Ubicomp Digital 2020 -- Handwriting classification using a convolutional recurrent network

TL;DR

This paper presents a CNN-LSTM neural network for classifying multi-variate handwriting time series data, achieving competitive accuracy in a challenge setting with Arabic letter data.

Contribution

It introduces a hybrid CNN-LSTM architecture specifically designed for handwriting classification of multivariate time series data.

Findings

Achieved 68% accuracy on writer-exclusive test set

Achieved 64.6% accuracy on blind challenge test set

Secured second place in the challenge

Abstract

The Ubicomp Digital 2020 -- Time Series Classification Challenge from STABILO is a challenge about multi-variate time series classification. The data collected from 100 volunteer writers, and contains 15 features measured with multiple sensors on a pen. In this paper,we use a neural network to classify the data into 52 classes, that is lower and upper cases of Arabic letters. The proposed architecture of the neural network a is CNN-LSTM network. It combines convolutional neural network (CNN) for short term context with along short term memory layer (LSTM) for also long term dependencies. We reached an accuracy of 68% on our writer exclusive test set and64.6% on the blind challenge test set resulting in the second place.