A comparison between hybrid models for classifying Bangla isolated basic characters

Abstract

Merging Histogram of Oriented Gradients (HoG) features with a typical Convolutional Neural Network (CNN) model demonstrates some improvements in Bangla numerals and compound characters classification. Despite the importance of Bangla isolated basic characters, a hybrid model for classifying those characters are still missing. In this work, a hybrid model has been introduced, where the proposed model has been trained and validated with 50 isolated Bangla basic character classes from an open sourced Bangla benchmark dataset. In addition, a comparison between CNN and three hybrid classifiers with different handcrafted features(HoG, Local Binary Patterns (LBP), and All-Pixel) has been shown. Moreover, the hybrid model with HoG feature extraction method outperforms the other classifier models. Hybrid-HoG models demonstrate a maximum accuracy of 83.69% on its 37 iterations, while the CNN model, LBP& All pixel based hybrid classifier models were able to obtain maximum accuracies of 83.64%, 83.62%, and 83.35% respectively.