Deep Learning Based OCR for Mathematical Expressions

Abstract

Mathematical expression recognition with the help of a deep neural network is proposed in this work. A latex sequence is generated from this model of a given mathematical expression image. An attention aggregation-based bi-directional mutual learning network is used along with high-level and low level feature maps to make it a multiscale feature with a multiscale attention aggregation model. A multiscale feature map is used to restore the fine grained details that use to drop by pooling operations of the Dense encoder. The model was trained and tested on CROHME dataset with different num bers of classes.