Recognition of Printed Mathematical Formula Symbols Based on Convolutional Neural Network

Zi-long WAN, Kai FAN, Qi-hui WANG, Sheng-nan ZHANG


Printed mathematical formula recognition is a topic research region in OCR. But the diversity of fonts and sizes of mathematical symbols, as well as incorrect segmentation and stroke destruction of touching symbols lead to the difficulty of feature extraction and low recognition rate. In this paper, we constructed a convolutional neural network to recognize formula symbols, and determined the optimal parameters of the network through a large number of comparative experiments. Two convolution layers and sampling layers deepened the number of network layers and improved the recognition rate to a certain extent. Convolution kernels with fixed size extracted gradient information effectively, and ReLU activation function and dropout connection mode reduced the degree of over-fitting and gained the better generalization ability of the network. The experimental results show that the presented method can improve the recognition of printed formula symbols.


Convolutional neural network, Formula symbol recognition, Deep learning.


Full Text:



  • There are currently no refbacks.