Gabor Features Assist Semantic Feature Learning for Handwritten Formula Symbol Recognition

Abstract

At present, the recognition of formula symbols is very challenging. Since a large number of similarities and a variety of styles in the standard library, the overall recognition rate of formula symbols is difficult to further improve. In order to deal with these problems, this paper proposes an off-line multi-directional feature fusion decision discriminant algorithm, called MFFD. The novelty lies in the construction of the MMFD based on the convolutional neural network. In addition, we explore the directional gradient feature that facilitates the classification of handwritten formula symbols. The standard mathematical formula symbol library provided by the CROHME would verify the proposed algorithm. The error rates of CROHME2014 and CROHME2016 are 8.28% and 6.88%, respectively, which is higher than that of existing algorithms.