Stacked Generalization Ensemble Method to Classify Bangla Handwritten Character

Abstract

Image classification is one of the fundamental tasks in the field of machine learning. This study focuses on Bangla handwritten character recognition (BHCR). Recognizing handwritten characters has numerous applications like optical character recognition, post office automation, signboard recognition, number plate recognition, etc. A system that can classify Bangla handwriting on the character level efficiently on a large scale would surely be a primary step to achieve those autonomous user experiences. This study has developed a stacked generalization ensemble system that consists of six convolutional neural networks (CNN). The ensembled CNN model of this study has been able to perform more accurately than the individual CNN models. It can recognize 122 different Bangla handwritten characters with a test accuracy of 96.72% and an f1-score of 96.70%. The system has also outperformed most of the existing works by achieving better performance and by recognizing more handwritten characters.