Recognition of Online Handwriting with Variability on Smart Devices

Abstract

Online handwriting recognition enables a convenient user interface for smart devices. However, readability of the handwriting varies based on the device user and the writing surface. State of art handwriting recognition systems are not yet robust with respect to different variability in writing text which results from user movement, writing speed, available space to write on, sloppiness etc. In this work we employ hybrid deep neural network architectures based on Bidirectional Long Short Term Memory (BLSTM) and Connectionist Temporal Classification (CTC) networks to recognize user independent dynamic handwriting with variability. Evaluated on IAMOnDB database, we show that our proposed model achieves state of the art accuracy with minimal data preprocessing and recognizes text with variability. We further compress the model using knowledge distillation to deploy on resource constrained smart devices. Our novel strategy of training student model for CTC networks enjoys significant model size reduction with moderate performance degradation.