赵路路1  殷泽轩1  陈红1  刘诚1,2

（1.广东省科学院智能制造研究所，广东 广州 510070

2.昆明理工大学机电工程学院，云南 昆明 650051）

摘要：为避免当归与独活2种中药材混淆，结合深度学习和近红外高光谱成像技术进行当归与独活的分类。首先，获取当归与独活样本的平均光谱数据，并采用显著图选出平均光谱数据中的20个波段作为特征波段，实现特征提取与降维；然后，在全波段（共181个波段）和特征波段（共20个波段）的光谱数据集上，分别采用一维卷积神经网络（1D-CNN）模型和支持向量机（SVM）模型对当归与独活进行分类。分类结果显示：利用全波段光谱数据集建模时，1D-CNN和SVM在测试集上的分类准确率分别为98.6%和98.1%；利用特征波段光谱数据集建模时，1D-CNN和SVM在测试集上的分类准确率分别为96.1%和95.5%。因此，将高光谱成像技术与深度学习相结合可以实现当归与独活的快速分类。

关键词：高光谱成像；一维卷积神经网络；支持向量机；特征波段；分类 

中图分类号：TP391.4           文献标志码：A           文章编号：1674-2605(2023)06-0006-07

DOI：10.3969/j.issn.1674-2605.2023.06.006

Classification of Angelicae and Heracleum Based on Hyperspectral Imaging 

ZHAO Lulu1  YIN Zexuan1  CHEN Hong1  LIU Cheng1,2

(1.Institute of Intelligent Manufacturing, Guangdong Academy of Science, Guangzhou 510070, China

2.Faculty of Mechanical and Electrical Engineering，Kunming University of Science and Technology, 

Kunming 650051, China)

Abstract: To avoid confusion between Angelicae and Heracleum, deep learning and near-infrared hyperspectral imaging techniques were combined to classify them. Firstly, obtain the average spectral data of Angelicae and Heracleum samples, and use saliency maps to select 20 bands from the average spectral data as feature bands to achieve feature extraction and dimensionality reduction; Then, a one-dimensional convolutional neural networks (1D-CNN) model and a support vector machine (SVM) model were used to classify Angelicae and Heracleum on spectral datasets with a total of 181 bands and 20 bands, respectively. The result of classification showed that when modeling using full band spectral datasets, the accuracy of 1D-CNN and SVM on the test set was 98.6% and 98.1% in classification, respectively; When modeling using the characteristic bands spectral datasets, the accuracy of 1D-CNN and SVM on the test set was 96.1% and 95.5% in classification, respectively. Therefore, combining hyperspectral imaging technology with deep learning can achieve rapid classification of Angelicae and Heracleum.

Keywords: hyperspectral imaging; one-dimensional convolution neural networks; support vector machine; characteristic bands; classification