Augmentation based Convolutional Neural Network for recognition of Handwritten Gujarati Characters
Pritesh J. Borad, Parth J. Dethaliya, Anand Mehta · 2020 IEEE International Conference for Innovation in Technology (INOCON) · 2020
The aim of this research is to construct handwritten Gujarati characters' dataset as well as their recognition using Convolutional Neural Network (CNN). Handwritten Character Recognition (HCR) is an electronic translation of handwritten text to editable machine text. It is more challenging due to lots of variation in writing style, characters' thickness, and curves of different age groups. The dataset is collected from the primary school in Gujarat, India and preprocessed in MATLAB using various techniques, such as Segmentation, Equalization, Skeletonization, Dilation, and Merging. Deep Learning techniques can be utilized to overcome the various challenges which are faced while recognizing handwritten characters. Therefore, CNN with Dropouts, Augmentation and Multi-Layer Perceptron (MLP) is employed as classifier. The proposed system has achieved maximum training accuracy of 98.6% and testing accuracy of 94.8%.