CALText: Contextual Attention Localization for Offline Handwritten Text

Recognition of Arabic-like scripts such as Persian and Urdu is more challenging than Latin-based scripts. This due to the presence a two-dimensional structure, context-dependent character shapes, spaces overlaps, placement diacritics. We present an attention based encoder-decoder model that learns read handwritten text in context. A novel localization penalty introduced encourage attend only one location at time when recognizing next character. In addition, we comprehensively refine complete publicly available dataset terms ground-truth annotations. evaluate on both Arabic datasets. For Urdu, contextual achieves $$82.06\%$$ recognition rate $$51.97\%$$ word which represent $$2\times $$ improvement over existing bi-directional LSTM models. Arabic, outperforms multi-directional models with $$77.47\%$$ $$37.66\%$$ without performing any slant or skew correction. Code pre-trained for this work are https://github.com/nazar-khan/CALText .