MSdocTr-Lite: A lite transformer for full page multi-script handwriting recognition

The Transformer has quickly become the dominant architecture for various pattern recognition tasks due to its capacity long-range representation. However, transformers are data-hungry models and need large datasets training. In Handwritten Text Recognition (HTR), collecting a massive amount of labeled data is complicated expensive task. this paper, we propose lite transformer full-page multi-script handwriting recognition. proposed model comes with three advantages: First, solve common problem scarcity, that can be trained on reasonable data, which case most HTR public datasets, without external data. Second, it learn reading order at page-level thanks curriculum learning strategy, allowing avoid line segmentation errors, exploit larger context reduce costly annotations. Third, easily adapted other scripts by applying simple transfer-learning process using only images. Extensive experiments different (French, English, Spanish, Arabic) show effectiveness model.