基于改进YOLO v10的复杂田间杂草检测模型

Improved YOLO v10-based Detection Model for Weeds in Complex Field Environments

摘要:

针对复杂农田环境下因遮挡与光照变化导致杂草检测精度低的问题，本文构建并扩充杂草图像数据集，并提出一种基于YOLO v10的改进检测模型SDE-YOLO v10。以YOLO v10模型为基准，采用激活函数为ReLu的SimSPPF（Simplified spatial pyramid pooling fast）替换骨干网络原有池化结构，以增强模型对多尺度杂草特征的提取能力;在颈部网络（Neck）C2f处引入可变形卷积网络v3（Deformable convolutional networks v3, DCNv3）构建C2f_DCNv3模块，提升模型对不规则形态杂草的自适应感知;嵌入具有空间和通道注意力的EMA（Efficient multi-scale attention）注意力机制，增强模型对图像不同特征维度的关注与加权能力。在自主构建并扩充的杂草数据集（3126幅）上进行试验，消融试验与多模型性能试验结果表明，SDE-YOLO v10模型精确率、召回率、mAP@0.5和mAP@0.5:0.95最高，分别为91.3%、80.2%、86.7%和73.1%，相比YOLO v10基线模型分别提升5.4、4.7、1.5、2.3个百分点。在光照变化和遮挡条件下，改进模型检测成功率分别为86.42%和82.72%，较原模型分别提高13.58、12.35个百分点，表明在复杂条件下具有良好鲁棒性。SDE-YOLO v10能实现对复杂形态杂草的准确识别，同时优化了网络结构，为智慧农业场景下实时杂草检测提供了技术参考。

Abstract:

Aiming to address the issue of low weed detection accuracy caused by occlusion and varying lighting conditions in complex farmland environments, a weed image dataset was constructed and augmented and an improved detection model, SDE-YOLO v10, was proposed based on YOLO v10. Firstly, the original pooling structure in the backbone network was replaced with the simplified spatial pyramid pooling fast (SimSPPF) module to enhance the extraction of multi-scale weed features. Secondly, the deformable convolutional networks v3 (DCNv3) was introduced into the C2f module in the neck network to improve adaptive perception of irregularly shaped weeds. Meanwhile, the efficient multi-scale attention (EMA) mechanism was embedded to strengthen the model's focus on key feature dimensions. Experiments were conducted on the self-constructed and augmented weed dataset (3126 images). Results from ablation experiments and multi-model comparisons indicated that the SDE-YOLO v10 model achieved the highest precision, recall, mAP@0.5, and mAP@0.5:0.95 scores of 91.3%, 80.2%, 86.7% and 73.1%, respectively. These represent improvements of 5.4, 4.7, 1.5 and 2.3 percentage points over the baseline YOLO v10 model. Under conditions of varying lighting and occlusion, the improved model attained detection success rates of 86.42% and 82.72%, corresponding to improvements of 13.58 and 12.35 percentage points compared with that of the original model, demonstrating its strong robustness in complex scenarios. SDE-YOLO v10 enabled accurate identification of weeds with complex morphology while optimizing the network architecture, providing a technical reference for real-time weed detection in smart agriculture scenarios.