Character-Based Manchu Segmentation and Recognition Method

In recent years, advancements in artificial intelligence have significantly improved Optical Character Recognition (OCR) for various languages and scripts. Manchu, the official language of the Qing Dynasty, holds valuable historical and cultural information but presents challenges for OCR due to its vertical writing style, character connectivity, and limited resources. This paper presents a character-based segmentation and recognition method for Manchu script. The method uses an improved DBNet algorithm for segmentation and an enhanced Vision Transformer (ViT) algorithm for recognition, ensuring efficient detection and accurate recognition of Manchu characters. We created a dataset for the models through systematic data collection and used data augmentation techniques, including geometric transformations and CycleGAN-based style transfer, to enhance the model's robustness across different styles. The proposed approach shows significant improvements in segmentation and recognition accuracy, outperforming traditional methods. Experimental results show a detection accuracy of 93.5%, a recognition accuracy of 95.5%, and an overall model recognition rate of 91.3%. These findings contribute to the digitalization and automation of Manchu document recognition, aiding the preservation and accessibility of Manchu cultural heritage.