【摘要】 针对洒水车智能喷洒中非机动车辆检测算法复杂度高、多尺度目标检测困难等问题，提出一种基于YOLO11n的轻量化目标检测模型。基于YOLO11n模型，设计了轻量化多尺度骨干网络，将传统下采样模块替换为轻量级自适应提取模块。提出多尺度卷积模块C3k2-MSCB，并引入大核注意力机制改进C2PSA，以减少模型的参数量，提升模型对多尺度信息的检测能力。使用基于卷积神经网络的跨尺度特征融合模块替换原模型颈部网络，进一步缩小模型大小，提高检测效率。实验结果表明，在自建的非机动车辆数据集上，相较于基准模型，改进的YOLO11n模型的的参数量、GFLOPs和模型大小分别降低了46.3%,20.3%和42.9%，同时mAP@50提升了0.3%，为洒水车智能喷洒系统提供了可靠的支持。更多还原

【Abstract】 A lightweight object detection model based on an improved YOLO11n is proposed to address the high computational complexity and the difficulty of detecting multiscale objects in the intelligent spraying of non-motor vehicles by sprinkler trucks. Building on the YOLO11n framework, a lightweight multiscale backbone network, termed lightweight multiscale backneck, is introduced, in which the conventional downsampling module is replaced with a lightweight adaptive ext module. The C3k2-multiscale convolution block incorporates a large-kernel attention mechanism, and a large separable kernel attention module is further proposed to enhance channel-and-spatial self-attention, thereby reducing model parameters and improving multiscale feature representation. In addition, convolutional neural network-based cross-scale feature fusion is employed to further reduce model size and improve detection efficiency. Experimental results demonstrate that, on a self-built non-motor vehicle dataset, the improved YOLO11n model reduces the number of parameters, GFLOPs, and model size by 46. 3%, 20. 3%, and 42. 9%, respectively. Compared with the baseline model, mAP@50 increases by 0. 3 percentage points, providing accurate and reliable support for intelligent spraying systems for sprinkler trucks.更多还原