Compression for Multiple Reconstructions

Yehuda Dar, Michael Elad, Alfred M. Bruckstein

Abstract

In this work, we propose a method for optimizing lossy compression for a network of diverse reconstruction systems. We focus on adapting a standard image compression method to a set of candidate displays, presenting the decompressed signals to viewers. Each display is modeled as a linear operator applied after decompression, associated with a specific probability of serving a network user. We formulate a complex operational rate-distortion optimization that trades off the network's expected mean-squared reconstruction error against the compression bit-cost. Using the alternating direction method of multipliers (ADMM), we develop an iterative procedure where the network structure is separated from the compression method, enabling the reliance on standard compression techniques like HEVC. Experimental results demonstrate that our method is the leading approach for adjusting high bit-rate image compression to a set of displays modeled as blur degradations.

Introduction

In modern multimedia distribution, a single image or video stream is broadcast to a wide variety of consumer devices, ranging from high-end 4K televisions to budget smartphones. Each device possesses unique display technologies, resolutions, and rendering pipelines that inevitably alter the final viewed image. Traditional lossy compression optimizes solely for the fidelity of the decompressed signal, ignoring the subsequent degradations introduced during rendering.

This research addresses the challenge of optimizing a single, non-layered compressed bitstream to ensure maximum quality across a known set of diverse displays. By factoring in the probability of a user employing a specific display type, we can optimize the source bitstream for the entire network's expected viewing experience.

Methodology

Problem Formulation

The core of our approach is a rate-distortion optimization framework that minimizes the expected mean-squared error (MSE) across all possible displays in a network. We model each display mathematically as a linear operator (a matrix applied post-decompression) with an associated probability of use. The goal is to find a compressed representation that balances the total bit-cost with the weighted average of reconstruction errors across these diverse operators.

Optimization via ADMM

Solving this global optimization directly is computationally prohibitive due to the coupling of the compression algorithm and the network-specific degradations. We overcome this by utilizing the Alternating Direction Method of Multipliers (ADMM). This mathematical framework allows us to decouple the optimization into two alternating, iterative steps:

  1. Standard Compression/Decompression: A network-independent codec (such as the state-of-the-art HEVC standard) is applied to an intermediate variable. This allows the codec to function as a "black box."
  2. Network-Aware Deconvolution: An analytic, $\ell_2$-constrained deconvolution step accounts for the specific display operators and their usage probabilities.

This decoupling is a significant breakthrough; it allows the use of standard, industry-compliant codecs without requiring modification to the underlying compression math or the end-user's decoding hardware.

Experimental Evaluation

Datasets and Setup

The proposed method was evaluated using 12 standard test images from three prominent databases:

  • TESTIMAGES: Almonds, Cards, Duck toys, Garden table
  • UCID: House & lawn, Tree, Garden, Teddy bear
  • Berkeley: Bears, Boats, Butterfly, Flower & Bugs

We simulated a network featuring three display types, modeled as Gaussian blurs with varying intensities (standard deviations of 0.6, 0.8, and 1.0) and usage probabilities of 60%, 30%, and 10%, respectively.

Key Insights

  • Pre-compensation: The encoder effectively "pre-compensates" for expected degradations. For instance, the encoder may intentionally over-sharpen an image so that the display's natural blur restores it to an optimal state.
  • Standard Decoding: Because the output is a standard binary compressed file, the end-user decodes the HEVC file normally. The optimization happens entirely on the server side during the encoding phase.

Results

Our experimental results show that the proposed method is highly effective for high bit-rate scenarios:

  • Quantitative Gains: The method achieved average Peak Signal-to-Noise Ratio (PSNR) gains ranging from 3.26 dB to 5.25 dB over standard HEVC compression.
  • Baseline Comparison: Our approach outperformed advanced baselines, such as Expected Patch Log Likelihood (EPLL) deblurring combined with HEVC, by 0.47 dB to 1.44 dB on average.
  • Qualitative Impact: While the intermediate decompressed images appear unnaturally sharp, the final viewed images—after passing through simulated display blurs—retain significantly more detail and clarity than those processed via standard methods.

Conclusion and Future Work

This work introduces a generic framework for network-aware compression that leverages existing standards through ADMM-based decoupling. While the iterative nature of the encoding process introduces computational overhead (requiring up to 40 iterations), the gains in visual quality for the end-user are substantial, particularly at high bit-rates.

Future research will focus on expanding this framework to non-linear display models and exploring its application in tasks where decompressed data is used for automated computation rather than human viewing. Additionally, optimizing the computational efficiency of the ADMM loop remains a priority for real-time applications.

Summary updated

Abstract

In this work we propose a method for optimizing the lossy compression for a network of diverse reconstruction systems. We focus on adapting a standard image compression method to a set of candidate displays, presenting the decompressed signals to viewers. Each display is modeled as a linear operator applied after decompression, and its probability to serve a network user. We formulate a complicated operational rate-distortion optimization trading-off the network's expected mean-squared reconstruction error and the compression bit-cost. Using the alternating direction method of multipliers (ADMM) we develop an iterative procedure where the network structure is separated from the compression method, enabling the reliance on standard compression techniques. We present experimental results showing our method to be the best approach for adjusting high bit-rate image compression (using the state-of-the-art HEVC standard) to a set of displays modeled as blur degradations.
BibTeX
@inproceedings{dar2018compression,
  author = {Yehuda Dar and Michael Elad and Alfred M. Bruckstein},
  title = {{Compression for Multiple Reconstructions}},
  booktitle = {ICIP},
  year = {2018},
  doi = {10.1109/icip.2018.8451743},
}
Figure