YOLOV8-ADS: A Lightweight Algorithm for Target Detection in Small-Area Unmanned Surface Vehicle Operations

Detecting surface targets is a significant challenge in unmanned surface vehicles (USVs) perception systems. Both detection accuracy and processing speed are crucial for effective deployment, yet current algorithms often struggle to identify floating objects with the necessary precision and efficiency. This paper proposes an improved, lightweight YOLOv8-ADS algorithm tailored for detecting floating objects in small urban waterways. Firstly, to reduce the interference from water surface noise, the ADown module is introduced, merging average and max pooling strengths. This method balances local averaging and maximum pooling, minimizing noise-induced instability and enhancing feature map robustness. Secondly, to preserve critical image details during upsampling, the standard upsampling module is replaced with the DySample module, which dynamically adjusts sampling strategies to produce clearer fused feature maps, thus improving detection accuracy of small targets in aquatic environments. Additionally, to address the challenges of small targets, varying scales, and complex backgrounds in urban water areas, the SEAM attention mechanism is added to the detection head to focus on essential features while reducing model parameters. Finally, the Layer-wise Adaptive Magnitude-based Pruning (LAMP) technique is employed to prune the optimized model, leading to a further decrease in its computational requirements. The findings from the Multi-Class Floating Waste (MCFW) dataset indicate that the proposed method significantly improves [email protected] by 5.2% and [email protected]:0.95 by 4.7% when compared to the baseline model. Additionally, it achieves a reduction in GFLOPs by 34.6% and lowers the number of model parameters by 33.9%. The refined algorithm establishes a harmonious equilibrium between detection precision and computational intricacy, rendering it particularly adept at identifying diminutive targets on water surfaces.