Extending OCR Model for Font and Style Classification

Abstract

Font and style classification aims to recognize which font and which style the characters in the input image belong to. The con- junction of font and style classification with traditional OCR systems is important in the reconstruction of visually-rich documents. However, the current text recognition systems have yet to take into account these tasks and focus solely on the recognition of characters from input images. The separation of these tasks makes the document reconstruction systems computationally expensive. In this paper, we propose a new approach that extends the current text recognition model to include font and style classification. We also present a dataset comprising input images and corresponding characters, fonts, and styles in Vietnamese. We evaluate the effectiveness of this extension on multiple recent OCR models, including VST [10], CRNN [8], ViSTR [1], TROCR [7], SVTR [4] . Our results demonstrate that our extension achieves decent accuracy rates of 98.1% and 90% for font and style classification, respectively. Moreover, our extension can even boost the performance of the original OCR models.