Handwritten Gurmukhi Akshara Recognition Using Convolutional Neural Network

Abstract

Handwritten Character Recognition Systems are the systems that are capable of identifying the symbols/characters drawn on graphical interface by hand using some input device. Online Handwriting Recognition System includes different stages, namely, Pre-processing, Feature Extraction, Classification and Recognizing. In Pre-processing phase, basic algorithms are used for segmentation of characters, centering/normalizing the coordinates of the character drawn, smoothing and slant correction etc. Feature Extraction is a process in which we try to extract only that relevant information from the pre-processed character that will be sufficient to classify the character drawn. It can be a low level feature or high level feature. Different classifiers are used to classify and recognize the symbol/character. This report presents a Convolution Neural Network based Gurmukhi akshara recognition system which is trained using Stochastic Diagonal Lavenberg-Marquardt Backpropogation algorithm. Convolution Neural Networks are analogous to neural networks having learnable weights and biases which evaluates raw digitized pixels to final winner class scores. A comprehensive recognition rate of 75.2% is achieved using 3-fold cross-validation strategy on a set of 10,500 images.