Oracle Bone Script Recognition Based on Multi-scale Feature Fusion and Knowledge Distillation

Jiaoyan Wang, Xuebin Xu, Alimjan Aysa, Jun Ding, Kurban Ubul

Abstract: Oracle bone scripts, being the oldest and most developed writing system discovered in China till date, have been meticulously studied by numerous scholars. This paper presents the Multi–scale Feature Fusion Attention Net (MFFA–Net) as a solution to the issue of variant characters in the oracle bone affecting recognition accuracy. The network is based on ResNet18 and employs asymmetric convolution alongside a refined coordinate attention technique to acquire image features. It also integrates perceptual field information of varying sizes through hierarchical bilinear pooling. Furthermore, knowledge from the Wide_ResNet101 and DenseNet169 is transferred to MFFA-Net using knowledge distillation techniques. Finally, to validate the effectiveness of our proposed method, we conducted rigorous experiments on the OBC306, OBC265, and EOBC datasets, comparing our results with those of existing methods. The experimental results demonstrate that our method obtains remarkable performance in oracle character recognition, reaching state-of-the-art Top–1 accuracy with 92.42%, 94.78%, and 98.82% on the OBC306, OBC265, and EOBC datasets, respectively.