KGAN: A Generative Adversarial Network Augmented Convolution Neural Network Model for Recognizing Kannada Language Digits

Abstract

AbstractKannada is a south Indian language with a history of two thousand years and spoken by more than sixty million people. Kannada language has its own script for alphabets and digit representations. So there is a need for convolution neural network (CNN) model to recognize Kannada language scripts. This paper presents a design of a CNN model to recognize Kannada digits. One of the challenges faced while designing a CNN model is data over fitting. Data over fitting is a phenomenon where the trained model arrives at parameter values such that they can classify only the instances provided during training resulting in reduction of accuracy for a new unseen test instance. To overcome this problem, datasets are split into train and test sets. The detriment of this system is lesser number of instances to train the CNN. Increasing the number of training instances is a good approach, but the complexity in data collection is to be answered. In this paper, we explore generative adversarial network (GAN) as an additional data generator and its suitability. Results of analysis on the experiment revealed the following advantages; first, the data augmentation has a positive impact on CNN, next, GAN-generated data meets qualitative requirement as train and test dataset and last, epoch value for training CNN has influence on data under fitting and data over fitting phenomenon.KeywordsKannada languageGenerative adversarial networkConvolution neural networkMNIST dataset