This project is the attempt to apply neural networks in Todo Bicig word recognition.

Dataset is based on existing Lamrim transliterated text. 
Lamrim page images were separated on line's and then on word's images with opencv lib and matched with transliterated words.

Validation of pairs "word/image" was made by hand with Yandex.Toloka service.

Network architecture is based on https://towardsdatascience.com/build-a-handwritten-text-recognition-system-using-tensorflow-2326a3487cd5 article.