Recognizing Handwritten Text Lines in Ancient Document Images Based on a Gated Residual Recurrent Neural Network

Over several decades, many archives and libraries have highlighted the growing need to assist them in the preservation and enrichment of the huge mass of digitized documentary heritage by using efficient handwritten text recognition (HTR) frameworks. To address this issue, we propose in this paper a deep learning based framework for recognizing handwritten text lines in historical document images. The proposed framework is based on a gated residual recurrent neural network, called G2R2N. G2R2N is composed of two modules: encoder and decoder. The encoder module is based on merging the gated and skip connection layers, while the decoder module is composed of the bidirectional long short-term memory (BLSTM), followed by the connectionist temporal classification (CTC) architectures. The proposed framework is evaluated using the same evaluation metrics computed in the context of the ICDAR2017 competition. Numerical and qualitative observations are reported on different benchmark datasets used in the most well-known HTR contests.