This paper addresses the challenges of limited detection accuracy for small objects and inefficient model deployment in complex water surface scenarios by proposing an enhanced object detection algorithm based on YOLOv8n, referred to as DSH-YOLO. This algorithm achieves a balanced optimization of accuracy and efficiency through a lightweight architecture design and multi-scale feature enhancement strategies. By integrating the DualConv module into the backbone, it leverages parallel \(3\times 3\) and \(1\times 1\) convolution kernels to combine group and heterogeneous convolutions, enabling lightweight feature extraction with reduced parameters while preserving feature representation. The Slim-neck in the neck network incorporates generalized spatial convolution and lightweight bottleneck structures, constructing a multi-scale feature fusion mechanism that reduces computational load and maintains high detection accuracy. To enhance small-object detection, a high-resolution detection head is added to capture fine-grained low-level features. Additionally, a dataset tailored for water surface environments with diverse scales, lighting conditions, and wave interference, providing precise annotations, is introduced. Experimental results on this dataset show that the proposed algorithm outperforms baseline models across multiple metrics while maintaining real-time detection speed, validating its effectiveness.
