古陶瓷作为中华文化的瑰宝, 自古以来不仅在国内受到追捧, 在国外同样被视若珍宝。 伴随着古代商贸的进行, 中国古陶瓷遍布全球各地, 辗转流传被私人或博物馆收藏, 还有部分古陶瓷经墓葬发掘以及沉船打捞后被收藏于博物馆, 这类古陶瓷的产地溯源一直以来都是陶瓷考古的重点, 对于研究古代商贸和文化交流有重要的意义。 通过便携式数码显微镜、 分光光度计、 X射线荧光等方法对从越窑后司岙、 越窑寺龙口、 龙泉枫洞岩窑、 耀州窑发掘出土的青釉瓷样品进行分析测量, 获得了来自四个窑青釉瓷样品的微观气泡尺寸分布特征、 紫外可见近红外光谱特征、 釉的成分数据。 将来自四个窑青釉瓷样品的这三种特征作为变量建立卷积神经网络分类模型进行训练和验证, 结果表明青釉瓷的微观气泡尺寸分布特征、 紫外可见近红外光谱特征以及瓷釉成分数据均有效, 但是不同特征的分类准确率差异非常明显。 三十次随机划分训练集与测试集的模型训练平均准确率: 微观气泡尺寸分布特征模型为75%, 紫外可见近红外光谱特征模型为89.2%, 成分数据模型为92.1%, 成分数据模型准确率最高且训练集与测试集准确率相差最小。 将基于不同特征训练好的模型参数保存进行融合后再训练发现基于紫外可见近红外光谱特征的模型与基于微观气泡尺寸分布特征模型融合后准确率提升至93.7%, 而将三种特征的模型融合后准确率提升至最高的97.4%。 五折交叉验证的结果表明多种特征融合后的模型可以有效避免出现单一特征模型对越窑后司岙以及越窑寺龙口样品交叉错判数较多的情况。 综合来看基于卷积神经网络探索更多的古陶瓷有效分类特征对于实现古陶瓷的精准溯源是可行的。

As a treasure of Chinese culture, ancient ceramics have been sought at home and abroad since ancient times. With the development of ancient commerce and trade, ancient Chinese ceramics spread worldwide and were collected by private individuals or museums. Some were collected in museums after being excavated from tombs and salvaged from sunken ships. Tracing the origin of such ancient ceramics has always been the focus of ceramic archaeology, which is of great significance to studying ancient commerce and cultural exchanges. Using a portable digital microscope, spectrophotometer, X-ray fluorescence and other methods, the celadon porcelain samples excavated from Housi'ao, Silongkou, Fengdongyan and Yaozhou kilns were analyzed, and the data of the microbubble size distribution characteristics, ultraviolet, visible near-infrared spectrum characteristics and glaze composition of the celadon porcelain samples from these four kilns were obtained. The convolution neural network classification model was established by using these three features as variables for training and verification. The results show that the microbubble size distribution features, ultraviolet, visible near-infrared spectral features and glaze composition data of celadon porcelain are effective, but the difference in classification accuracy is very obvious.The average accuracy of model training of 30 randomly divided training sets and test sets: the model of microbubble size distribution features is 75%, the model of ultraviolet, visible near-infrared spectrum features is 89.2%, and the model of component data is 92.1%.The accuracy of the glaze composition data model is the highest, and the difference between the accuracy of training sets and test sets is the smallest. After saving the model parameters trained based on different features for fusion and retraining, it was found that the accuracy of the model after the fusion of ultraviolet, visible and near-infrared spectral features and microbubble size distribution features was improved to 93.7% and the accuracy of the model after the fusion of the three features was improved to the highest 97.4%. The results of the five-fold cross-validation show that the model, after the fusion of multiple features, can effectively avoid the case that the single feature model has many cross misjudgments on the samples from Housi'ao and Silongkou. In general, it is feasible to explore more effective classification features of ancient ceramics based on convolutional neural networks to trace the source of ancient ceramics accurately.