Handwritten Mathematical Symbol Recognition using Neural Network Architectures

Abstract

Mathematical symbol recognition is a topic of attention that convert physical documents into digital format. Despite the existing techniques to recognize handwritten characters and symbols, recognition accuracy is unstable. The main objective of this work is to build an intelligent system to recognize handwritten characters or symbols written in different styles with improved and stable accuracy. The proposed system can read handwritten mathematical characters or symbols as input and recognize them with corresponding characters or symbol names. The proposed implementation uses various machine learning and deep learning algorithms like Logistic Regression, Convolutional Neural networks, and Dense net. The dataset used for this research is 46 MB, which contains images of numerical values from 0 to 9, mathematical symbols, and alphabets that are available in the Kaggle open-source platform. Each data category has around 500 plus handwritten images. The implementation uses various machine learning and deep learning algorithms like Logistic Regression, Convolutional Neural networks, and Dense net to address the symbol recognition challenges. The comparative study is implemented with these algorithms and the Dense net has presented exceptional results during the training and testing phase with an accuracy of 99% and 94.2% respectively. This improved accuracy is due to the utilization of Densenet over other CNN architectures, as the DenseNet concatenates the output of the predecessor layer with the successor layer and it weakens the vanishing gradient problem. Also, the Dynamic feature propagation helps in the regulated flow of information in the dense network architecture.