MixUp Data Augmentation for Handwritten Arabic Mathematical Symbols Recognition

Handwritten mathematical expression and symbol recognition is a subfield of document image analysis that aims to convert images of handwritten mathematical formulas into a machine-readable format. Despite decades of research, recognition of handwritten mathematical expressions and symbols, particularly those written in Arabic, remains a challenging problem. To address this issue, we propose a Deep Neural Network (DNN)-based approach for Handwritten Arabic Mathematical Symbol Recognition. DNNs are powerful tools for data analysis and image classification, but they are susceptible to overfitting. Additionally, there are limited large-scale databases available for Arabic mathematical symbol recognition. To overcome this, we propose using MixUp data augmentation to increase the diversity of available data. MixUp involves training a CNN on convex combinations of pairs of examples and their corresponding labels. Furthermore, we integrate our new framework with both Cross Entropy loss and triplet loss on the augmented samples, which significantly improves the classification accuracy. The experimental results conducted on the Handwritten Arabic Mathematical Dataset (HAMF) [1] demonstrate that our proposed framework yields a significant improvement in accuracy compared to the standard CNN.