Radical-based extract and recognition networks for Oracle character recognition

• Published in: International journal on document analysis and recognition Vol. 25; no. 3; pp. 219 - 235
• Main Authors: Lin, Xiaoyu, Chen, Shanxiong, Zhao, Fujia, Qiu, Xiaogang
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.09.2022; Springer Nature B.V
• Subjects: Algorithms; Annotations; Character recognition; Computer Science; Datasets; Feature extraction; Feature maps; Image Processing and Computer Vision; Machine learning; Original Paper; Pattern Recognition; Attention mechanism; Text detection and recognition; Oracle radical; MSER
• ISSN: 1433-2833; 1433-2825
• DOI: 10.1007/s10032-021-00392-2

The recognition of Oracle bone inscription (OBI) is one of the most fundamental aspect of OBI study. However, the complex glyph structure and many variants of OBI, which hinder the advancement of automatic recognition research. In order to solve these problems, this paper designs an Oracle radical extract and recognition framework(ORERF) based on deep learning. First, combining the maximally stable extremal regions(MSER) algorithm and self-defined post-processing algorithm to generate Oracle single radical data annotation; then, the generated Oracle radical-level annotation data set is input into the detection network, the detection network integrates multi-scale features, and uses the attention mechanism to implicitly extract Oracle single radical features, and then feeds the feature map to the detection module for radical detection; finally, we put the detected radicals to the auxiliary classifier network for recognition. The method of treating an OBI character as a composition of radicals rather than as a character category is a human-like method that can reduce the size of the vocabulary, ignore redundant information among similar characters. The experimental results are highlighted and compared to demonstrate the efficiency of the method. Furthermore, we also introduce two new datasets containing Oracle radical character dataset(ORCD) and Oracle combined-character dataset(OCCD).