| 摘要: |
| 【目的】建立改进的轻量化玉米苗期杂草识别模型，为玉米苗期杂草的快速精准识别和激光除草等提供技术支持。【方法】提出了一种基于YOLOv8s网络框架改进的轻量化玉米苗期杂草识别模型YOLOv8s GEF，将轻量级网络EfficientViT引入YOLOv8主干网络，以降低模型参数量；在颈部网络中引入全局注意力机制（GAM），降低玉米苗期图片信息弥散、放大全局交互，从而提高深度神经网络对玉米杂草识别的精准度；运用基于动态非单调聚焦机制的边界框损失（WIoU）替代CIoU损失函数，减少玉米苗期图片数据集中低质量示例对识别结果的影响；采用结合跨阶段部分卷积的空间金字塔池化（SPPCSPC）替换YOLOv8中的SPPF，提高模型对玉米、杂草及土壤背景中对象的识别精度。【结果】改进后的YOLOv8s-GEF模型对玉米苗期杂草识别的平均精度值（mAP@0.5）达到了78%，相较于原YOLOv8s模型提升了3.3个百分点，计算量降低了52.5%。与经典模型YOLOX、YOLOv5s、YOLOv6s、YOLOv7和YOLOv8s相比，YOLOv8s-GEF模型的参数量分别降低了36.7%，18.6%，66.9%，84.6%和48.7%，平均精度值分别提高了8.6，4.9，5.4，4.7和3.3个百分点。【结论】YOLOv8s-GEF模型在参数量控制、检测速度以及检测精度方面较其他经典目标检测算法有显著提升，可实现田间玉米苗期植株与杂草的实时精准识别。 |
| 关键词: 玉米苗期 杂草识别 YOLOv8 模型优化 |
| DOI：10.13207/j.cnki.jnwafu.2025.11.006 |
| 分类号: |
| 基金项目:国家自然科学基金项目“玉米光谱图像特征模型构建及氮截获机制研究”（32360432） |
|
| Lightweight maize seedling weed identification model based on improved YOLOv8s |
|
QI Yongjie, MA Jilong, HAN Linpu, GAO Jiaqi, JIA Biao
|
|
College of Agriculture,Ningxia University,Yinchuan,Ningxia 750021,China
|
| Abstract: |
| 【Objective】The study aimed to establish an improved lightweight model for identifying weeds in the seedling stage of maize,providing technical support for the rapid and accurate identification of weeds in the seedling stage of maize,as well as for laser weeding,etc.【Method】A lightweight model for identifying weeds in the seedling stage of maize,named YOLOv8s-GEF,which was improved based on the YOLOv8s network framework,was proposed.The lightweight network EfficientViT was introduced into the backbone network of YOLOv8 to reduce the number of model parameters.The Global Attention Mechanism (GAM) was introduced into the neck network to reduce the diffusion of image information in the seedling stage of maize and amplify the global interaction,thereby improving the accuracy of the deep neural network in identifying maize weeds.The bounding box loss (WIoU) based on the dynamic non-monotonic focusing mechanism was used to replace the CIoU loss function,reducing the impact of low-quality examples in the image dataset of the seedling stage of maize on the recognition results.The spatial pyramid pooling with cross stage partial convolution (SPPCSPC) combined with cross-stage partial convolution was used to replace the SPPF in YOLOv8,improving the model’s accuracy in identifying objects in maize,weeds,and the soil background.【Result】The mean average precision (MAP) of the improved YOLOv8s-GEF model for identifying weeds in the seedling stage of maize reached 78%,which was increased by 3.3 percentage points than that of the original YOLOv8s model,and the computational load was reduced by 52.5%.Compared with the algorithms of classic models such as YOLOX,YOLOv5s,YOLOv6s,YOLOv7,and YOLOv8s,the number of parameters of the YOLOv8s-GEF model was reduced by 36.7%,18.6%,66.9%,84.6%,and 48.7% respectively,and the mean average precision values increased 8.6,4.9,5.4,4.7,and 3.3 percentage points respectively.【Conclusion】The YOLOv8s-GEF model has significant improvements in terms of parameter quantity control,detection speed,and detection accuracy compared with other classic object detection algorithms,and can achieve real time and accurate identification of plants and weeds in the field during the seedling stage of maize. |
| Key words: maize seedling weed identification YOLOv8 model optimization |
