基于图像结构纹理信息的当归药材产地识别研究  

Recognition of producing areas of Angelicae Sinensis Radix based on structure-texture image decomposition

作　　者：王天舒[1] 严辉[2] 胡孔法[1] 祝蕾 郭盛[2] 段金廒[2] WANG Tian-shu;YAN Hui;HU Kong-fa;ZHU Lei;GUO Sheng;DUAN Jin-ao(College of Artificial Intelligence and Information Technology,Nanjing University of Chinese Medicine,Nanjing 210023,China;Key Laboratory of Chinese Medicinal Resources Recycling Utilization,National Administration of Traditional Chinese Medicine,Jiangsu Collaborative Innovation Center of Chinese Medicinal Resources Industrialization,Nanjing University of Chinese Medicine,Nanjing 210023,China)

机构地区：[1]南京中医药大学人工智能与信息技术学院,江苏南京210023 [2]南京中医药大学江苏省中药资源产业化过程协同创新中心国家中医药管理局中药资源循环利用重点研究室,江苏南京210023

出　　处：《中国中药杂志》2021年第16期4096-4102,共7页China Journal of Chinese Materia Medica

基　　金：国家自然科学基金面上项目(81773848,82074580);中央本级重大增减支项目(2060302);现代农业产业技术体系建设专项(CARS-21)。

摘　　要：不同产地的当归药效参差不齐,实现当归产地的准确判别对其质量评价具有参考价值。通过图像视觉信息与机器学习的方法能够对当归的产地进行智能识别。采用数码相机拍摄不同产地当归的高清图像,构建当归图像数据库。基于图像相邻像素点间灰度关系提取纹理特征,并通过支持向量机训练模型,得到当归产地预测模型。当模型训练集占比80%,测试集占比20%,相邻像素点采样半径为2时,预测准确率高达98.49%。当训练集占比仅为10%时,预测准确率也能达到93%以上。当归的3个产地中,出错比例最高的为青海互助县,最低的为云南鹤庆县。甘肃岷县与青海互助县出错当归均被误判为云南鹤庆县所产。青海互助县出错当归绝大部分均被误判为甘肃岷县所产。因此,该文的当归药材产地识别方法,能够准确预测对当归的产地进行预测,具有快速无损、识别准确率高以及稳定性强的优势。甘肃岷县与青海互助县的当归具有明确的形态差异,甘肃岷县与青海互助县出错当归与云南鹤庆县当归具有相似的形态特征,云南鹤庆县的大部分出错当归与甘肃岷县当归具有相似的形态特征。The pharmacological effects of Angelicae Sinensis Radix from different producing areas are uneven. Accurate identification of its producing areas by computer vision and machine learning(CVML) is conducive to evaluating the quality of Angelicae Sinensis Radix. This paper collected the high-definition images of Angelicae Sinensis Radix from different producing areas using a digital camera to construct an image database, followed by the extraction of texture features based on the grayscale relationship of adjacent pixels in the image. Then a support vector machine(SVM)-based prediction model for predicting the producing areas of Angelicae Sinensis Radix was built. The experimental results showed that the prediction accuracy reached up to 98.49% under the conditions of the model training set occupying 80%, the test set occupying 20%, and the sampling radius(r) of adjacent pixels being 2. When the training set was set to 10%, the prediction accuracy was still over 93%. Among the three producing areas of Angelicae Sinensis Radix, Huzhu county, Qinghai province exhibited the highest error rate, while Heqing county, Yunnan province the lowest error rate. Angelicae Sinensis Radix from Minxian county, Gansu province and Huzhu county, Qinghai province were both wrongly attributed to Heqing county, Yunnan province, while most of those from Huzhu county, Qinghai province were misjudged as the samples produced in Minxian county, Gansu province. The method designed in this paper enabled the rapid and non-destructive prediction of the producing areas of Angelicae Sinensis Radix, boasting high accuracy and strong stability. There were definite morphological differences between Angelicae Sinensis Radix samples from Minxian county, Gansu province and those from Huzhu county, Qinghai province. The wrongly predicted samples from Minxian county, Gansu province and Huzhu city, Qinghai province shared similar morphological characteristics with those from Heqing county, Yunnan province. Most wrongly predicted samples from Heqing county,

关 键 词：当归 机器学习 图像处理 特征提取 

分 类 号：R282.5[医药卫生—中药学]