[Paper Review] Hierarchical Bayesian sparse image reconstruction with application to MRFM

Nicolas Dobigeon, Alfred O. Hero+1 more|Open Archive Toulouse Archive Ouverte (University of Toulouse)|Sep 22, 2008

Sparse and Compressive Sensing Techniques4 citations

TL;DR

This paper proposes a hierarchical Bayesian sparse image reconstruction method for MRFM imaging, using a mixture prior of a point mass at zero and a positive exponential distribution to enforce sparsity and positivity. By employing Gibbs sampling with fully Bayesian hyperparameter marginalization, the method provides full posterior inference and outperforms existing sparse reconstruction techniques, especially at high SNR, as validated on synthetic and real MRFM data of a tobacco virus.

ABSTRACT

This paper presents a hierarchical Bayesian model to reconstruct sparse images when the observations are obtained from linear transformations and corrupted by an additive white Gaussian noise. Our hierarchical Bayes model is well suited to such naturally sparse image applications as it seamlessly accounts for properties such as sparsity and positivity of the image via appropriate Bayes priors. We propose a prior that is based on a weighted mixture of a positive exponential distribution and a mass at zero. The prior has hyperparameters that are tuned automatically by marginalization over the hierarchical Bayesian model. To overcome the complexity of the posterior distribution, a Gibbs sampling strategy is proposed. The Gibbs samples can be used to estimate the image to be recovered, e.g. by maximizing the estimated posterior distribution. In our fully Bayesian approach the posteriors of all the parameters are available. Thus our algorithm provides more information than other previously proposed sparse reconstruction methods that only give a point estimate. The performance of our hierarchical Bayesian sparse reconstruction method is illustrated on synthetic and real data collected from a tobacco virus sample using a prototype MRFM instrument.

Motivation & Objective

• To address the challenge of reconstructing naturally sparse, positive images from linear, noisy observations in MRFM imaging.
• To overcome the instability of hyperparameter tuning in sparse Bayesian deconvolution methods.
• To develop a fully Bayesian framework that provides full posterior distributions for all parameters, not just point estimates.
• To enable automatic hyperparameter estimation through marginalization, avoiding reliance on penalized likelihood or empirical Bayes methods.

Proposed method

• Proposes a hierarchical Bayesian model with a prior that combines a point mass at zero and a positive exponential distribution to enforce sparsity and positivity in image pixels.
• Introduces conjugate priors on hyperparameters to enable full Bayesian inference and automatic hyperparameter marginalization.
• Employs Gibbs sampling to generate posterior samples, allowing estimation of the image and all unknown parameters via posterior modes or means.
• Uses an efficient recursive computation strategy to avoid recomputing the bilinear observation operator T(κ, x) at each Gibbs iteration, significantly improving computational efficiency.
• Applies a Bernoulli-truncated Gaussian distribution for pixel state sampling, with auxiliary variables to handle the mixture structure.
• Derives a simulation scheme that leverages the linear structure of the observation model to update T(κ, x) incrementally during Gibbs sampling.

Experimental results

Research questions

• RQ1 Can a hierarchical Bayesian model with automatic hyperparameter tuning improve sparse image reconstruction in MRFM compared to existing penalized likelihood or empirical Bayes methods?
• RQ2 How does the proposed mixture prior—combining a point mass at zero and a positive exponential distribution—enhance sparsity and positivity enforcement in image reconstruction?
• RQ3 To what extent does the full posterior inference from Gibbs sampling provide more robust and informative estimates than point estimates from traditional sparse deconvolution methods?
• RQ4 How does the proposed method perform in reconstructing real MRFM data of a tobacco virus, especially at high signal-to-noise ratios where previous methods fail?
• RQ5 Can the recursive computation of the observation operator T(κ, x) be effectively exploited to reduce computational cost in Gibbs sampling for large-scale image reconstruction?

Key findings

• The proposed hierarchical Bayesian method provides more stable and accurate image reconstructions than existing penalized likelihood approaches, particularly at high signal-to-noise ratios where previous methods exhibit instability.
• The automatic marginalization of hyperparameters via a second-level hierarchy eliminates the need for iterative optimization or tuning, improving robustness.
• Gibbs sampling generates full posterior distributions over all parameters, including the image, noise variance, and hyperparameters, enabling richer inference than point estimates.
• The recursive update strategy for the bilinear observation operator reduces computational cost, enabling efficient sampling even for large images.
• On real MRFM data of a tobacco virus, the method successfully recovers fine structural details with improved contrast and reduced noise compared to baseline methods.
• The use of a positive exponential prior combined with a point mass at zero better captures the natural sparsity and positivity of MRFM images than standard Laplacian or Gaussian mixture priors.