Segmentation-free Traditional Mongolian OCR and its Public Dataset

The cursive nature of traditional Mongolian script poses a challenge for accurate character segmentation, and the presence of internal spaces within words can result in incorrect word segmentation. Consequently, an segmentation-free approach to Mongolian Optical Character Recognition (OCR) is preferred as it recognizes the entire text line, overcoming the complexities associated with both character and word segmentation. In this study, we introduce an OCR model for recognizing whole-line Mongolian text. To facilitate the development and evaluation of this model, we have compiled a database encompassing 80,000 lines of Mongolian text and 10,000 scanned images of real-world Mongolian text lines, carefully annotated with their corresponding text content. Additionally, we leverage data synthesis and augmentation techniques to augment the dataset. Furthermore, we make this dataset publicly available, alongside the text utilized for generating synthetic images, data generation code, and OCR model training configurations, solely for research purposes.