Auto-encodeur optimisé au sens débit-distorsion : indépendant de la quantification?

Thierry Dumas, Aline Roumy, Christine GuillemotHAL 2017

TL;DR

This paper explores how to make image compression using deep autoencoders more flexible by decoupling the learning of the image transform from the quantization step. They propose methods to either learn quantization parameters jointly with the transform or to adapt quantization during testing, achieving competitive compression performance with fewer learned models.

What Problem Does This Paper Address?

Traditional image and video compression standards rely on fixed, linear transforms like the Discrete Cosine Transform (DCT). However, deep learning, specifically autoencoders, has shown promise in learning more powerful, non-linear transforms that can capture image content more effectively. A key challenge in autoencoder-based compression is the quantization step, which discretizes the learned representation before encoding. Current state-of-the-art autoencoders often fix the quantization parameters during training, meaning a separate autoencoder must be trained for each desired compression rate (or bitrate). This is inefficient and limits flexibility. The authors question whether this fixed quantization is optimal and explore ways to make the autoencoder's compression independent of a specific, learned quantization step.

Methods

The paper proposes two main approaches:

  1. Joint Learning of Transform and Quantization: Instead of fixing quantization, the authors propose to learn the quantization parameters ($\delta_i$) alongside the autoencoder's weights ($\theta, \phi$). This is achieved by approximating the quantization step with a differentiable function during training, allowing for gradient-based optimization. The objective function is modified to minimize both reconstruction error and the entropy of the quantized representation, while also learning the quantization step sizes.
  2. Quantization Adaptation at Test Time: This approach focuses on learning a single autoencoder with a fixed quantization step (or one that is not explicitly learned for different rates). However, during testing, the paper investigates whether different quantization steps can be applied to the learned representation without retraining. They analyze the distribution of the learned representations and find they often follow Laplace distributions. This observation suggests that a single learned transform might be compatible with various quantization levels at inference time.

Datasets

  • Training: 24,000 luminance images of size 256x256, created from ImageNet.
  • Testing: 24 luminance images of size 512x768 from the Kodak dataset.

Key Insights and Ideas

  • Quantization is Crucial: The paper highlights that quantization, a fundamental step in all coding standards, is surprisingly often fixed in deep learning compression models.
  • Learned Transforms Adapt to Quantization: The authors demonstrate that autoencoders trained with an entropy constraint learn representations that naturally adapt to different quantization levels, often exhibiting Laplace-like distributions.
  • Test-Time Quantization Flexibility: The most significant insight is that a single autoencoder, once trained, can be effectively used with various quantization steps during testing, achieving performance comparable to models trained for specific bitrates. This drastically reduces the need for multiple trained models.
  • Joint Learning vs. Test-Time Adaptation: While joint learning is explored, the results suggest that adapting quantization at test time is more promising for achieving bitrate flexibility with a single model.

Key Results

  • The proposed method of adapting quantization at test time achieves compression performance (rate-distortion curves) comparable to the state-of-the-art approach that requires training a separate autoencoder for each bitrate.
  • This test-time adaptation approach requires only one autoencoder to be trained, significantly reducing the computational cost and complexity compared to training multiple models.
  • The paper shows that autoencoders trained with an entropy constraint learn representations that are well-approximated by Laplace distributions, which is key to their adaptability to different quantization levels.

Limitations and Future Work

  • Joint Learning Efficiency: The paper suggests that the joint learning of the autoencoder and uniform scalar quantization parameters might be "little interesting" in terms of performance gains, though it does address the first research question.
  • Generalization to Non-Uniform Quantization: The authors note that the current approach is focused on uniform scalar quantization and suggest that generalizing this to non-uniform quantizations would be a valuable future direction.
  • Training Data Bias: The paper acknowledges that the training dataset is derived from JPEG-encoded images and raises the question of whether this might bias the learned autoencoder. Analyzing this potential bias is a direction for future work.

Summary updated

Abstract

National audience
BibTeX
@inproceedings{dumas2017auto,
  author = {Thierry Dumas and Aline Roumy and Christine Guillemot},
  title = {{Auto-encodeur optimisé au sens débit-distorsion : indépendant de la quantification?}},
  booktitle = {HAL},
  year = {2017},
}
Figure