HANDWRITTEN WORD GENERATION USING GENERATIVE ADVERSARIAL NETWORKS

Abstract

We propose a framework for generating handwritten samples using Generative Adversarial Networks. This is a slightly different version of the Generative Adversarial Networks in which two networks are pitted against each other in an adversarial format to maximize the “accuracy” of the generated content. Instead of that, we introduce a third network that controls the output of the generator and enforces legibility. We train this architecture on the famous IAM dataset, in order to replicate handwritten words. The generated words bear some resemblance to the words in the dataset. Although not perfect, we plan to shed some light on the topic using this architecture and further improve the model using advanced training methods and new techniques to help the model converge faster.