A new classification method of ancient Chinese ceramics based on machine learning and component analysis
A new classification method of ancient Chinese ceramics based on machine learning and component analysis
复制标题
基于机器学习和成分分析的中国古陶瓷新分类方法
DOI：
10.1016/j.ceramint.2019.12.037
复制
发表时间：
2020
影响因子：
5.2
通讯作者：
Feng XiangQian
中科院分区：
文献类型：
--
作者：
Sun Heyang;Liu Miao;Li Li;Yan LingTong;Zhou Yue;Feng XiangQian
中国古代青瓷因其实用价值和艺术价值而受到世界各国的追捧。研究古代青瓷对于了解文化交流具有重要意义，而古代青瓷的分类是其中的重要组成部分。本文的目的是建立一个可靠的青瓷分类模型的基础上EDXRF，机器学习算法和马氏距离。用于训练机器学习模型的数据集由陶瓷坯体和釉料中的12种成分组成，这些成分通过EDXRF测量。比较四种机器学习模型的预测结果，随机森林算法在所有评价指标上表现最好。因此，随机森林是最适合青瓷分类的算法，平均准确率为96.41%，Kappa系数为0.985。样品的化学成分的含量被确定为在预测类别的相应成分范围内。选取随机森林中对古陶瓷类别识别影响较大的化学成分作为特征参数。总结了样本到类别中心的马氏距离的一般规律，并将其用于描述样本与预测类别之间的相似度。将这两种方法结合起来建立的青瓷分类模型，可以做出更具体、更准确的预测。并利用青瓷分类模型对吉州窑和楚州遗址出土的青瓷进行了分类预测。将预测结果与相应的样本背景信息进行比较，验证了模型的良好预测能力。
Ancient Chinese celadon is sought after all over the world for practical and artistic values. The study of ancient celadon is of great significance for understanding the cultural exchange, of which the classification of ancient celadon is an important part. The goal of this work was to establish a reliable celadon classification model based on EDXRF, machine learning algorithm and Mahalanobis distance. The data set for training machine learning models is constructed of 12 components in the ceramic body and glaze, which are measured by EDXRF. Comparing the predicted results of four machine learning models, the Random forest algorithm performed best on all evaluation indicators. Therefore, the Random forest was the most suitable algorithm for celadon classification with an average accuracy of 96.41% and a Kappa coefficient of 0.985. The contents of the chemical compositions of the sample were determined to be within the corresponding composition ranges of the predicted category. The chemical compositions with greater influence in identifying the categories of ancient ceramics in Random forest were chosen as the characteristic parameters. The general rules of the Mahalanobis distance from the sample to the category center were summarized and used to describe the similarity between the sample and the predicted category. The celadon classification model established by combining these two methods can make a more specific and accurate prediction. The celadon classification model was also adopted to predict the categories of samples excavated from the Jizhou kiln and Chuzhou site. The excellent prediction capability of the model was verified by comparing results with the corresponding background information of samples.