Gated Convolutional Recurrent Neural Networks for Multilingual Handwriting Recognition

In this paper, we propose a new neural network architecture for state-of-the-art handwriting recognition, alternative to multi-dimensional long short-term memory (MD-LSTM) recurrent neural networks. The model is based on a convolutional encoder of the input images, and a bidirectional LSTM decoder predicting character sequences. In this paradigm, we aim at producing generic, multilingual and reusable features with the convolutional encoder, leveraging more data for transfer learning. The architecture is also motivated by the need for a fast training on GPUs, and the requirement of a fast decoding on CPUs. The main contribution of this paper lies in the convolutional gates in the encoder, enabling hierarchical context-sensitive feature extraction. The experiments on a large benchmark including seven languages show a consistent and significant improvement of the proposed approach over our previous production systems. We also report state-of-the-art results on line and paragraph level recognition on the IAM and Rimes databases.