基于改进YOLO v8s的小麦小穗赤霉病检测研究

Wheat Spikelet Detection of Fusarium Head Blight Based on Improved YOLO v8s

为实现大田复杂背景下小麦小穗赤霉病快速准确识别,构建了包含冬小麦开花期、灌浆期和成熟期3个生育期共计640幅的小麦赤霉病图像数据集,并提出一种基于改进YOLO v8s的小麦小穗赤霉病识别方法。首先,利用全维动态卷积ODConv替换主干网络中的标准Conv,提高网络对目标区域特征的提取;然后,在Neck网络使用改进Efficient RepGFPN特征融合网络实现低层特征与高层语义信息的融合,使模型能够提取更丰富的特征信息;最后,采用EIoU损失函数替换CIoU损失函数,加快模型收敛速度,进一步提高模型准确率,实现对小麦小穗赤霉病的快速、准确识别。在自建的数据集上进行模型验证,结果表明,改进模型(OCE-YOLO v8s)对小麦小穗赤霉病的检测精度达到98.3%,相比原模型提高2个百分点;与Faster R-CNN、CenterNet、YOLO v5s、YOLO v6s、YOLO v7模型相比分别提高36、25.7、2.1、2.6、3.9个百分点。提出的OCE-YOLO v8s模型能有效实现小麦小穗赤霉病精确检测,可为大田环境下农作物病虫害实时监测提供参考。

To achieve rapid and accurate identification of fusarium head blight on wheat spikelets in complex field background,a wheat fusarium head blight image dataset comprising a total of 640 images across three growth stages:flowering,grain filling,and ripening of winter wheat was constructed.Additionally,a wheat spikelet fusarium head blight recognition method based on an improved YOLO v8s model was proposed.Firstly,using the omni-dimensional dynamic convolution(ODConv)to replace the standard convolution in the backbone network enhanced the network's extraction of features from target regions and suppressed interference from cluttered background information.Secondly,an improved Efficient RepGFPN feature fusion network was utilized in the neck network to integrate low-level features with high-level semantic information,enabling the model to extract richer feature information.Lastly,the enhanced intersection over union(EIoU)loss function was employed instead of the complete intersection over union(CIoU)loss function to accelerate model convergence speed and further improve model accuracy,thus achieving rapid and accurate identification of fusarium head blight on wheat spikelets.Model validation on a self-built dataset revealed that the improved model(OCE-YOLO v8s)achieved a detection accuracy of 98.3%for fusarium head blight on wheat spikelets,which was an improvement of 2 percentage points compared with the original model.Compared with Faster R-CNN,CenterNet,YOLO v5s,YOLO v6s,and YOLO v7 models,the OCE-YOLO v8s model achieved improvements of 36 percentages,25.7 percentages,2.1 percentages,2.6 percentages,and 3.9 percentages,respectively.The OCE-YOLO v8s model effectively met the requirements for precise detection of fusarium head blight on wheat spikelets and could provide valuable insights for real-time monitoring of crop diseases and pests in complex backgrounds of field environments.