Isolated Kannada Character Recognition using Densely Connected Convolutional Network

Abstract

Handwritten Character Recognition and Identification are one of the most interesting problem statements in the present digitized world because of its variety of applications. It has leveraged its potential in reducing the manual work of converting the documents containing handwritten characters to machine-readable texts. Recognition of handwritten characters is challenging due to various reasons like high variance in the writing styles across the globe, poor quality of the handwritten text compared to the printed text and the size of the handwritten text. Kannada language has a history of over 1000 years. Kannada vowels and consonants are curvy and symmetric in nature and hence recognition in an offline system becomes difficult. Hence, recognition of Handwritten Kannada characters effectively serves as the main objective of this work. This work proposes a DenseNet121 based Character Recognition model that effectively recognizes the Handwritten Kannada characters. Transfer Learning is used to improve the overall performance of the model. The proposed model achieved a training accuracy of 96.7% and test accuracy of 96.28%, hence proving the effectiveness of the model. © 2022 IEEE.