Abstract
Automated age estimation of archaeological artifacts is crucial for categorization and dating, yet challenging due to variations in characteristics, degradation, and limited chronological information. This study investigates the performance of Convolutional Neural Network (CNN) architectures and loss functions for accurate age estimation. Using a dataset of about 10,000 labeled images from distinct archaeological sites, spanning 16 periods ranged from the Paleolithic to the Late Islamic periods, our results demonstrate top-5 accuracy above 90%. Notably, our empirical results revealed that InceptionV3, while known for its strong performance in object recognition tasks, outperformed other architectures in this classification task. Additionally, we found that conventional cross-entropy loss functions can, in some architectures, outperform ordinal cross-entropy, challenging conventional wisdom. Our findings not only advance the computational methodologies available for artifact dating but also provide critical insights into the nuanced selection of neural network architectures and loss functions, thereby opening new avenues for research in computational archaeology.