Improving Handwritten Mathematical Expression Recognition with CNN by Exploring Augmentation, Regularization, and Training Strategies

Abstract

This scholarly investigation constructs a framework predicated on Convolutional Neural Networks (CNN) for the recognition of handwriting in the context of mathematical problem-solving at the elementary educational tier, encompassing numerals 0-9 and fundamental mathematical symbols including addition, subtraction, multiplication, and division. Employing a CNN-based Sequencing Architecture, the developed system demonstrates the capability to identify, organize, and structure handwritten mathematical expressions authored by students with an impressive accuracy rate of 99.07%. The implementation of this system accelerates the evaluation process, mitigates human errors associated with grading, and furnishes immediate feedback. Furthermore, this research delves into the formulation and examination of diverse data augmentation methodologies, regularization techniques, and training strategies aimed at enhancing the model’s performance. The application of augmentations such as rotation, flipping, and image displacement, along with regularization measures incorporating dropout and batch normalization, significantly bolsters the model’s capacity for generalization. The overarching objective of this research is to augment the efficiency of mathematics learning assessments and to advocate for the integration of artificial intelligence technologies within primary educational settings.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint