Improving Handwritten Cyrillic OCR by Font-based Synthetic Text Generator

Abstract

In this paper, we propose a straight-forward and effective Font-based Synthetic Text Generator (FbSTG) to alleviate the need for annotated data required for not just Cyrillic handwritten text recognition. Unlike standard GAN-based methods, the FbSTG does not have to be trained to learn new characters and styles; all it needs is the fonts, the text, and sampled page backgrounds. In order to show the benefits of the newly proposed method, we train and test two different OCR systems (Tesseract, and TrOCR) on the Handwritten Kazakh and Russian dataset (HKR) both with and without synthetic data. Besides, we evaluate both systems' performance on a private NKVD dataset containing historical documents from Ukraine with a high amount of out-of-vocabulary (OoV) words representing an extremely challenging task for current state-of-the-art methods. We decreased the CER and WER significantly by adding the synthetic data with the TrOCR-Base-384 model on both datasets. More precisely, we reduced the relative error in terms of CER / WER on (i) HKR-Test1 with OoV samples by around 20% / 10%, and (ii) NKVD dataset by 24% CER and 8% WER. The FbSTG code is available at: https://github.com/mhlzcu/doc_gen.