Oracle Character Recognition Based on Attention Enhancement and Multi-level Feature Fusion

Zhiwang Han, Nurbiya Yadikar, Xuebin Xu, Alimjan Aysa, Kurban Ubul

Abstract: Oracle bone characters represent the earliest inscriptions in China. Recognizing and deciphering these characters is significant. Despite some progress made by recent methods, their recognition accuracy remains limited by two major issues: 1) how to focus on characters features within complex background noise images, and 2) how to effectively fuse shallow detail information with deep semantic information. To address these issues, we propose a novel deep learning model called Character Feature Enhancement Network (CFE-Net). The model consists of two key components: Character Feature Enhancement (CFE) and Adaptive Multi-level Classifier Fusion (AMCF). Specifically, CFE utilizes the Spatial Focus Attention Module (SFAM) to focus on extracting foreground character features and suppressing background noise, thereby significantly enhancing high-level semantic representation capabilities. AMCF, on the other hand, achieves multi-level feature fusion by adaptively fusing the outputs of different classifiers, effectively avoiding information loss or interference that simple fusion strategies might cause. We evaluated the CFE-Net on two rubbing oracle bone characters benchmark datasets, OBC306 and Oracle-MNIST. The experimental results demonstrate that CFE-Net significantly outperforms several existing methods in terms of Top-1 accuracy, establishing it as the new state-of-the-art.