A study on image acquisition methods to improve classification accuracy of excavated ceramics

A wide variety of ceramics have been excavated from archaeological sites in Japan. These ceramics vary widely in production locations, eras, types, and decorative patterns, making classification difficult for non-experts. This study investigates image acquisition methods to improve the classification accuracy of excavated ceramics using convolutional neural networks (CNNs). To enhance classification accuracy, the effects of ceramic fragment presence and background trimming on classification performance using a data set containing both intact ceramics and fragments are experimentally evaluated. The results suggest that training with non-fragmented ceramics and applying background trimming may improve classification accuracy.