文献类型：期刊文献

中文题名：基于深度学习的5种树皮纹理图像识别研究

英文题名：Research on image recognition of five bark texture images based on deep learning

作者：刘嘉政[1] 王雪峰[1] 王甜[1]

第一作者：刘嘉政

机构：[1]中国林业科学研究院资源信息研究所

年份：2019

卷号：0

期号：4

起止页码：146-154

中文期刊名：北京林业大学学报

外文期刊名：Journal of Beijing Forestry University

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD2019_2020】；

基金：国家重点研发计划项目(2017YFC0504106)

语种：中文

中文关键词：深度学习;树皮图像;卷积神经网络;树种智能识别

外文关键词：deep learning;bark image;convolutional neural network;tree species intelligent recognition

分类号：TP391.41;TP18

摘要：【目的】针对在树皮图像识别时,现有的算法和识别过程过于复杂的问题,提出了基于深度学习的方法来对不同树种的树皮图像进行识别。【方法】本文以5种常见树种的树皮纹理图像为例,采用基于卷积神经网络的深度学习方法,将原始图像直接作为输入,通过卷积和池化层对图像的低级、高级特征进行自动提取,解决了手动提取纹理特征的困难和问题;在此基础上,对CNN模型结构进行改进,采用带Maxout的ELU激励函数来代替ReLU函数,解决模型的偏移和零梯度问题;对损失函数进行改进,通过添加规范项来优化结构参数,并使用分段常数衰减法对学习率进行动态调控;最后采用softmax分类器对图像类别进行输出。【结果】对5个树种的树皮图像共计10 000张图像进行实验,其中每类选取200张图像作为测试集。最终训练准确率达到93.80%,测试集识别准确率为97.70%。另外,为验证本文方法的可行性,与传统人工特征提取法,提取HOG特征、Gabor特征和灰度共生矩阵统计法,训练SVM分类器。通过实验比较,本文方法识别准确率最高。【结论】本文提出的基于深度学习的树皮纹理图像识别方法是可行的,提高了识别效率和精度,为树种的智能化识别提供新的参考。
[Objective]Aiming at the problem that the existing algorithm and recognition process are too complicated in bark image recognition,a method based on deep learning was proposed to identify the bark images of different tree species.[Method]Taking the bark texture image of five common tree species as an example,the deep learning method based on convolutional neural network was used to input the original image directly as input,and the low-level and high-level features of the image were performed by convolution and pooling layer.Automatic extraction solved the difficulty and problem of manually extracting texture features.On this basis,the CNN model structure was improved,and the ELU excitation function with Maxout was used instead of the ReLU function to solve the model offset and zero gradient problem.The loss function was improved,the specification parameters were added to optimize the structural parameters,and the learning rate was dynamically adjusted using the piecewise constant attenuation method.Finally,the softmax classifier was used to output the image categories.[Result]A total of 10 000 images of 5 types of bark texture images were tested,and 200 images of each type were selected as test sets.The final training accuracy rate reached 93.80%,and the test set recognition accuracy rate was 97.70%.In addition,in order to verify the feasibility of the proposed method,and the traditional artificial feature extraction method,the HOG feature,the Gabor feature and the gray level co-occurrence matrix statistical method were extracted,and the SVM classifier was trained.Through experimental comparison,the method identification accuracy was the highest.[Conclusion]The tree-skin image recognition method based on deep learning proposed in this paper is feasible,which improves the recognition efficiency and precision,and provides a new reference for the intelligent identification of tree species.