Recognizing offline handwritten mathematical expressions efficiently

Abstract

In this study, we propose a system that can recognize offline handwritten mathematical expressions using limited training data. The purpose of the system is to exhibit a high recognition accuracy with a few training data and to allow everyone to form a recognition module with a small sample of his or her own handwriting. The system comprises three main parts: segmentation, symbol recognition, and structural analysis. A recursive cortical network is used to form the recognition part of the system and a new type of linked list is proposed to analyze the complex structure of the expressions. We prepared 400 real handwritten mathematical expressions from 20 different people, containing a total of 60,103 symbols from 100 symbol classes to evaluate the performance. The system was trained using one image per class and achieved 80% accuracy on the correct segmentation result.