Hierarchical Quantized Autoencoders

TL;DR

This paper introduces a hierarchical VQ-VAE framework that enhances lossy image compression by maintaining perceptual quality and semantic features at very low bitrates through a novel training objective.

Contribution

It proposes a hierarchical structure of VQ-VAEs with stochastic quantization and a new training objective to improve high-factor compression and image quality.

Findings

Produces high-perceptual quality images with semantic features

Achieves better compression efficiency on CelebA and MNIST datasets

Introduces a novel hierarchical VQ-VAE training scheme

Abstract

Despite progress in training neural networks for lossy image compression, current approaches fail to maintain both perceptual quality and abstract features at very low bitrates. Encouraged by recent success in learning discrete representations with Vector Quantized Variational Autoencoders (VQ-VAEs), we motivate the use of a hierarchy of VQ-VAEs to attain high factors of compression. We show that the combination of stochastic quantization and hierarchical latent structure aids likelihood-based image compression. This leads us to introduce a novel objective for training hierarchical VQ-VAEs. Our resulting scheme produces a Markovian series of latent variables that reconstruct images of high-perceptual quality which retain semantically meaningful features. We provide qualitative and quantitative evaluations on the CelebA and MNIST datasets.