Bangla Handwritten Character Recognition Using Deep Convolutional Autoencoder Neural Network

Handwritten Character Recognition (HCR) is one of the most challenging tasks in the field of pattern recognition and computer vision. In image classification, performance can be improved by representing data from a high feature space to a low feature space. The Convolutional Neural Network (CNN) aims to reduce the images into a form without losing any features. In this paper, we introduce an Autoencoder with a Deep CNN, which we call DConvAENNet for recognizing Bangla Handwritten Character (BHC). A total of 22 experiments were performed on the three-character datasets (BanglaLekha-Isolated, CMATERdb 3.1, Ekush). All attempts acquire satisfying results up to 90% accuracy for the recognition of Bangla handwritten numerals, vowels, consonants, compound characters, modifiers, and all characters set unitedly. Using this supervised and unsupervised learning technique, our proposed DConvAENNet model achieved 95.21% on BanglaLekha-Isolated for 84 classes, 92.40% on CMATERdb 3.1 for 238 classes, and 95.53% on Ekush for 122 classes. Most errors in our model were caused due to the similarity and high curvature nature of the BHC sets.