基于多网络深度学习图像分类模型的进境木材常见害虫鉴定识别系统研究  

Research on the identification and recognition system of common pests in imported timber based on multi-network deep learning image classification model

作　　者：夏来群 翁伟钊 王新国 蔡文天 谭思明 周琳 Xia Laiqun;Weng Weizhao;Wang Xinguo;Cai Wentian;Tan Siming;Zhou Lin(China E-port Data Center Huangpu Branch/Guangdong Provincial Key Laboratory of Intelligent Port Security Inspection,Guangzhou 510700,China;School of Computer Science and Engineering,South China University of Technology;Huangpu Customs District Technical Center)

机构地区：[1]中国电子口岸数据中心黄埔分中心/广东省口岸安全智能化检测重点实验室,广东广州510700 [2]华南理工大学计算机科学与工程学院 [3]黄埔海关技术中心

出　　处：《植物检疫》2025年第2期24-31,共8页Plant Quarantine

基　　金：广东省口岸安全智能化检测重点实验室资助(2023B1212010011)。

摘　　要：进境原木和木材常会携带大量林木害虫,为提升进境木材常见害虫快速鉴别能力以及海关初筛实验室对进境木材中害虫的智能精准识别、快速处理能力,本研究采用多网络深度学习图像分类训练方法,综合利用卷积网络(CNN)、注意力机制(Transformer)网络、生成对抗网络(GAN)、特征复制器对抗训练网络,选取进境木材中截获的7种常见害虫,拍摄1200多张图像进行训练实验,构建了昆虫鉴定模型。实验研究表明,本模型对7种害虫的鉴定识别准确率达到99.17%,并基于该方法,设计了一套以人工智能模型为“大脑”的进境木材常见害虫智能鉴定识别系统。Incoming logs and timber often carry large quantities of forest pests.To enhance the rapid identification capability of common pests in imported timber and improve the intelligent recognition and rapid processing efficiency of customs primary screening laboratories,this study adopted a multi-network deep learning image classification training method integrating convolutional neural networks(CNN),attention mechanism(Transformer)networks,generative adversarial networks(GAN),and feature replicator adversarial training networks.Over 1200 images of seven common intercepted pests in imported timber were captured for training experiments,and an insect identification model was developed.Experimental results demonstrated that the identification accuracy rate of this model for 7 kinds of pests had reached 99.17%.Based on this methodology,an intelligent pest identification system for imported timber was designed,featuring an artificial intelligence model as its"core brain."

关 键 词：进境木材害虫 多网络深度学习图像分类方法 开集图像分类 鉴定识别 

分 类 号：S41[农业科学—植物保护]