卷积神经网络在红木树种识别中的应用  

Application of convolutional neural network in rosewood species identification

作　　者：黄鹏桂 赵璠 李晓平[2] 吴章康[2] 汤正捷[2] 张严风 HUANG Penggui;ZHAO Fan;LI Xiaoping;WU Zhangkang;TANG Zhengjie;ZHANG Yanfeng(College of Big Data and Intelligent Engineering,Southwest Forestry University,Kunming 650224,Yunnan,China;National Forestry and Grassland Administration Center for Quality Supervision and Testing of Wood and Bamboo Products,Southwest Forestry University,Kunming 650224,Yunnan,China)

机构地区：[1]西南林业大学大数据与智能工程学院,云南昆明650224 [2]西南林业大学国家林业和草原局木材与木竹制品质量检验检测中心,云南昆明650224

出　　处：《浙江农林大学学报》2020年第6期1200-1206,共7页Journal of Zhejiang A&F University

基　　金：国家自然科学基金资助项目(31870551);西南林业大学科研启动基金(111807)。

摘　　要：【目的】不同类型的红木由于生长周期和木材特性的不同,导致商业价格差异悬殊,其中还包含有国家保护木种。本研究旨在找到能准确地识别红木种类的方法,以防止交易中的欺诈行为和保护树种。【方法】以国家林业和草原局木材与木竹制品质量检验检测中心(昆明)实际检测中累积的黄檀属Dalbergia和紫檀属Pterocarpus中的交趾黄檀D.cochinchinensis、刀状黑黄檀D. cultrata、卢氏黑黄檀D. louvelii、巴里黄檀D. bariensis、奥氏黄檀D. oliveri、大果紫檀P. macrocarpus、檀香紫檀P. santalinus等7种红木的376个样本作为基本数据,使用计算机算法扩展样本数量,提出自动化识别红木的卷积神经网络模型。【结果】该方法能够自动提取适合模型分类识别的特征,使用更为便捷,相比其他传统方法识别效果更准确的,结果证明平均识别精度达99.4%。【结论】自建的卷积神经网络可以有效识别红木树种,虽然在调参优化与训练时间大于VGG16等迁移学习方法,但泛化能力更强,证明了自建模型在红木识别应用上优于迁移学习模型。[Objective] Due to the differences in growth cycles and wood characteristics, the market price of rosewood species including some nationally protected species varies sharply. To prevent fraud and better protect tree species, this paper is aimed at the accurate identification of rosewood species. [Method] With 376 samples of 7 species of rosewood collected in the actual test by the National Forestry and Grassland Administration Center for Quality Supervision and Testing of Wood and Bamboo Products(Kunming) selected as the research subject, a convolutional neural network model for automatic recognition of redwood is proposed using the computer algorithm to expand the number of samples. [Result] This method can automatically identify features that fit with the model of classification and recognition, making it more convenient to utilize. Compared with traditional methods, it secures a more accurate identification with an average identification accuracy of 99.4%. [Conclusion] The self-built convolutional neural network can effectively identify redwood species. Although ittakes a long time to achieve the parameter adjustment and train than VGG16 and other transfer learning methods, it shows stronger competence of generalization. It is proved that the self-built model is superior to the transfer learning model when applied in rosewood recognition.

关 键 词：红木切片 自动识别 卷积神经网络 黄檀属 紫檀属 

分 类 号：S781.1[农业科学—木材科学与技术]