Nepali Handwritten Letter Generation using GAN

Abstract

The generative adversarial networks seem to work very effectively for training generative deep neural networks. The aim is to generate Nepali Handwritten letters using adversarial training in raster image format. Deep Convolutional generative network is used to generate Nepali handwritten letters. Proposed generative adversarial model that works on Devanagari 36 classes, each having 10,000 images, generates the Nepali Handwritten Letters that are similar to the real-life data-set of total size 360,000 images. The generated letters are obtained by simultaneously training the generator and discriminator of the network. Constructed discriminator networks and generator networks both have five convolution layers and the activation function is chosen such that generator networks generate the image and discriminator networks check if the generated image is similar to a real-life image dataset. To measure the quantitative performance, Frechet Inception Distance (FID) methodology is used. The FID value of 18 random samples, generated by networks constructed, is 38413677.145. For a qualitative measure of the model let the reader judge the quality of the image generated by the generator trained model. The Nepali letters were generated by the adversarial network as required. The evaluation helps the generative model to be better and further enables a better generation that humans have not thought of.