基于改进YOLOv7的航拍图像下松材线虫病疫木识别  

Improved YOLOv7 based aerial images of pine nematode epidemic wood recognition

作　　者：胡丹妮 吴红玉 叶振[3] HU Danni;WU Hongyu;YE Zhen(School of Sciences,Zhejiang Sci Tech University,Hangzhou 310018,China;Faculty of Engineering,Lishui University,Lishui 323000,China;School of Mathematics and Computer Science,Lishui University,Lishui 323000,China)

机构地区：[1]浙江理工大学理学院,杭州310018 [2]丽水学院工学院,浙江丽水323000 [3]丽水学院数学与计算机学院,丽水323000

出　　处：《林业工程学报》2025年第2期147-155,共9页Journal of Forestry Engineering

基　　金：浙江省自然科学基金探索项目(LTGN23F020001)。

摘　　要：松材线虫病是一种危害程度极高的传染性松树病害。为精确掌握大尺度范围松材线虫病疫木的数量和分布,提出了一种基于改进YOLOv7的无人机航拍图像下松材线虫病疫木检测模型。首先,针对因航拍图像背景复杂而导致的疫木错检漏检问题,在模型主干特征提取部分引入SimAM注意力机制,以便模型更好地聚焦松材线虫病疫木颜色、纹理等关键特征;其次,用ConvNeXt网络对Head部分的ELAN-W网络进行替换,以提高模型对单株疫木的特征提取效率,在降低模型参数量的同时提升模型检测速度;然后,引入SPD-Conv以提高低分辨率航拍图像下小目标的检测精度;最后,将颈部网络的卷积替换为CoordConv,以更好地感受特征图中疫木的位置信息。在自建的松材线虫病疫木数据集中进行了大量验证,结果表明:经改进后的YOLOv7模型检测精确度为91.1%,召回率为93.5%,F_(1)分数为92.3%,与原YOLOv7模型及其他当前主流模型相比,各项主要指标均有一定提升。在选取的两块不同区域测试样地上的实验结果表明,本模型具有较好的适应性,可有效应用于大尺度松材线虫病疫木普查任务中。Pine nematode disease is a highly harmful infectious disease of pine caused by pine wood nematode,which causes serious losses to forestry resources in China.To accurately comprehend the quantity and distribution of pine nematode infected trees in a large-scale range,this study proposed a model for detecting pine wilt infected trees under UAV aerial images based on the improved YOLOv7 algorithm.For the problem of wrong detection and omission of infected trees caused by the complex background of aerial images,this study introduced the SimAM attention mechanism in the main feature extraction part,so that the model can better focus on key features such as color,texture,and other key features of the pine nematode infected trees.Secondly,the ELAN-W network in the Head part was replaced with ConvNeXt network,which improved the efficiency of the model for single infected tree feature extraction,and reduced the number of model parameters and improved the model detection speed.Then,SPD-Conv was introduced to improve the detection accuracy and recall rate of small targets in low resolution aerial images.Finally,the convolution of the neck network was replaced by CoordConv to better feel the position information of the epidemic trees in the feature map.The experiment result showed that the improved YOLOv7 model could effectively detect the pine nematode infected trees,with a precision of 91.1%,a recall rate of 93.5%,and an F_(1)score of 92.3%,which were 1.6,3.4,and 2.5 percentage points higher than the original YOLOv7 model,respectively.Compared with other current mainstream models,the improved model showed some improvement in all the main indexes.In addition,in the ideal confidence threshold range,the precision,recall rate and F 1 score of the improved model were always better than the original model for the detection of two test plots with different degrees of disaster.The above experimental results showed that the improved model had good adaptability and can be effectively applied in large-scale pine nematode epidemic tre

关 键 词：松材线虫病 大尺度范围疫木识别 无人机航拍图像 目标检测 YOLOv7 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] S763[自动化与计算机技术—控制科学与工程]