Applying Long-Short Term Memory Recurrent Neural Networks for Real-Time Stroke Recognition

Abstract

This note discusses how to build a real-time recognizer for stroke gestures based on Long Short Term Memory Recurrent Neural Networks. The recognizer provides both the gesture class prediction and the completion percentage estimation for each point in the stroke while the user is performing it. We considered the stroke vocabulary of the $1 and $N datasets, and we defined four different architectures. We trained them using synthetic data, and we assessed the recognition accuracy on the original $1 and $N datasets. The results show an accuracy comparable with state of the art approaches classifying the stroke when completed, and a good precision in the completion percentage estimation.