A Bronze Inscription Recognition Method Based on Multi level Feature Fusion Network

Automatic recognition of bronze inscription characters is a challenging yet critical task in digital humanities, hindered by irregular shapes, stylistic variations, and significant class imbalance. In this paper, we propose a Multi-Level Feature Fusion Network (MLFFN) specifically designed to address these challenges. Our method employs a hierarchical backbone integrated with a Dual Attention Module (DAM), effectively capturing fine-grained details and global contextual information via parallel standard and depthwise separable convolutions. Additionally, we introduce a multi-scale feature fusion strategy that combines complementary features across multiple abstraction levels. Extensive experiments on a specially curated dataset consisting of 8,700 images and 167 character classes demonstrate the superior performance of MLFFN, achieving an accuracy of 89.2% and a macro-averaged F1-score of 86.4%, significantly outperforming state-of-the-art CNN architectures. Ablation studies further confirm the effectiveness of our design choices.