一种复杂背景下的苹果树病害多维动态自适应检测模型

Multi-Dimensional Dynamic Adaptive Apple tree Disease Detection Model for Complex Backgrounds

针对目前苹果树病害数据集病害种类不全以及图片背景单一的问题,本文构建了复杂背景下包括苹果树不同部位7种常见病害的苹果树病害数据集。针对目前主流苹果树病害检测模型检测精度不高,模型复杂等问题,提出了一种基于YOLOv8-NANO改进得到的复杂背景下的苹果树病害多维动态自适应检测模型(MDD-YOLOv8)。首先,设计并使用了高效多尺度注意力卷积(EMAC)模块来更换骨干网络(Backbone)中的卷积(Conv)模块来捕捉细节信息,以提高模型的特征提取能力。其次,设计了ODC2F模块并用其更换颈部网络(Neck)中的所有跨阶段局部层卷积(C2F)模块,减少模型的参数量和计算量的同时提高了模型的检测精度。最后引入SIoU损失函数(SIoU Loss)来提高模型对于苹果果实或叶片上密集小目标病害的检测性能。在自制数据集上MDD-YOLOv8相对原始YOLOv8-NANO模型平均精度均值(mAP)提升了6.8%,同时模型的参数量和计算量分别下降了8.54%和13.7%。对比主流的目标检测模型,MDD-YOLO也具有一定的优越性。

To address the issues of disease types and the uniformity of background images in the existing apple tree disease dataset,we constructed a new dataset includes seven common diseases of apple trees in different parts of the tree under complex backgrounds.Additionally,we designed MDD-YOLOv8(Multi-Dimensional Dynamic-YOLOv8)based on YOLOv8-NANO(You Only Look Once v8-NANO)for the detection of apple tree disease,aiming to solve the problems of low accuracy and complex models.Firstly,we designed a novel EMAC module to replace the Conv module in the backbone network to capture fine-grained information.This improved the feature extraction ability of the model.Secondly,the ODC2F module was designed to replace all C2F modules in the neck network.This reduced the model's parameters and computation while improving the model's detection accuracy.Finally,the SloU loss was introduced to improve the detection performance of the model for dense small target diseases on apple fruits or leaves.On the self-made dataset,MDD-YOLOv8 achieved a 6.8%mAP improvement over the original YOLOv8-NANO model,while the model's parameters and computation decreased by 8.54%and 13.7%,respectively.MDD-YOLOv8 also exhibits a certain degree of superiority compared to mainstream apple tree disease detection models.