HRRegionNet: Chinese Character Segmentation in Historical Documents with Regional Awareness

Abstract

Human beings, as the only species capable of developing high levels of civilization, the transmission of knowledge from historical documents plays an indispensable role in this process. The amount of historical documents accumulated in the last centuries is not to be belittled, and the knowledge they contain is not to be underestimated. However, these historical documents are also facing difficulties in preservation due to various factors. The digitization process was mostly performed manually in the past, but the costs made the process very slow and challenging, so how to automate the digitization process has been the focus of much research previously. The digitization of Chinese historical documents can divide into two main stages: Chinese character segmentation and Chinese character recognition. This study will only focus on Chinese character segmentation in historical documents because only accurate segmentation results can achieve high accuracy in Chinese character recognition. In this research, we further improve the model based on our previously proposed Chinese character detection model, HRCenterNet, by adding a transposed convolution module to restore the output to a higher resolution and use multi-resolution aggregation combine features in different resolutions. In addition, we also propose a new objective function such that the model can more comprehensively consider the features needed to segment Chinese characters during the learning process. In the MTHv2 dataset, our model achieves an IoU score of 0.862 and reaches state-of-the-art. Our source code is available on https://github.com/Tverous/HRRegionNet.