Manchu ancient books carry the historical context and cultural essence of the Qing Dynasty, serving as an indispensable and important heritage in the development of Chinese civilization. Due to the challenges in interpreting these texts, particularly the difficulty in recognizing Manchu words, the study of Manchu word recognition based on deep learning is particularly important. However, deep learning models highly dependend on high-quality and large-scale annotated datasets during training. To address this, this study constructs a high-quality and diverse dataset of Manchu ancient book words. Given the limitations of traditional segmentation methods for data collection, such as the high costs and time-consuming nature of manual annotation, this dataset adopts a semi-automated approach, combining computer-assisted technology and manual verification to achieve effective word extraction. On this basis, we further introduced an image quality screening and annotation review mechanism to systematically eliminate blurred, tilted and miscut samples to ensure the clarity of images and the accuracy and consistency of annotations. Based on the Series of Rare Ancient Books in Manchu and Chinese Housed in the National Library, this study extracts a total of 24,280 word images from Manchu ancient books, covering 2,428 unique Manchu words, with 10 images corresponding to each word. In the process of data collection and collation, we found that this dataset is the largest collection of Manchu word images so far, providing a solid data foundation for related research.
SUN et al. (Sun,) studied this question.