Advancing Archaeological Ceramic Image Classification via Few-Shot Learning

Ancient Chinese ceramic classification faces challenges due to limited data and fine-grained texture analysis. To overcome this challenge, we propose a few-shot learning method, integrating CLIP-like networks with LLM-generated textual knowledge. By combining cross-entropy and self-consistency losses with Gaussian-weighted prompt fusion, our approach enhances the incorporation of prior knowledge, thereby addressing the challenge of fine-grained texture analysis in cultural relic classification. We introduce the Chinese Ceramic Vessel Dataset (CCVD), a carefully curated collection of 1,020 samples across 27 subclasses featuring diverse shapes. Through comprehensive experimental evaluation, we demonstrate the superiority of our integrated approach over existing few-shot learning methods in ceramic image classification.