Offline handwritten mathematical expression evaluator using convolutional neural network

Abstract

Recognition of Offline Handwritten Mathematical Expression (HME) is a complicated task in the field of computer vision. The proposed method in this paper follows three steps: segmentation, recognition and evaluation of the HME image (which may include multiple mathematical expressions and linear equations). The segmentation of symbols from image incorporates a novel pre-contour filtration technique to remove distortions from segmented symbols. Then, recognition of segmented symbols is done using Convolutional Neural Network which is trained on an augmented dataset prepared from EMNIST and custom-built dataset giving an accuracy of 97% in recognizing the symbols correctly. Finally, the expressions/equations are evaluated by tokenizing, converting into postfix expressions and then solving using a custom-built parser.