Comparative Analysis of Convolutional Neural Network and Character Recognition Techniques for Handwritten Mathematical Equation Solver

Main Article Content

Navaneetha Krishnan M, Swetha G, Saravanakumar R, Nandhinishri S.N, Arthi R

Abstract

Solving mathematical equations manually can pose various difficulties, depending on the complexity of the equation and the level of mathematical knowledge and skills of the individual like Lack of understanding, arithmetic errors, incorrect assumptions, and complex equations. This process involves recognizing handwritten characters, which is a difficult task digitizing handwritten text due to a variety of factors. This project aims to develop a system that can automatically identify and solve handwritten mathematical equations using the CROHME dataset by CNN and LeNet character recognition technique. The CROHME dataset contains handwritten characters, including mathematical symbols, which will be used to train a CNN. The LeNet architecture was one of the earliest successful applications of deep learning for image acknowledgement and has since become a classic model in the arena of computer vision. The system will take an image of a handwritten mathematical equation as input and use the trained CNN to recognize the characters and symbols in the equation. Once the equation is recognized, it will be solved using mathematical operations, and the output will be displayed. The results of this project have the potential to be used in educational and professional settings to help individuals solve mathematical problems more efficiently. In experimental analysis, the comparative analysis was conducted by both learning and deep learning algorithms such as CNN, SVM, Decision Tree, and Naive Bayes, and character recognition techniques such as LeNet, OCR, VGG, and ResNet. However, the CNN algorithm gives a better performance rate with an accuracy of 97.56%, specificity of 96%, sensitivity of 96%, and F-score of 97% respectively. The LeNet character recognition technique gives better performance and it produces 98% accuracy compared with OCR, ResNet, and VGG.

Article Details

Section
Articles