A neural network-based approach for recognizing multi-font printed English characters

Abstract In this paper, we propose a method for recognizing English characters in different fonts. The proposed method based on neural network is resistant to font variant. When the samples in new fonts are added to the database, the accuracy of existing methods rapidly decreases and they are not resistant to font variant but to the accuracy of proposed method that almost stays constant and does not much decrease. A similarity measure neural network is used to identify characters and similarity measure compares the features of characters and the features of the indicators associated with the characters from A to Z obtained in the training stage. We use similarity measure instead of distance measure in SOM neural network because a person learns font-independent and a literate can read without knowing the font of the written note. In fact he/she measures similarity between the notes in new fonts and learned notes in his/her mind. Therefore, we use two samples for training the network as representative of all fonts such as default notes in man's mind. We could obtain 98.56% accuracy of recognizing a database that includes 24 different fonts in 11 different sizes.