDA) module, which leverages the sparsity of self-attention mechanisms at multiple scales. By applying different dilation rates across multiple heads, MSDA captures richer contextual information, enhancing the model's global perception capability. To further optimize computational efficiency for low-resource platforms, we utilize MobileNetV 4 as the backbone network, significantly reducing the model's computational complexity while maintaining high performance. Experimental results on the SeaShips and Singapore Maritime datasets demonstrate that the proposed Mo-Yolo algorithm achieves a 2.1% and 2.9% improvement in detection accuracy, respectively, compared to the baseline model. These findings indicate that Mo- Yolo not only enhances precision but also offers superior efficiency, making it well-suited for real-world ship detection applications in resource-constrained environments.