Recognition of Basic Handwritten Math Symbols Using Convolutional Neural Network with Data Augmentation

Abstract

In the ever-updating digital world, automatic handwritten math symbols classification (HMC) plays many vital roles in the advancement of computer-aided systems. It is the main foundation of perfecting one of the most challenging tasks out there: recognizing handwritten mathematical formulas. As with the other similar automated handwritten characters classifications tasks, HMC also faces various difficulties while attempting to correctly classify images. As people tend to have distinct types of handwriting styles and unique ways to write symbols, a simple character may have infinite versions of itself. In our research, we focused on the classification of such images of numerous handwritten mathematical symbols. For this classification, we have developed a convolutional neural network (CNN) model and worked with three different datasets to test our model’s efficiency. We introduced different data augmentation techniques to construct various versions of the already available images. This created a virtual mimicry of people’s tendency to write the same character in many styles. Our CNN model of 11 layers (6 were convolutional layers) worked to classify 16 classes (each denoting a mathematical symbol or digit) and had an accuracy of 98.71%, 99.01%, and 99.85% respectively on three publicly available datasets. To our knowledge, our model performed better than every other research work in this field. Considering this remarkable success, we are bent on working further on this and creating a fully working app that would eventually be able to automatically classify handwritten mathematical formulas.