Bangla Handwritten Words Recognition using Convolutional Neural Network

Convolutional Neural Network has become the advanced stage of technology for image recognition. It has been widely used for offline recognition of handwritten words of different languages such as English, Arabic, etc. This paper proposes a system for recognizing Bengali handwritten isolated characters and words using Convolutional Neural Network (CNN). To introduce an automated process that can be used to recognize a particular Bengali handwritten word to find the equivalent machine-readable format of that word is the actual purpose of this system. Based on the internal construction of the system, three different recognizing models have been introduced in this experiment. The 2-Layer CNN model, the Pure CNN model, and the 3-Layer CNN model have undergone several evaluation processes throughout the whole experiment to determine a stable model with satisfactory recognition accuracy. The 2-Layer CNN model has the highest accuracy of 72% when trained with the selected dataset. When trained with the same dataset, the Pure CNN model has produced 80% recognition accuracy for only ten epochs and the 3-Layer CNN model has produced a maximum 97% recognition accuracy. The models have also been trained and evaluated using the MNIST and Bangla-Lekha-Isolated datasets.