Real-Time Mobile Application for Handwritten Digit Recognition Using MobileNet

Abstract

Real-time mobile application for handwritten digit recognition helps children learn how to write a single digit at their own pace, in a fun and straight forward approach. It can improve and strengthen the children’s knowledge and skills in writing digits. One of the problems children have in learning how to write digits is mirror-writing in which the digits or numbers are written as if they are a reflection from a mirror. Convolutional Neural Networks (CNNs) have shown tremendous performance on mobile devices, including Android and iOS, with low computational cost and yet producing high recognition accuracy. Two popular CNNs for mobile applications are MobileNet and ShuffleNet. This project reports our preliminary investigation, comparing the real-time mobile application performance of MobileNet and ShuffleNet for handwritten digit recognition. The preliminary experiment involving training these models with some randomly selected images from the MNIST dataset indicates that MobileNet produces the accuracy of 0.9442 while ShuffleNet only achieves 0.6883. Thus, our mobile application employs MobileNet for the handwritten digit recognition with a simple and user-friendly interface to be tested by children of four to six years old. In this application, children can write using one of their fingers or a stylus using the mobile phone. The findings show that MobileNet is beneficial for a real-time mobile application for children to learn writing digits.