An Improved Upper Bound on the Rate-Distortion Function of Images

TL;DR

This paper presents an improved upper bound on the image rate-distortion function using a new VAE architecture, variable-rate techniques, and a stabilization method, demonstrating significant potential for enhancing lossy image compression.

Contribution

Introduces a novel VAE model and training stabilization method to better estimate the image rate-distortion function, enabling more effective lossy compression.

Findings

Achieves at least 30% BD-rate reduction compared to VVC intra prediction.

Demonstrates the effectiveness of the new VAE architecture and training method.

Provides publicly available code for reproducibility.

Abstract

Recent work has shown that Variational Autoencoders (VAEs) can be used to upper-bound the information rate-distortion (R-D) function of images, i.e., the fundamental limit of lossy image compression. In this paper, we report an improved upper bound on the R-D function of images implemented by (1) introducing a new VAE model architecture, (2) applying variable-rate compression techniques, and (3) proposing a novel \ourfunction{} to stabilize training. We demonstrate that at least 30\% BD-rate reduction w.r.t. the intra prediction mode in VVC codec is achievable, suggesting that there is still great potential for improving lossy image compression. Code is made publicly available at https://github.com/duanzhiihao/lossy-vae.