Deep Learning based Handwritten Arithmetic Equation Solver

Abstract

Manual recognition of different styles of handwritten characters, digits, symbols, and operators is quite intricate. This paper attempts to automate the process of recognition of digits (0 to 9), decimal point, opening and closing parenthesis (‘(’ and ‘)’) brackets, and five binary arithmetic operators: sum (+), difference (-), product (×), exponent (A) and division in input handwritten equation followed by evaluation of expression. Firstly the segmentation based on binary threshold and contouring is applied to segregate the components of input handwritten equation. The individual segregated components are further input to Convolution Neural Networks(CNN) for recognition as part of second step, followed by evaluation of valid expression as the final step. The publicly available MNIST dataset handwritten digits (0–9) are considered and rest of the images for parenthesis, decimal point and five arithmetic operators are manually generated. The CNN model identified the handwritten expression with a high accuracy of 99% followed by a correct evaluation of the valid expression.