Non-uniform source modeling for distributed video coding

V. Toto-Zarasoa, Aline Roumy, Christine Guillemot2010

TL;DR

This paper tackles the challenge of efficiently compressing video data using Distributed Source Coding (DSC) by introducing a more accurate model for the correlation between video bit planes. They propose a "predictive" channel model that better captures the non-uniform nature of these bit planes, leading to improved compression rates and a more robust decoder.

What Problem Does This Paper Address?

Distributed Source Coding (DSC) is a powerful technique for compressing correlated data sources separately and decoding them jointly. In the context of Distributed Video Coding (DVC), this means encoding intermediate video frames (Wyner-Ziv frames) independently and then decoding them using information from previously decoded frames (side information). Traditional DSC methods often assume that the binary data representing video bit planes are uniformly distributed and that the correlation between the source and side information can be modeled by a simple "additive" Binary Symmetric Channel (BSC). However, real-world video bit planes are often non-uniform, and their correlation might not strictly follow the additive model. This mismatch can lead to suboptimal compression performance. This paper aims to bridge this gap by developing a more accurate correlation model and a decoder that can leverage the non-uniformity of video bit planes for better DVC.

Methods

The core of this paper lies in two main contributions:

  1. Novel Correlation Model: Predictive BSC: The authors introduce a "predictive" Binary Symmetric Channel (BSC) model. In this model, the source (X) is related to the side information (Y) by X = Y + Z, where Z is a noise term. This contrasts with the traditional "additive" BSC where Y = X + Z. The predictive model is motivated by the observation that in DVC, the side information often represents a prediction of the current frame, and the noise is more likely to be independent of the side information itself rather than the source. They show that while these models are equivalent for uniform sources, the predictive model offers advantages for non-uniform sources.
  2. Adaptive LDPC Decoder: They propose a modified Low-Density Parity-Check (LDPC) decoder. This decoder is designed to:
  3. Handle Non-Uniform Sources: It iteratively estimates the probability of a '1' bit (px) for each bit plane, rather than assuming uniformity.
  4. Adapt to Correlation Models: It can operate with either the additive or the predictive BSC model, and crucially, it includes a criterion to dynamically switch between these models during decoding. This is achieved by attempting decoding with the predictive model first and, if it fails, resorting to the additive model.

Datasets

The paper evaluates its proposed methods using synthetic binary sources and real video sequences. The video sequences used for evaluation are:

  • Hall monitor
  • Foreman
  • Soccer

These sequences are used to generate Wyner-Ziv (WZ) frames and key frames (KF). The WZ frames are then processed bit plane by bit plane, and the non-uniformity of these bit planes is analyzed.

Key Insights and Ideas

  • Non-uniformity Matters: The paper strongly emphasizes that real video bit planes are not uniformly distributed. Exploiting this non-uniformity is crucial for achieving better compression.
  • Predictive Model Advantage: For non-uniform sources, the predictive BSC model can offer a significant advantage over the additive BSC, especially when the side information is a good predictor.
  • Model Mismatch is Costly: A mismatch between the true correlation model (predictive or additive) and the model assumed by the decoder leads to a significant loss in performance, particularly for non-uniform sources.
  • Adaptive Decoding is Key: A decoder that can dynamically assess and adapt to both the non-uniformity of the source and the type of correlation channel (predictive vs. additive) is essential for optimal DVC performance.

Key Results

  • Rate Gain: The proposed adaptive decoder, which exploits non-uniformity and switches between correlation models, achieves noticeable rate gains compared to the standard DISCOVER codec. For the Soccer sequence, a rate reduction of up to 5.7% was observed.
  • Parameter Estimation Accuracy: The decoder's ability to estimate the non-uniformity parameter (px) was shown to be accurate when the decoding was successful, with a loss of less than 0.02 bits per symbol at low Bit Error Rates (BER).
  • Model Mismatch Penalty: Figures 1 and 2 clearly illustrate that when the decoder assumes the wrong correlation model (e.g., additive when it's predictive), the BER performance degrades significantly.

Limitations and Future Work

  • Decoder Complexity: While the proposed LDPC decoder is designed for efficiency, the iterative estimation of the non-uniformity parameter and the model switching mechanism likely introduce some computational overhead compared to a standard decoder. The paper mentions that 100 iterations is a good compromise between performance and complexity.
  • Channel Model Assumption: The paper primarily focuses on BSC models. While they mention that the correlation between WZ and SI bit planes can be approximated by a Laplacian channel, the core decoder modifications are presented within the BSC framework. Future work could explore more sophisticated channel models.
  • Rate Gain Variability: The observed rate gains vary across different video sequences and quality levels. While significant for Soccer, it was less pronounced for Hall monitor. Further investigation into the factors influencing this variability could be beneficial.
  • Future Work: The authors suggest that the proposed approach of exploiting non-uniformity and adapting to correlation models can be further improved. They also hint at exploring more complex channel models beyond BSC.

Summary updated

Abstract

We introduce a novel correlation model, called predictive Binary Symmetric Channel (BSC), for Distributed Source Coding (DSC). We then consider non-uniform binary sources and show that, for the predictive BSC model, the nonuniformity does not reduce the compression rate in asymmetric DSC. We assess the minimum achievable rate loss induced by a mismatch between the assumed and the true correlation models. We also propose an LDPC-based decoder adapted to both the classical (additive) BSC and the new (predictive) BSC correlation models. Finally, for Distributed Video Coding (DVC) application, we propose a criterion that allows us to switch between the two correlation models. 1.
BibTeX
@misc{totozarasoa2010non,
  author = {V. Toto-Zarasoa and Aline Roumy and Christine Guillemot},
  title = {{Non-uniform source modeling for distributed video coding}},
  year = {2010},
}

Loading related content...

Figure