基于陶瓷显微特征与深度学习的窑口分类方法研究  

Using Microscopic Imaging and Deep Learning to Classify the Provenance of Archaeological Ceramics

作　　者：王倩 Wang Qian(Jingdezhen Ceramic University,Jingdezhen,333403)

机构地区：[1]景德镇陶瓷大学,景德镇333403

出　　处：《陶瓷研究》2024年第2期36-39,共4页Ceramic Studies

基　　金：江西省研究生创新基金(项目编号:YC2022-s903)——《基于深度迁移学习的古陶瓷碎片的检测与分类研究》。

摘　　要：针对传统人工鉴定陶瓷类别存在误差,基于化学特征或光谱特征分析的方法可能对陶瓷造成损坏性等问题,提出采用显微特征与深度学习模型的陶瓷分类方法,并且不破坏陶瓷结构和艺术价值。该方法运用VGG-16,Inception-v3,GoogLeNet三种深度学习模型进行图像特征提取,采用随机梯度下降(SGD)算法,通过冻结和解冻模型层来实现模型参数的最优拟合;实验采用准确率、精确度、召回率和F1评分标准来提供综合的分类结果,在5折交叉验证和独立测试实验下,对比这三种深度学习模型,得到Inception-v3模型效果更好,此外,我们运用对应分析来探讨不同窑口的陶瓷显微图像的分布。该方法可应用于陶瓷文物鉴定领域,有助于提高诊断的准确性和效率,为相关研究领域提供新的思路和方法。Aiming at the problems of traditional manual identification of ceramic categories with errors and the possible damaging nature of ceramics by methods based on chemical features or spectral characterization,we propose a ceramic classification method using microscopic features and deep learning models without destroying the ceramic structure and artistic value.The method utilizes three deep learning models,VGG-16,Inception-v3,and GoogLeNet,for image feature extraction,and a stochastic gradient descent(SGD)algorithm for optimal fitting of model parameters by freezing and unfreezing the model layers;the experiments use the accuracy,precision,recall,and F1 scoring criteria to provide comprehensive classification results,and the classification results are compared under a 5-fold cross-validation and independent test experiments,comparing these three deep learning models,the Inception-v3 model is obtained to be more effective;in addition,we apply correspondence analysis to explore the distribution of ceramic micrographs from different kilns.This method can be applied to the field of ceramic artifact identification,which helps to improve the accuracy and efficiency of diagnosis and provides new ideas and methods for related research fields.

关 键 词：显微图像 来源 分类 深度学习 无损检测 

分 类 号：J527[艺术—艺术设计]