Generate High Fidelity Images With Generative Variational Autoencoder
share
In this work, we address the problem of blurred images which are often generated using Variational Autoencoders and the problem of mode collapse in Generative Adversarial Networks using a single model architecture. We use the encoder of VAE as it is while replacing the decoder with a discriminator. The encoder is fed data from a normal distribution while the generator is fed that from a gaussian distribution. The combination from both is then fed to a discriminator which tells whether the generated images are correct or not. We tested the model on 3 different datasets increasing in complexity MNIST, fashion MNIST and TCIA Pancreas CT dataset. On training the model for 300 iterations, it was able to generate much sharper images as compared to those of VAEs. This work is potentially very exciting as we are able to combine the advantages of generative models and inference models in a bayesian approach. As there is a shortage of medical data, this approach could be revolutionary given that we are able to reason from the bayesian approach taking into account the uncertainty of generation.
READ FULL TEXT