A Script Independent Hybrid Feature Extraction Technique for Offline Handwritten Devanagari and Bangla Character Recognition

Abstract

Recognizing handwritten characters plays a significant role in different applications of pattern recognition. That is why the digital representation of character images is much necessary to design an efficient offline Optical Handwritten Character Recognition System (Offline HCR). Here a hybrid feature representation method is suggested for two Indic scripts, such as Devanagari and Bangla. The method utilizes three different features to represent any character images. Those are angular motion of character shape-based feature, center to the thin text of character shape-based feature, and center to edge text of character shape-based feature. After collecting all these three features, these are applied to various machine learning algorithms, including two modified neural network models. One simple traditional convolutional neural network is also designed which takes immediate images and recognizes the character images. Although the two modified neural network models are unable to hit the peak in terms of accuracy like the traditional CNN, it is found that two of our modified NN models take quite less time to execute upon the character datasets.