南京林业大学学报(自然科学版) ›› 2020, Vol. 44 ›› Issue (1): 138-144.doi: 10.3969/j.issn.1000-2006.201809004

• 研究论文 • 上一篇    下一篇

基于深度学习的树种图像自动识别

刘嘉政(), 王雪峰*(), 王甜   

  1. 中国林业科学研究院资源信息研究所,北京 100091
  • 收稿日期:2018-09-07 修回日期:2019-03-28 出版日期:2020-02-08 发布日期:2020-02-02
  • 通讯作者: 王雪峰
  • 作者简介:刘嘉政( liujiazheng0919@163.com)。
  • 基金资助:
    国家重点研发计划(2017YFC0504106)

Automatic identification of tree species based on deep learning

LIU Jiazheng(), WANG Xuefeng*(), WANG Tian   

  1. Research Institute of Resource Information Techniques, Chinese Academy of Forestry, Beijing 100091, China
  • Received:2018-09-07 Revised:2019-03-28 Online:2020-02-08 Published:2020-02-02
  • Contact: WANG Xuefeng

摘要:

【目的】为了探究深度学习方法用于林业树种图像智能识别的可行性,提出一种基于深度学习方法的自动识别树种新方法。在TensorFlow框架下,对卷积神经网络(CNN)模型进行改进,对7类树种图像进行自动识别研究。【方法】首先,在图像库建立时,为增加特征选择多样性,选择树木的树皮和树叶图像,保留自然背景;另外,考虑到同一树种在不同树龄条件下树皮图像存在差异,因此加入不同树龄的树皮图像,并用胸径指标来表示树龄大小。其次,对每类树种图像随机挑选100张作为测试集,剩余数据集全部作为训练集。通过反复试验比较不同CNN结构设置、卷积层数量、全连接层层数、学习率等对结果的影响。采用Adam算法代替传统的随机梯度下降(SGD)算法,对模型进行优化,用指数衰减法对学习率进行调节,在交叉熵函数中加入L2正则项对权重进行惩罚,并采用Dropout策略和ReLU激励函数,以避免训练过程中过拟合现象。最后,确定适合试验要求的13层CNN结构,同时比较深度学习方法和传统人工特征识别方法的差异,与已有的树种图像识别方法做对比。【结果】提出的13层树种图像识别模型,对训练集和测试集取得了理想的识别效果,识别率分别为96.78%、91.89%,在未参与训练的验证集上取得了96%的平均准确率。相对于已有的人工特征识别方法,所提出的方法识别效率和准确度更高。【结论】基于改进的卷积神经网络树种识别模型识别效果明显高于传统方法,说明所提出的方法能够应用于树种识别,可为林业树种图像自动识别提供一条新思路。

关键词: 深度学习方法, 树种图像, 卷积神经网络, 自动识别

Abstract:

【Objective】 In order to examine the feasibility of developing a deep learning method for the intelligent identification of forestry tree species, we propose a new method of automatic tree recognition tree based on deep learning. In this study, we improved a convolutional neural network model under the framework of TensorFlow, and used this to examine the automatic identification of images of seven tree species. 【Method】 Initially, when creating an image library, in order to increase the feature selection diversity, the bark and leaf images of the trees are selected to preserve the natural background. In addition, considering that the same tree species can have different bark images depending on tree age, the images of bark from trees of different ages are added, and the DBH index is used to indicate the age of the tree. Thereafter, 100 images of each tree species were randomly selected as test sets, and the remaining data sets were all used as training sets. Through repeated experiments to compare the effects of different convolutional neural network structure settings, the number of convolutional layers, the number of fully connected layers, and the learning rate on the experimental results, we propose the use the Adam algorithm instead of the traditional SGD algorithm to optimize the model and use the index. The attenuation method adjusts the learning rate. The L2 regular term is added to the cross entropy function to penalize the weight, and the Dropout strategy and ReLU excitation function are used. Finally, a 13-layer convolutional neural network structure suitable for the experimental requirements of this study was determined. In addition, in order to compare differences between the deep learning method and the traditional artificial feature recognition method, we compared performance with that of an existing tree species image recognition method.【Result】 The 13-layer tree species image recognition model proposed in this study achieved an ideal recognition effect on both training and test sets, with recognition rates of 96.78% and 91.89%, respectively. Furthermore, we obtained an average 96% recognition for the verification set that was not included in the training, and overall found the recognition efficiency and accuracy to be higher than those of the existing artificial feature recognition methods. 【Conclusion】 Compared with the traditional method, the proposed method, based on an improved convolutional neural network tree species recognition model, can be applied to tree species identification and provides a new approach for forestry tree species identification.

Key words: deep learning method, tree species image, convolution neural network, automatic identification

中图分类号: