Generator Based Methods for Off-line Handwritten Character Recognition

We present a novel method using Convolutional Autoencoder and Generative Adversarial Network for off-line classification of handwritten alphabets and numerals of different scripts namely English, Devanagari and Kannada. Here we have used the concept of generation of a predefined view from any novel view as used in facial recognition problem, for handwritten character recognition (HCR) challenge. Our investigations in Autoencoder based method show that the reduced dimension representation of the dataset is not quite clusterable thus reducing the classification accuracy. So we used Convolutional Autoencoder for dimension reduction of the data. We then further added Adversarial Learning to the Convolutional Autoencoder model using GAN architecture. On average we achieved better results with the GAN based methods. A detailed discussion justifying the reasons behind this is presented with the help of observations from our experiments in the main article. Quantitative comparison of our results with state of the art is also presented in the article.