欢迎访问林业科学,今天是

林业科学 ›› 2021, Vol. 57 ›› Issue (9): 152-159.doi: 10.11707/j.1001-7488.20210915

• 论文与研究报告 • 上一篇    下一篇

iWood: 基于卷积神经网络的濒危珍贵树种木材自动识别系统

何拓,刘守佳,陆杨,张永刚,焦立超,殷亚方*   

  1. 中国林业科学研究院木材工业研究所 中国林业科学研究院木材标本馆 北京 100091
  • 收稿日期:2020-09-29 出版日期:2021-09-25 发布日期:2021-11-29
  • 通讯作者: 殷亚方
  • 基金资助:
    中央级公益性科研院所基本科研业务费专项资金重点项目"木材标本资源及其科学数据平台建设"(CAFYBB2021ZD002)

iWood: An Automated Wood Identification System for Endangered and Precious Tree Species Using Convolutional Neural Networks

Tuo He,Shoujia Liu,Yang Lu,Yonggang Zhang,Lichao Jiao,Yafang Yin*   

  1. Research Institute of Wood Industry, CAF Wood Collections(WOODPEDIA), CAF Beijing 100091
  • Received:2020-09-29 Online:2021-09-25 Published:2021-11-29
  • Contact: Yafang Yin

摘要:

目的: 构建基于卷积神经网络的木材识别系统,实现木材树种在多场景条件下的自动精准识别,为我国提升CITES履约执法能力、加强林产品产业链监管以及保障木材安全提供科技支撑。方法: 采集15种黄檀属和11种紫檀属木材标本横切面构造特征图像,建立图像数据集Rosewood-26;构建AlexNet、VGG16、DenseNet-121和ResNet-50共4种卷积神经网络模型,基于ImageNet图像数据集对模型进行迁移学习,采用Rosewood-26图像数据集训练、测试和比较模型,优选识别性能较好的卷积神经网络模型,并进行木材树种分类;在此基础上,构建包含15种黄檀属和11种紫檀属树种的木材自动识别系统iWood,利用市场木材样品对系统进行应用测试和评价。结果: 在构建的4种卷积神经网络模型中,ResNet-50模型表现出最高的识别精度(98.33%)、最少的权重数量和较低的模型复杂性,适用于木材树种准确快速识别;ResNet-50模型对9种黄檀属和3种紫檀属木材的识别精度达100%,并可成功鉴别构造特征极其相似的檀香紫檀和染料紫檀;基于ResNet-50模型构建的木材自动识别系统iWood,在"属"和"种"水平的识别精度分别为91.8%和77.3%。结论: 基于卷积神经网络的木材识别系统iWood适用于海关执法、木材贸易和质量监督检验等多场景下的木材自动精准识别,能够为我国提升CITES履约执法能力、加强林产品产业链监管以及保障木材安全提供科技支撑。

关键词: iWood, 卷积神经网络, 木材构造特征, 图像数据集, 木材自动识别, 识别精度

Abstract:

Objective: An image recognition method based on convolutional neural networks(CNNs) for endangered and precious tree species was developed in this study to reach the automatic and accurate identification of timbers at the species level in multi-scene scenarios. Method: Images representing wood anatomical features were collected from transverse section of authentic wood specimens for 15 Dalbergia and 11 Pterocarpus species, and an image data set of Rosewood-26 was established. Four CNN models, i.e. AlexNet, VGG16, DenseNet-121 and ResNet-50 were constructed and pre-trained with ImageNet for transfer learning. And then the image data set of Rosewood-26 was deployed to re-train and test these models, which were comparatively analyzed and evaluated to obatin an optimal one for wood identification. A wood identification system was developed for identifying those samples collected from timber market for application test. Result: Among the four CNN models, ResNet-50 showed the highest identification accuracy(98.33%) and lower model complexity, which is preferable in the context of accurate and rapid wood identification. The ResNet-50 model achieved 100% accuracy when identifying 9 Dalbergia and 3 Pterocarpus species, and successfully discriminated P. santalinus from its look-alike species, P. tinctorius. The automated wood identification system based on ResNet-50 model exhibited an identification accuracy of 91.8% at the genus level and 77.3% at the species level. Conclusion: The wood identification system, iWood, developed in this study using CNNs is applicable in broad fields, i.e. customs enforcement, timber trade and quality inspection, and can reach automatic and accurate identification of wood species. This study will provide scientific support for promoting regulation of forest product industry chain, enhancing CITES enforcement capabilities and protecting forest species diversity.

Key words: iWood, convolutional neural networks, wood anatomical features, image dataset, automated identification, identification accuracy

中图分类号: