一种弱监督细粒度深度网络的木材分类方法  

A Wood Classification Method Based on Weakly Supervised Fine-grained Deep Network

作　　者：戴天虹[1] 谢千程 黄建平 孙春雪 丛士杰 黄新望 李克新 DAI Tianhong;XIE Qiancheng;HANG Jianping;SUN Chunxue;CONG Shijie;HUANG Xinwang;LI Kexin(College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,China;Artificial Intelligence Academy,Wuxi Vocational College of Science and Technology,Wuxi Jiangsu 214000,China)

机构地区：[1]东北林业大学机电工程学院,哈尔滨150040 [2]无锡科技职业学院人工智能学院,江苏无锡214000

出　　处：《西南大学学报（自然科学版）》2022年第10期161-172,共12页Journal of Southwest University(Natural Science Edition)

基　　金：中央高校基本科研业务费专项资金资助项目(2572019CP17,2572019CP19);黑龙江省自然科学基金项目(C201414,TD2020C001);哈尔滨市科技创新人才项目(2014RFXXJ086).

摘　　要：针对使用木材的微观结构对木材树种进行识别分类,提出一种新的基于细粒度图像识别的深度卷积神经网络自动分类算法,在阐述Navigator Teacher Scrutinizer Network算法的基础上,首先,利用分段线性激活函数对特征的存在程度和缺失程度的选择能力进行改进,在改进后的特征选择模型算法中搜寻最优α参数;其次,在改进后的算法中加入一个全局K-max池化层并应用在木材分类中,获得最佳的分类结果.实验结果表明,相比于原始NTS神经网络,本文所提算法能够更准确地实现数据分类,该模型的实验准确率为88.36%,准确率高,实用性强,可以提高木材树种分类精度,为木材树种快速分类提供参考.Aiming at the identification and classification of the species of wood with the microstructure of wood,a new automatic classification algorithm of deep convolutional neural network based on fine-grained image recognition was proposed.On the basis of elaborating the Navigator Teacher Scrutinizer Network algorithm,the piecewise linear activation function was adopted firstly to improve the selection ability of the existence and lack degrees of the feature,and optimal parameters were searched in the improved feature selection model algorithm.Secondly,a global K-max pooling layer was added to the improved algorithm and applied to wood classification and obtained the best classification results.Finally,experimental results showed that the proposed algorithm can achieve more accurate data classification in comparison with original NTS neural network.The experimental accuracy rate of this model was 88.36%.This accuracy rate is high,and with strong practicability.It can improve the accuracy of wood species classification and provide a reference for rapid classification of wood species.

关 键 词：导航教师审查网络 弱监督卷积神经网络 木材显微识别 分类 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]