Study of Transfer Learning Technique for Handwritten Devanagari Character Recognition

In recent years, many researchers have investigated the techniques for recognition of handwritten letters from images. It is an important application for historical document analysis and other applications like content reading from images. The recognition of characters from handwritten images is a challenging task due to huge variation in handwriting from person to person. The transfer learning technique have been used for performing such classification with higher accuracy. This paper proposes Devanagari script recognition using the Convolutional Neural Network (CNN) model. The transfer learning technique is used to tune the pretrained models for classification. Three CNN architectures, VGG16, VGG19 and ResNet have been used for classification and their performances have been compared. The proposed model using VGG16,VGG19 and ResNet is observed to demonstrate higher accuracy of 99.07%, 99.17%, and 99.17% as compared to the existing works.