Shape Prior Fusion for Oracle Bone Inscriptions Detection

The methods for detecting texts in different modern scenes automatically have achieved rapid development, however, they are not able to directly apply to another ancient but important type of Chinese characters, Oracle Bone Inscriptions(OBIs). Different from modern texts which are written or printed in a specific place and represent specific meaning, OBIs were just carved on animal bones with few vague relations and were similar to other noise areas which only contain cracks. In this paper, we aim to detect OBIs from the digital images of rubbings by using deep learning methods. We propose an attention-based model and leverage the latent features of ancient characters to guide the model. First, we used a classification network to acquire the shape prior of inscriptions. Then, a module based on the attention mechanism is used to merge the information from the prior to the high-dimension feature map. With this auxiliary module, the prior information is introduced to the backbone network and existing models can be trained from scratch to complete the task of OBIs detection without a mass of pre-training works and elaborate design of fine-tuning. We conduct experiments on a newly divided and rigorous dataset and the results prove that our network with a lightweight scale of parameter finally achieves preferable performance.