基于残差网络的树叶分类  

Leaf Classification Based on Residual Network

作　　者：蔡圣杰 郑成勇[1] 陈伟杰 CAI Sheng-jie;ZHENG Cheng-yong;CHEN Wei-jie(School of Mathematics and Computer Science,Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学数学与计算科学学院,广东江门529020

出　　处：《五邑大学学报（自然科学版）》2022年第1期21-27,共7页Journal of Wuyi University(Natural Science Edition)

基　　金：广东省教育科学规划课题(2021GXJK308)。

摘　　要：现有的树叶分类方法的精确率已超过90%,但可分类的树叶种类较为有限.为此,本文提出一种基于残差网络迁移学习的大规模树叶分类方法.首先使用大规模数据集预训练残差网络;然后在保留其他节点参数的基础上,对已预训练好残差网络进行部分结构调整,使之适用于树叶分类;最后,使用树叶数据集对调整后的残差网络进行再训练,以使网络具备树叶分类能力.所提算法可以兼顾准确率与运行速度.实验结果表明,所提方法可分类176种树叶,树叶分类准确率超过95.6%,且识别速度可保持在212.2帧/秒,可有效应用于农、林业.Quickly identifying plant species by taking photos of leaves is of great significance to agriculture and forestry.The accuracy of traditional leaf classification methods has exceeded 90%,but the types of leaves that can be recognized are relatively limited.To this end,a large-scale leaf classification method based on residual network transfer learning is proposed,which can classify 176 species of leaves.First,large-scale data sets are used to pre-train the residual network.Then based on retaining of other node parameters,part of the structure of the pre-trained residual network is adjusted to suit leaf classification.Finally,leaf data sets are used to retrain the adjusted residual network to give it the ability to classify leaves.The proposed algorithm takes into account both accuracy and running speed.Experimental results show that the proposed method can maintain a recognition speed of 212.2 leaves per second while the accuracy exceeds 95.6%,and can be effectively applied to agriculture and forestry.

关 键 词：树叶分类 残差网络 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]