基于轻量型网络的无人机遥感图像中茶叶枯病检测方法  

Lightweight Network for Tea Leaf Blight Detection in UAV Remote Sensing Images

作　　者：胡根生[1] 谢一帆 鲍文霞[1] 梁栋[1] HU Gensheng;XIE Yifan;BAO Wenxia;LIANG Dong(National Engineering Research Center for Agro-Ecological Big Data Analysis and Application,Anhui University,Hefei 230601,China)

机构地区：[1]安徽大学农业生态大数据分析与应用技术国家地方联合工程研究中心,合肥230601

出　　处：《农业机械学报》2024年第4期165-175,共11页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(32372632、62273001);安徽省高等学校自然科学研究重大项目(KJ2020ZD03);安徽省自然科学基金项目(2208085MC60)。

摘　　要：针对无人机采集的茶叶枯病图像中病斑差异大,病斑和背景之间相似性高等问题,设计了一个轻量型网络LiTLBNet,用于准确、实时地检测野外茶园无人机图像中的茶叶枯病。LiTLBNet使用轻量型的M-Backbone作为骨干网络,用来提取茶叶枯病病斑的可区分特征,减少因图像中病斑的尺度、颜色和形状的巨大差异而导致的漏检。在LiTLBNet的LNeck结构中引入了SE和ECA模块,帮助网络在通道维度上学习目标的综合特征,减少因病斑和背景之间的相似性造成的误检,同时删除原基线网络最大的特征图,以减少计算量和模型大小。此外,本研究还通过旋转、加噪声、构建合成图像等方式来扩充训练样本数量,提高小样本条件下LiTLBNet网络泛化能力。实验结果表明,利用LiTLBNet检测无人机遥感图像中茶叶枯病的精度为75.1%,平均精度均值为78.5%,与YOLO v5s接近。然而,LiTLBNet内存占用量仅2.0 MB,是YOLO v5s网络的13.9%。LiTLBNet网络可用于对茶叶枯病进行实时、准确的无人机遥感监测。Aiming at the problems of large differences in disease spots and high similarity between disease spots and background in tea leaf blight(TLB) disease images collected by UAV,a lightweight network LiTLBNet for the accurate and real-time detection of TLB disease in UAV images of tea gardens in the field was designed.A lightweight M-Backbone was used to extract the distinguishing features of the TLB spots,which reduced missed detections caused by the large differences in the scales,colors,and shapes of the disease spots in the images.The SE and ECA modules were introduced into the LNeck of LiTLBNet to help the network learn more comprehensive features in the channel dimension and reduce false detections caused by the similarities between disease spots and backgrounds.The largest feature maps were deleted to reduce the calculations and the network size,and furthermore,the training samples were also augmented by rotating them by different angles,adding noise to the images,and constructing synthetic images to improve the generalization of LiTLBNet by using a small number of samples.Experimental results showed that the precision of LiTLBNet was 75.1%,and the mAP was 78.5%,which was similar to that of YOLO v5s.However,the size of LiTLBNet was only 2.0 MB,which was 13.9% of the size of YOLO v5s.The proposed method can be effectively used for the real-time and accurate UAV remote sensing monitoring of TLB disease in tea gardens with a relatively large area.

关 键 词：茶叶病害 目标检测 无人机遥感 轻量型网络 LiTLBNet 

分 类 号：S127[农业科学—农业基础科学]