Learned transform compression with optimized entropy encoding

Magda Gregorova, Marc Desaules, Alexandros Kalousis

Keywords: lossy compression, learned transform coding, vector quantization, entropy minimization

TL;DR: We learn a transform coding compression model with relaxation of the the vector-quantization to enable backpropagation and hence full optimization of the rate but also the entropy term in a novel way.

Abstract:

We consider the problem of learned transform compression where we learn both, the transform as well as the probability distribution over the discrete codes.
We utilize a soft relaxation of the quantization operation to allow for back-propagation of gradients and employ vector (rather than scalar) quantization of the latent codes.
Furthermore, we apply similar relaxation in the code probability assignments enabling direct optimization of the code entropy.
To the best of our knowledge, this approach is completely novel.
We conduct a set of proof-of concept experiments confirming the potency of our approaches.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2104.03305/code)