Manchu script recognition surges with combined data

Researchers achieved up to 98.27% accuracy in recognizing Manchu script by combining artificial and real historical images.

Manchu was a key language during China’s Qing dynasty (1636-1912) and a wealth of archival documents in the language are now being digitised. Accessing these records, however, requires accurate Optical Character Recognition (OCR) technology, which is challenging given the limited availability of labelled Manchu text. This research explored how to best utilise both computer-generated and genuine historical images to train OCR systems.

The team tested four different approaches using 60,000 synthetic and 20,306 real Manchu word images. They evaluated three vision-language models and a convolutional recurrent neural network (CRNN) using training that relied solely on synthetic data, solely on real data, a combination of both, and sequential training beginning with synthetic data. Adding real images to the training process significantly boosted accuracy, with the best configurations reaching between 95.09% and 96.28%.

While synthetic data alone achieved up to 87.92% accuracy, it consistently improved the performance of all three vision-language models when combined with real images. The CRNN model showed a smaller benefit from synthetic data, depending on how the training was structured. Importantly, the study found that a smaller, more efficient CRNN model could perform as well as larger models when trained with real images.

The researchers also demonstrated that combining the outputs of multiple high-performing recognisers, using a voting system, could push accuracy even further, to 98.27%, without requiring any additional training. They used an eighteenth-century Manchu dictionary to resolve disagreements between the systems. This approach offers a promising pathway to unlock the vast historical information contained within Manchu archives.