陈舒迪　柴琴琴　张勋　黄泽豪　林羽　徐伟

摘要： 金線莲是中国珍稀中草药，不同品系的金线莲具有细微的形态差异和显著的药效差异。针对金线莲的单一特征贡献能力不足以及传统分类器泛化能力不佳的问题，提出使用形状、颜色和纹理特征对金线莲叶片图像进行特征提取与融合，再使用表现性能更优的LightGBM（轻量级梯度提升机）构建分类器，以提高金线莲识别正确率。LightGBM具有精确高效等优点，将提取得到的高层次特征导入LightGBM进行训练预测，可以有效提高分类准确性。对金线莲数据集中的6个品系共368幅叶片图像进行试验，结果表明，相比于传统的分类方法，基于多特征融合和LightGBM的模型识别效果最好，10次随机试验的平均识别率比传统方法KNN、SVM和GBDT高，并且在分类评价指标精确率、召回率、综合评价指标上有较优表现，该研究结果可为中药材品系识别提供参考。

关键词： 金线莲;多特征融合;LightGBM;叶片识别

中图分类号： TP391 文献标识码： A 文章编号： 1000-4440（2021）01-0155-08

Identification of Anoectochilus roxburghii strains based on multi feature fusion and LightGBM

CHEN Shu-di1，2， CHAI Qin-qin1，2， ZHANG Xun3， HUANG Ze-hao3， LIN Yu3， XU Wei3

（1.College of Electrical Engineering and Automation， Fuzhou University， Fuzhou 350108， China;2.Ministry of Education Key Laboratory of Medical Instrument and Pharmaceutical Technology， Fuzhou University， Fuzhou 350108， China;3.College of Pharmacy， Fujian University of Traditional Chinese Medicine， Fuzhou 350122， China）

Abstract： Anoectochilus roxburghii is a rare Chinese herbal medicine， different strains of A.roxburghii have slight morphological difference and significant variance in medicinal effects. To solve the problems of insufficient contribution ability of single-feature in A.roxburghii and poor generalization ability of traditional classifiers， it was proposed to use shape， color and texture features to extract and fuse features of A.roxburghii leaf images. Then LightGBM （light weight class elevator） with a better performance was used in building classifier， so as to improve the recognition accuracy of A.roxburghii. LightGBM had the advantages of accurate and efficient， and the prediction accuracy could be improved effectively by importing the extracted high-level features into LightGBM to forecast the training. A total of 368 leaf images from six strains of A.roxburghii dataset were trained and tested. The results showed that model based on multi-feature fusion and LightGBM had the best recognition effect compared with traditional classification methods. The average recognition rate of ten random experiments was higher than traditional methods such as KNN， SVM and GBDT， and it showed good performance in classification evaluation indices like precision， recall rate and comprehensive evaluation index. The result can provide reference for the identification of different strains of traditional Chinese medicine.

Key words： Anoectochilus roxburghii;multi feature fusion;LightGBM;leaf recognition

金线莲属于兰科开唇兰属植物[1-2]，含糖类、多种氨基酸及无机元素等，常用于糖尿病、肺热咳嗽、急慢性肝炎等疾病的治疗[3-4]。由于金线莲药用价值高、市场需求量大，而野生金线莲的繁殖率低、生存条件受限等原因，导致目前市场上销售的金线莲基本是人工栽培的。不同品系的金线莲虽外观相似但生物产量积累和药用化合物组成有着很大的差异，市面上以低药用价值的金线莲冒充优质品系的金线莲的现象尤为常见。……