An innovative approach to optimizing optical character recognition (OCR) systems has been developed at NUST MISIS. By using machine learning and modern generative AI models, it was possible to increase the accuracy of text recognition in Russian and reduce training time from several weeks to 72 hours.
In a context where companies are actively digitizing documents — from invoices to archives — OCR technologies are particularly relevant. However, standard systems do not always cope with real scans containing seals, signatures, or non-standard fonts. To improve quality, training is required, which traditionally takes a significant amount of time — up to two months.
Scientists at NUST MISIS have proposed a method based on a combination of machine learning and generative AI models. They implemented a closed loop of interaction between OCR engines and language models: the system independently analyzes recognition results and corrects errors. This reduced the training process to 72 hours of continuous operation.
One of the key results was an increase in recognition quality — it exceeded 90% for the Russian language, which meets modern requirements for document management automation. In addition, the cost of training models was reduced by almost a third, and the use of generative neural networks reduced the required amount of test data.
The developers tested the system on ideal documents and real scans with uneven signatures and seals. The data obtained helped to identify the most effective combinations of technologies. Promising methods based on neural networks will accelerate the creation of accurate and accessible OCR solutions for business and science.