欢迎访问林业科学,今天是

林业科学 ›› 2021, Vol. 57 ›› Issue (6): 134-143.doi: 10.11707/j.1001-7488.20210615

• 论文与研究报告 • 上一篇    下一篇

基于卷积神经网络模型的木材宏、微观辨识方法

赵子宇1,杨霄霞1,郭慧2,葛浙东1,周玉成1,2,*   

  1. 1. 山东建筑大学信息与电气工程学院 济南 250101
    2. 中国林业科学研究院木材工业研究所 北京 100091
  • 收稿日期:2020-08-17 出版日期:2021-06-25 发布日期:2021-08-06
  • 通讯作者: 周玉成
  • 基金资助:
    泰山学者优势特色学科人才团队(2015162);基于X射线断层成像技术的树木年轮识别与分析(ZR2020QC174)

Recognition Method of Wood Macro- and Micro-Structure Based on Convolution Neural Network

Ziyu Zhao1,Xiaoxia Yang1,Hui Guo2,Zhedong Ge1,Yucheng Zhou1,2,*   

  1. 1. School of Information and Electrical Engineering, Shandong Jianzhu University Jinan 250101
    2. Research Institute of Wood Industry, CAF Beijing 100091
  • Received:2020-08-17 Online:2021-06-25 Published:2021-08-06
  • Contact: Yucheng Zhou

摘要:

目的: 提出一种基于卷积神经网络模型——PWoodIDNet模型的木材宏、微观辨识方法,以有效提高木材辨识精度和速度,为海关、进出口检疫检验、家具企业等法定部门和企业提供先进的辨识方法和仪器,推动我国木材进出口检疫检验行业和木材加工制造企业的科技进步。方法: 首先,选择16种木材样本,每种样本获取50张高分辨率显微CT图像和工业相机图像,共1 600幅;然后,截取具有木射线、薄壁组织、轴向管胞、纹孔和纹理等特征的目标区域,共4 800幅,通过水平翻转、垂直翻转、镜像、亮度变换等图像增强算法后将图像集扩充至19 200幅。构建基于卷积神经网络的木材宏、微观辨识模型——PWoodIDNet模型,采用加入动量的随机梯度下降(SGDM)方法优化模型,并利用GPU优化并行运算库,对木材宏、微观结构数据集进行分类准确率对比。结果: 相比现行GoogLeNet模型,PWoodIDNet模型准确率提高1.49%,速度提高59.69%;相比现行AlexNet模型,PWoodIDNet模型准确率提高3.76%,速度提高2.63%。结论: PWoodIDNet模型突破现有辨识方法木材辨识种类范围窄、准确率低和辨识速度慢的难点,能够有效辨识木材,并可在更短的训练时间内实现最佳辨识效果,为我国木材辨识提供一种新的方法和思路。

关键词: 木材辨识, 卷积神经网络, 微观结构, 宏观结构

Abstract:

Objective: In order to effectively improve the accuracy and speed of wood identification, a new method of wood identification based on PWoodIDNet model, a convolutional neural network model, was proposed with expects to provide advanced identification method and instruments for customs, import and export quarantine inspection, furniture enterprises and other legal departments and enterprises, and to promote the scientific and technological progress of China's wood import and export quarantine inspection industry and wood processing and manufacturing enterprises. Method: Firstly, 16 kinds of wood samples were selected, and 50 high-resolution microscopic CT images and industrial camera images were obtained from each sample with a total of 1 600 images. Then, a total of 4 800 target regions with wood rays, parenchyma, axial tracheids, pits and textures were intercepted. The image set was expanded to 19 200 images through image enhancement algorithms, such as horizontal flipping, vertical flipping, mirroring and brightness transformation. The PWoodIDNet model for the identification of tree species based on convolutional neural network was constructed. The stochastic gradient descent (SGDM) method with momentum added was used to optimize the model, and GPU was used to optimize the parallel operation library. The classification accuracy of wood macro- and micro-structure data sets was compared. Result: Compared with the existing GoogLeNet identification method, the accuracy and speed of PWoodIDNet model were increased by 1.49% and 59.69%, respectively. Compared with the existing AlexNet identification method, the accuracy and speed of PWoodIDNet model were increased by 3.76% and 2.63%, respectively. Conclusion: PWoodIDNet model could break through the difficulties of existing identification methods, such as narrow range of wood species, low accuracy and low identification speed, effectively identify wood, and could achieve the best identification effects in less training time. It would be expected to provide a new method and thought for the identification of wood in China.

Key words: wood identification, convolutional neural network, micro-structure, macro-structure

中图分类号: