Evaluation of Bengali Handwritten Mathematical Equationusing Convolutional Neural Network

Abstract

Abstract Humans are naturally capable of solving mathematical expressions, but machines lack the abilityto comprehend an issue through a visual context. Computers are gradually becoming moreadvanced and catching up with the subtlety and inaccuracy of real life. The need for an automatedsystem to check answer scripts of mathematical equations has become unparallel, especiallyfor Bengali handwritten scripts. This study checks each line of the solution of a mathematicalequation to evaluate its correctness using a deep learning approach. In contrast to earliermethods, this paper introduces a CNN architecture to verify the accuracy of a handwritten mathematicalequation in addition to solving the problem. The model reads a handwritten equationand validates its mathematical symbols and operations. A dataset has been created to evaluatethe models performance which is named ”BHQED”. The experimental result shows that theaccuracy of the proposed CNN architecture is 92.25% and the recall is 90.65% on our solelycreated dataset. To further boost the performance, this study applies the pretrained ResNet18model and substantially outperforms the CNN with an accuracy of 94.57% and recall of 93.69%.