Exploration of Deep Convolutional Neural Networks(Via Transfer Learning) for Handwritten Character Recognition

Recognition of handwritten characters is one of the difficult and challenging task because of the variation of characters in size, style and pattern. The complexity increased further with dictionary (alphabets, numerals, special characters), more individuals, age groups, and also with working environment. The exploration of open-source pre-trained networks for the classification of characters was minimal. This motivated us to explore the pre-trained deep convolutional networks (Alexnet, VGG-16, Resnet-50), and fine-tune them to recognize the handwritten characters using transfer learning. The experimentation results of widely used database EMNIST using pre-trained networks are in-par with the results of the state-of-art customized networks,which is specific to database and language. The classification accuracy of Resnet-50 for EMNIST (By-class: 87.24%, By-merge: 90.64%, Balanced: 89.18%, Letters: 94.90%, Digits: 99.57%).