Variational Autoencoded Compositional Pattern Generative Adversarial Network for Handwritten Super Resolution Image Generation

Abstract

Since generative adversarial training has been decleared as one of the most exciting topics of the last 10 years by the pioneers, many researchers have focused on the Generative Adversarial Network (GAN) in their studies. On the otherhand, Variational Autoencoders (VAE) had gain autoencoders' popularity back. Due to some restrictions of GAN models and their lack of inference mechanism, hybrid models of GAN and VAE have emerged for image generation problem in nowadays. With the influence of these views and improvements, we have focused on addressing not only generating synthetic handwritten images but also their high-resolution version. For these tasks, Compositional Pattern Producing Networks (CPPN), VAE and GAN models are combined inspired by an existing model with some modification of its objective function. With this model, the idea behind the inspired study for generating high-resolution images are combined with the feature-wise reconstruction objective of a VAE/GAN hybrid model instead of pixel-like reconstruction approach of traditional VAE. For evaluating the model efficiency, our VAE/CPGAN model is compared with its basis models (GAN, VAE and VAE/GAN) and inspired model accoording to inception score. In this study, it is clearly seen that the proposed model is able to converge much faster than compared models for modeling the underlying distribution of handwritten image data.