Handwritten Character Recognition of Kannada Language Using Convolutional Neural Networks and Transfer Learning

Abstract

Recognition of Kannada Handwritten Characters in recent times is one among the active research fields of study. It is one of the challenging topics in the pattern recognition field due vocabulary in large scale, complicated structural hierarchy and different people have different handwriting styles. In this study, we put forward a technique for recognition of characters in kannada which are handwritten based on convolutional neural networks and transfer learning. We have first pre-processed each character to remove noise, cropped each character image and resize the images. After the pre-processing the enhanced pixel values in the image helps in training of the neural network for an efficient classification of characters to their respective classes. We have considered vowels, consonants and numerals of Kannada language which are handwritten. Even though some of the characters in kannada language have similarities in structures our model is capable of classifying the character to their correct respective class. We have also used transfer learning to further improve the model so that any new style of handwritten character can be correctly predicted. Our method is efficient in neural network approach and it achieves competitive performance on the rest of the other traditional recognition methods in the literature.