Unconstrained online handwritten Uyghur word recognition based on recurrent neural networks and connectionist temporal classification

This paper conducts the first experiments applying recurrent neural networks-RNN accompanied with connectionist temporal classification (CTC) to build end-to-end online Uyghur handwriting word recognition system. The traced pen-tip trajectory is fed to network without conducting segmentation and feature extraction. The network is trained to transcribe handwritten word trajectory to a string of characters in alphabet which has total 128 character forms. In order to avoid overfitting during training and improve generalisation of the model, dropout technique is implemented. An online handwritten word dataset has been established and used for model training and evaluation in writer independent manner. Recognition results are evaluated by calculating the Levenshtein-edit distance and 14.73% character error rate CER on test set of 3,600 samples for 900 word classes has been observed without help of any lexicon search and language model.