Recognition of Online Handwritten Gurmukhi Strokes using Convolutional Neural Networks

In this paper, we attempt to explore and experiment multiple variations of Convolutional Neural Networks on the basis of their distributions of trainable parameters between convolution and fully connected layers, so as to achieve a state-of-the-art recognition accuracy on a primary dataset which contains isolated stroke samples of Gurmukhi script characters produced by 190 native writers. Furthermore, we investigate the benefit of data augmentation with synthetically generated samples using an approach called stroke warping on the aforementioned dataset with three variants of a Convolutional Neural Network classifier. It has been found that this approach improves classification performance and reduces over-fitting. We extend this finding by suggesting that stroke warping helps in estimating the inherent variances induced in the original data distribution due to different writing styles and thus, increases the generalisation capacity of the classifier.