Classification of Bangla Compound Characters Using a HOG-CNN Hybrid Model

Abstract

Automatic handwriting recognition is challenging task due to its sheer variety of acceptable stylistic differences. This is especially true for scripts with large character sets. Bangla, the sixth most widely spoken language in the world has a complex, large and rich set of compound characters. In this study, a hybrid deep learning model is proposed which combines the use of the manually designed feature Histogram of Oriented Gradients (HOG), with the adaptively learned features of a Convolutional Neural Networks (CNN). The proposed hybrid model was trained on the CMATERDB 3.1.3.3, a Bangla compound character data set which divides Bangla compound characters into 177 broad classes and 199 specific classes. The results demonstrate that CNN-only models achieve over 91% and 92% test accuracy respectively. Furthermore, it is shown that the proposed model, which incorporates HOG features with a CNN, achieves over 92.50% test accuracy on each division. While there is still room for improvement, these results are significantly better than currently published state of art on this data set.