Performance Evaluation of Different Algorithms for Handwritten Isolated Bangla Character Recognition

Handwritten character recognition is one of the most emerging fields within the optical character recognition area. Bangla handwritten character recognition is a complex task, it is challenging due to extensive size and diversity within the alphabets. Currently, convolutional neural network (CNN) has been proven to have the ability to classify complex dataset. The convolutional neural network does not require any predefined feature extraction method, but it requires a large dataset to gain accuracy. This work proposes a convolutional neural network model for classifying Bangla handwritten alphabets and compares the performance with the other widely used models for classification. Each model is trained with a large dataset, which is augmented to have diversity in data and features. After training, we have tested the models with 7500 sample images and it shows an accuracy of 97.87% for the proposed model. In this work, we also find out the weights of the CNN network for best performance and used that weights to evaluate the performance from other data set for cross-validation of our model. The weighted model accuracy for two different independent data set is 95.23% and 94.22%.