Research On Ancient Chinese Character Recognition Based On Object Detection And Knowledge Graph

Ancient Chinese characters are important materials for historical research,they have a long history and have played an important role in the development of Chinese characters.The recognition of ancient Chinese characters has long been a problem that researchers are committed to studying.Although modern technologies and methods continue to advance,various problems such as long-tail effect,insufficient sample size and complex glyphs still exist in ancient character recognition,and these problems bring many challenges to ancient character recognition.Most studies mainly use deep learning classification algorithms or detection algorithms to recognize ancient Chinese characters,but these studies have ignored the following issues:(1)Ancient Chinese characters have complex glyphs,significant differences between characters,and lack of connection between characters,while lacking exploration of common features.In addition,there is a lot of overlap between the parts that make up ancient characters,and it is difficult to achieve good accuracy with detection or classification methods based on ordinary deep learning;(2)Due to the long distance between ancient Chinese characters and modern times,most characters have only one image,and the unbalanced sample size leads to a serious problem of long-tail effect.At the same time,the construction of some ancient Chinese character knowledge bases is not perfect,while part of the knowledge base of ancient characters is not well constructed,and many fonts do not have fonts corresponding to modern Chinese characters and corresponding textual information..In order to solve the above problems,this paper proposes a network framework that integrates target detection and knowledge mapping,and mainly studies ancient Chinese character component recognition,character recognition,and knowledge base construction on oracle bone rubbings,the main research contents and contributions of this paper are as follows:1.In this paper,a comprehensive oracle bone detection dataset is firstly constructed.To address the problems of few connections between ancient characters and the lack of obvious common features,this paper transforms the character recognition problem into a fine-grained component recognition problem and proposes a detection network for ancient Chinese character component recognition.The Radical Attention Block(RAB)is proposed to enhance the extraction of the oracle part features by comparing different advanced attention mechanism models,and the SE(Squeeze-and-Excitation)attention mechanism is selected to join the network to alleviate the problems of part adhesion and overlap in oracle characters.Based on the existing pyramid network,the feature fusion network is improved as(Adaptively Spatial Feature Fusion,ASFF),which effectively fuses the feature maps of different scales of the model.In terms of loss function and non-maximum suppression,the original loss function and non-maximum suppression algorithm are replaced by SIOU loss function and SIOU-NMS non-maximum suppression algorithm.In addition,this paper proposes a lightweight detection network applied to mobile and other scenarios.Through a large number of comparison experiments,it is shown that the method proposed in this paper has a good effect on component recognition,comprehensively improving the performance of Oracle component recognition,enhancing the relationship between characters,and providing a good foundation for subsequent Oracle character recognition.2.To address the problem of long-tail effect in Oracle dataset,this paper proposes a network structure integrating target detection and knowledge graph Oracle Recognition(KGOR)to improve the number of Oracle character recognition.The entire network model consists of three parts,including the component detection module,the position relationship determination module and the knowledge graph inference module.To address the problem of imperfect oracle knowledge base,this paper constructs a comprehensive oracle knowledge graph based on the existing oracle text information and proposes a position relationship determination module,which takes the position coordinates and detected components as input and the structure of oracle text as output.Finally,the recognized characters are inferred by using other textual information such as parts and text structure in conjunction with the knowledge graph.Through experiments,it is found that the network model proposed in this paper has significantly improved the effect compared with other oracle classification models.For uninterpreted oracle bone pictures,the method in this paper can construct modern Chinese character fonts corresponding to the oracle bones.This paper plays a role in promoting the preservation of the oracle bone knowledge base and the research of other ancient scripts.