Research on Dynasty Identification of Yue Kiln Celadon Using Multi-Scale Feature Fusion and Deep Learning

Yue Kiln celadon, an early Chinese porcelain of significant archaeological value, provides crucial insights into ancient techniques, cultural exchange, and diffusion. However, traditional expert-based dynasty authentication faces limitations in efficiency and objectivity. To address this challenge, we developed a deep learning approach with multi-scale feature fusion for efficient dynasty identification and feature extraction of Yue Kiln celadon spanning the Tang, Five Dynasties, and Bei Song periods. A curated dataset of 3,000 high-resolution celadon images was constructed and expanded to 9,000 images through standardized preprocessing and data augmentation. We designed a novel multi-scale feature fusion framework (HDBN-LGA) based on convolutional neural networks (CNNs), integrating an attention mechanism to accentuate discriminative features including texture, glaze characteristics, and morphology. Transfer learning strategies were further employed to mitigate sample imbalance and enhance model generalization capability. Experimental results demonstrate exceptional classification performance, achieving an average accuracy of 95.7%—specifically reaching 96.8%, 94.7%, and 95.5% for the Tang, Five Dynasties, and Bei Song periods, respectively, significantly outperforming conventional methods. The model exhibited notable robustness, maintaining accuracy declines within 3% when processing noisy or blurred images. This research establishes an efficient, objective methodology for dynasty authentication and scholarly analysis of Yue Kiln celadon, offering substantial academic and practical value through deep learning-driven multi-scale feature fusion.