Character type based online handwritten Uyghur word recognition using recurrent neural network

Abstract

This paper presents a character type based approach for online handwritten Uyghur word recognition. According to the writing characteristics of Uyghur, a word can have two different Unicode transcriptions, either based on 32\u2009+\u20092 basic character types or 128 specific character forms. However, the previous studies on unconstrained Uyghur handwritten word recognition have been based on character forms, while the character types have been out of attention. This paper carries out comparative handwriting Uyghur word recognition experiments. The character type and character form based recognition systems are built using recurrent neural networks and connectionist temporal classification. Each system is trained to transcribe the handwritten word trajectory to a string of characters, either based on the character types or the character forms. The handwritten trajectory is fed to the recognition system without explicit or implicit character segmentation. Dropout regularization is implemented to avoid over-fitting during training and improve model generalization. The training process and recognition results show Character type based system outperformed the character form based system in terms of speed and recognition results. Character type based system provided 13.96% character error rate on the our test set which exceeded the character form based system by 0.76 points.