Handwritten Mathematical Expression Recognition Using Convolutional Neural Network

Giang Son Tran, Chi-Kien Huynh, Thanh-Sach Le, Tan-Phuc Phan, Khanh-Ngoc Bui

Abstract: Recognizing mathematical expressions on raster images usually consists of two steps: detecting individual symbols and analyzing their spatial structure to form a coherent equation. In this work, we focus on the first step and propose a detection method that is able to locate small and difficult handwritten symbols. We use a deep convolutional neural network with robust detection performance. It is able to achieve a mean average precision score of 0.65 for 106 different mathematical symbols on the dataset we created. For structural analysis, we use the DRACULAE parser since it has high accuracy given that the symbols were correctly detected.