LCINet: Local Cross-position Interaction Network for Oracle Bone Inscriptions Recognition

Oracle Bone Inscriptions (OBIs) are a type of ancient Chinese hieroglyphs, which are precursors of many Asian characters. Automatic OBIs recognition can assist archaeologists in understanding the history and evolution of hieroglyphs. Recently, deep learning methods have been gradually applied to OBIs recognition, but these methods often fail to leverage the glyphological feature information of oracle bone characters. In this paper, we propose a novel Local Cross-position Interaction (LCI) module, which dynamically adjusts the relative importance of feature maps in convolutional neural networks (CNNs) by exploiting the potential of the glyphological information. LCI extracts the glyphological context information by orientation pooling, then constructs the local dependencies between glyphological context features via a cross-position interaction. Subsequently, we combine the LCI module with the residual module to form the Residual-LCI module and then build an LCINet for automatic OBIs recognition by stacking multiple Residual-LCI modules. In addition, we construct an OBIs dataset named OBI316 to verify the effectiveness of LCINet, which will be released soon. The comprehensive experiments on the OBI316 dataset demonstrate that our LCINet outperforms baselines and state-of-the-art attention-based networks. The CIFAR datasets are used to further demonstrate the generalization ability of our method.