ComNet: A Deep Convolutional Neural Network Capable of Classifying Compound Bengali Handwritten Characters High Number of Classes in a Data-Scarce State

Handwritten character recognition has been a well-studied problem for decades among researchers. Even though there exist character recognition systems for some non-cursive languages like English, this hasn’t been as much successful for cursive languages yet. Bengali, being a cursive language, has another complication up its sleeve which is numerous number of compound characters. These facts make it very difficult to classify accurately even for a modern Deep Convolutional Neural Network (CNN). Deep CNNs thrive when plenty of data are available to optimize their parameters and don’t perform so well in a data-scarce situation. In this paper, we studied both the aspects, of data scarcity and the cursive compound nature of the Bengali language. We created a Deep CNN called ComNet with EfficientNetB7 and an additional regularization layer, which achieved an accuracy of 92.42% and outperformed baseline DenseNet201 architecture which managed only 79.42% accuracy on our in-house 267 class augmented compound handwritten character dataset called Juktakkhor. Without augmentation, deep learning algorithms suffer. ComNet also suffers since the dataset has a large number of classes with quite a low number of data per class. ComNet achieved 41.16% accuracy on the non-augmented Juktakkhor dataset, while baseline DenseNet201 achieved 39.89%.