STanH : Parametric Quantization for Variable Rate Learned Image Compression

TL;DR

A differentiable quantizer designed around a parametric sum of hyperbolic tangents, called STanH, that relaxes the step-wise quantization function is proposed that enables variable rate coding with comparable efficiency to the state-of-the-art, yet with significant savings in terms of ease of deployment, training time, and storage costs.

Abstract

In end-to-end learned image compression, encoder and decoder are jointly trained to minimize a $R + λD$ cost function, where $λ$ controls the trade-off between rate of the quantized latent representation and image quality. Unfortunately, a distinct encoder-decoder pair with millions of parameters must be trained for each $λ$, hence the need to switch encoders and to store multiple encoders and decoders on the user device for every target rate. This paper proposes to exploit a differentiable quantizer designed around a parametric sum of hyperbolic tangents, called STanH , that relaxes the step-wise quantization function. STanH is implemented as a differentiable activation layer with learnable quantization parameters that can be plugged into a pre-trained fixed rate model and refined to achieve different target bitrates. Experimental results show that our method enables variable rate coding with comparable efficiency to the state-of-the-art, yet with significant savings in terms of ease of deployment, training time, and storage costs

Figures (8)

• Figure 1: The reference learned image compression architecture Zou22zou (CNN-based architecture) with two STanH layers for quantizing the main latent space $\textbf{y}$ and the hyperprior latent space $\textbf{z}$.
• Figure 2: STanH activation function with $L=5$ quantization levels and for increasing values of inverse temperature $\beta$.
• Figure 3: Rate-distortion performance of Zou22 on Kodak dataset using different number of anchors, from six (a) to one (f). For our proposed approach, red crosses represent trained anchors, whereas red circles the refined derivation(s).
• Figure 4: Rate-PSNR plots for the proposed STanH -based method and relative reference for Kodak (top row), Clic (central row), and Tecnik (bottom row) datasets and for 3 anchors and 3 derivations.
• Figure 5: MS-SSIM for the proposed STanH -based method on Zou22, for 3 anchors and 3 derivations.