CNN-BLSTM Model for Arabic Text Recognition in Unconstrained Captured Identity Documents

Optical Character Recognition (OCR) for Arabic text (printed and handwritten) has been widely studied by researchers in the last two decades. Some commercial solutions have emerged with good recognition rates for printed text (on white or uniform backgrounds) or handwritten text with limited vocabulary. In addition to being naturally cursive, the Arabic language comes with additional challenges due to its calligraphy resulting in a variety of fonts and styles. In this work, recent advances in recurrent neural networks are explored for the recognition of Arabic text in identity documents captured in the wild. The unconstrained captures bring additional difficulties as the text has to be first localized before being able to recognize it. Various pre-processing steps are introduced to overcome the difficulties related to the Arabic text itself and also due to the capturing conditions. The presented approach outperforms existing solutions when evaluated using a private dataset and also using the recent MIDV2020 dataset.