[Paper Review] Detection of Markov Random Fields on Two-Dimensional Intersymbol Interference Channels

Ying Zhu, Taikun Cheng+2 more|ArXiv.org|Sep 27, 2006

Cellular Automata and Applications12 references3 citations

TL;DR

This paper proposes a novel iterative detection algorithm that combines a soft-input/soft-output 2D intersymbol interference (ISI) detector with a Markov random field (MRF) detector to exploit image correlation in 2D optical and magnetic storage systems. By leveraging a first-order binary MRF model and iterative extrinsic information exchange, the method achieves 0.5–2.0 dB SNR gain over ISI-only detection, with robustness to MRF parameter mismatch and strong performance on natural binary images.

ABSTRACT

We present a novel iterative algorithm for detection of binary Markov random fields (MRFs) corrupted by two-dimensional (2D) intersymbol interference (ISI) and additive white Gaussian noise (AWGN). We assume a first-order binary MRF as a simple model for correlated images. We assume a 2D digital storage channel, where the MRF is interleaved before being written and then read by a 2D transducer; such channels occur in recently proposed optical disk storage systems. The detection algorithm is a concatenation of two soft-input/soft-output (SISO) detectors: an iterative row-column soft-decision feedback (IRCSDF) ISI detector, and a MRF detector. The MRF detector is a SISO version of the stochastic relaxation algorithm by Geman and Geman in IEEE Trans. Pattern Anal. and Mach. Intell., Nov. 1984. On the 2 x 2 averaging-mask ISI channel, at a bit error rate (BER) of 10^{-5}, the concatenated algorithm achieves SNR savings of between 0.5 and 2.0 dB over the IRCSDF detector alone; the savings increase as the MRFs become more correlated, or as the SNR decreases. The algorithm is also fairly robust to mismatches between the assumed and actual MRF parameters.

Motivation & Objective

• To address the performance limitations of existing 2D ISI detectors that assume i.i.d. image inputs, which do not reflect real-world image correlation.
• To exploit the inherent spatial correlation in natural and uncompressed images by modeling them as first-order binary Markov random fields (MRFs).
• To develop a soft-input/soft-output (SISO) MRF detector that can be concatenated with an existing 2D ISI detector to improve detection performance.
• To evaluate the robustness of the concatenated detector to mismatches between assumed and actual MRF parameters.
• To demonstrate the practical utility of the MRF model on real natural binary images such as 'chair' and 'man'.

Proposed method

• The system uses a concatenated iterative detector based on the 'turbo principle,' combining an iterative row-column soft-decision feedback (IRCSDF) ISI detector with a SISO MRF detector.
• The ISI detector employs the BCJR algorithm on rows and columns with soft decision feedback, using weighted LLR estimates to mitigate intersymbol interference.
• The MRF detector is a SISO version of the stochastic relaxation algorithm by Geman and Geman, used to estimate the original MRF from noisy, corrupted observations.
• Extrinsic log-likelihood ratios (LLRs) are exchanged iteratively between the ISI and MRF detectors to refine estimates and improve detection accuracy.
• The input image is interleaved and level-shifted from {0,1} to {-1,1} before transmission to ensure i.i.d. pixel assumptions in the ISI detector.
• The MRF is modeled using the Ising model with a two-parameter energy function, and the detector assumes knowledge of the MRF parameters (including β).

Experimental results

Research questions

• RQ1 Can exploiting image correlation via a first-order MRF model lead to significant SNR gains in 2D ISI detection?
• RQ2 How does the performance of the concatenated MRF-ISI detector vary with the degree of image correlation (i.e., MRF parameter β)?
• RQ3 How robust is the detector to mismatches between the assumed and actual MRF parameters?
• RQ4 Can a simple first-order MRF model effectively model real natural binary images for 2D ISI mitigation?
• RQ5 What is the performance gain of the proposed method compared to ISI-only detection on natural images?

Key findings

• On the 2×2 averaging-mask ISI channel, the concatenated detector achieves 0.5–2.0 dB SNR gain over the IRCSDF detector alone at a BER of 10⁻⁵, with gains increasing as image correlation increases.
• The algorithm is robust to MRF parameter mismatch: at high SNR, mismatched β values (e.g., β = -0.5, -0.75, -3.0) incur only 0.2–0.3 dB penalty compared to correct β, while still outperforming ISI-only detection.
• At low SNR, even large mismatches (e.g., β = -4.5 when true β = -1.5) yield performance nearly identical to the correct model, indicating strong robustness.
• For highly correlated MRFs (e.g., β = -3.0), a wide range of assumed β values (from -1.5 to -10.0) yield gains exceeding 1 dB over ISI-only detection.
• On natural binary images 'chair' and 'man', the concatenated detector achieves 1–2 dB SNR savings over ISI-only detection across a broad range of β values, demonstrating practical utility.
• The results suggest that a first-order MRF model is a highly effective and robust approximation for 2D-ISI mitigation in real-world images, even without precise model estimation.