Using microscopic imaging and ensemble deep learning to classify the provenance of archaeological ceramics.

Considering the substantial inaccuracies inherent in the traditional manual identification of ceramic categories and the issues associated with analyzing ceramics based on chemical or spectral features, which may lead to the destruction of ceramics, this paper introduces a novel provenance classification of archaeological ceramics which relies on microscopic features and an ensemble deep learning model, overcoming the time consuming and require costly equipment limitations of current standard methods, and without compromising the structural integrity and artistic value of ceramics. The proposed model includes the following: the construction of a dataset for ancient ceramic microscopic images, image preprocessing methods based on Gamma correction and CLAHE equalization algorithms, extraction of image features based on three deep learning architectures-VGG-16, Inception-v3 and GoogLeNet, and optimal fusion. This latter is based on stochastic gradient descent (SGD) algorithm, which allows optimal fitting of the fusion model parameters by freezing and unfreezing model layers. The experiments employ accuracy, precision, recall and F1 score criteria to offer a comprehensive of the classification outcomes. Under 5-fold cross-validation and independent testing, the proposed fusion-based model performed excellently after comparing above three typical deep learning model. The predictive results of the ensemble deep learning are very stable at about 0.9601, 0.9615, 0.9607 and 0.9583 in precision, recall, F1-score, and accuracy on the independent testing dataset, respectively. This indicates that our model is robust and reliable. Furthermore, we use correspondence analysis to explore the distribution of the ceramic microscopic images from different kilns. This method can be applied in the field of ceramic cultural relic identification, contributing to improved diagnostic accuracy and efficiency, and providing new ideas and methods for related research areas.