基于卷积神经网络的油茶害虫生态识别  

Ecological Identification of Camellia spp.Pests Based on Convolutional Neural Networks

作　　者：梁秀豪 杨丽萍 廖旺姣[1] 黄丽芸 陈健武 阳文林 蒙芳[1] 黄超航 韦维[1] 王国全[3] Liang Xiuhao;Yang Liping;Liao Wangjiao;Huang Liyun;Chen Jianwu;Yang Wenlin;Meng Fang;Huang Chaohang;Wei Wei;Wang Guoquan(Guangxi Forestry Research Institute,Guangxi Key Laboratory of Special Non-wood Forests Cultivation and Utilization,Guangxi Forest Pests Natural Enemies Breeding Research Center of Engineering Technology,Nanning,Guangxi 530002,China;Zhejiang Normal University,Jinhua,Zhejiang 321004,China;College of Agriculture,Guangxi University,Nanning,Guangxi 530004,China)

机构地区：[1]广西壮族自治区林业科学研究院、广西特色经济林培育与利用重点实验室广西林业有害生物天敌繁育工程技术研究中心,广西南宁530002 [2]浙江师范大学,浙江金华321004 [3]广西大学农学院,广西南宁530004

出　　处：《广西林业科学》2023年第3期361-366,共6页Guangxi Forestry Science

基　　金：广西林业科技推广示范项目(桂林科研[2022]第17号);中国—东盟(华为)人工智能创新中心补贴项目(桂数发[2022]20-1-14)。

摘　　要：害虫是影响油茶(Camellia spp.)产量的主要因素之一,对其进行准确识别有助于及时防控,减少损失。目前,油茶害虫识别研究缺少相关的数据集,限制了深度学习技术在油茶害虫识别中的应用。为给在生态环境下准确识别油茶害虫提供1种新范式,构建包含1116张7类害虫的油茶害虫识别图像数据集,采用4种目标检测算法(SSD、YOLOv3、YOLOX和RetinaNet)在该数据集上进行试验。结果表明,IOU阈值为0.5时,SSD的平均精度为93.50%,YOLOX为93.50%,RetinaNet为86.80%,YOLOv3为96.60%;SSD的平均召回率为73.20%,YOLOX为75.10%,RetinaNet为78.00%,YOLOv3为76.80%。综合分析,YOLOv3的检测和分类能力最优。Pests are one of the most significant factors affecting yields of Camellia spp.,and accurate identifica⁃tion of pests is helpful to control timely and reduce losses.However,relevant datasets were lacked in identifica⁃tion researches of Camellia spp.pests,which limited application of deep learning technology.To provide a new model for accurate identification of Camellia spp.pests in ecological environment,a Camellia spp.pests recog⁃nition dataset was constructed,which contained 1116 images of 7 classes of pests.Object detection algorithms SSD,YOLOv3,YOLOX and RetinaNet were experimented based on the dataset.Results showed that when IOU threshold was 0.5,average accuracy of SSD was 93.50%,YOLOX was 93.50%,RetinaNet was 86.80%,and YO⁃LOv3 was 96.60%;average recall of SSD was 73.20%,YOLOX was 75.10%,RetinaNet was 78.00%,and YO⁃LOv3 was 76.80%.YOLOv3 had the best abilities of detection and classification by comprehensive analysis.

关 键 词：害虫识别 目标检测算法 油茶 

分 类 号：S794.4[农业科学—林木遗传育种]