Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
133 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
46 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Compressive Imaging using Approximate Message Passing and a Markov-Tree Prior (1108.2632v1)

Published 12 Aug 2011 in cs.CV

Abstract: We propose a novel algorithm for compressive imaging that exploits both the sparsity and persistence across scales found in the 2D wavelet transform coefficients of natural images. Like other recent works, we model wavelet structure using a hidden Markov tree (HMT) but, unlike other works, ours is based on loopy belief propagation (LBP). For LBP, we adopt a recently proposed "turbo" message passing schedule that alternates between exploitation of HMT structure and exploitation of compressive-measurement structure. For the latter, we leverage Donoho, Maleki, and Montanari's recently proposed approximate message passing (AMP) algorithm. Experiments with a large image database suggest that, relative to existing schemes, our turbo LBP approach yields state-of-the-art reconstruction performance with substantial reduction in complexity.

Citations (213)

Summary

  • The paper introduces a novel compressive imaging technique combining Approximate Message Passing (AMP) with a Markov-Tree Prior (HMT) via "turbo" message passing.
  • The approach integrates a turbo message passing schedule exploiting HMT structure with AMP for efficient and high-quality image reconstruction.
  • Experimental results demonstrate state-of-the-art reconstruction performance and significantly reduced computational complexity compared to existing methods.

Overview of "Compressive Imaging using Approximate Message Passing and a Markov-Tree Prior"

The paper authored by Subhojit Som and Philip Schniter explores the application of advanced probabilistic algorithms for compressive imaging, which is the problem of estimating an image from fewer measurements than its pixel count. The research introduces an innovative approach that combines Approximate Message Passing (AMP) with a Markov-Tree Prior, specifically exploiting the statistical dependencies and sparsity present in wavelet-transformed images. This paper stands out by incorporating a "turbo" message passing schedule which iteratively exploits structure within both compressive measurement data and a Hidden Markov Tree (HMT) model to achieve enhanced reconstruction performance.

Key Methodological Innovations

The central methodological contribution of the paper is the innovative application of Loopy Belief Propagation (LBP) within the framework of a Hidden Markov Tree (HMT) model, which captures the persistence across scales (PAS) property of wavelet-transformed natural images. More specifically, the research builds on prior work by integrating LBP with a message passing strategy dubbed "turbo," which alternates between the leveraging of HMT and compressive measurement structures. This approach leads to optimized image reconstruction performance by iteratively refining the estimated states and coefficients of the wavelet transform.

Another critical component is the use of Approximate Message Passing (AMP), an efficient algorithm known for its low computational cost and ability to statistically leverage high-dimensional, sparse structures. The paper adapts AMP for the non-identical, non-independent wavelet coefficients typical in natural images, achieving robust image reconstruction with a substantial reduction in complexity compared to traditional methods.

Experimental Results and Concrete Outcomes

The paper reports experimental validation using a large image database, highlighting the superior performance of the proposed Turbo AMP method compared to existing state-of-the-art schemes. Numerical results demonstrate that the Turbo LBP approach not only enhances reconstruction quality but also significantly reduces computational complexity. These experiments establish the proposed method as providing "state-of-the-art reconstruction performance with substantial reduction in complexity," a claim supported by competitive numerical results in terms of normalized mean squared error (NMSE) and runtime analysis.

Implications and Future Prospects

From both a theoretical and practical viewpoint, the integration of AMP with a Markov-tree prior within a turbo LBP framework has several implications. Theoretically, it opens paths to investigate the convergence properties of LBP in high-dimensional structured spaces, particularly when coupled with AMP under non-traditional priors. Practically, the approach could be extended to various imaging modalities beyond simple compressive measurements, such as those involving redundancy in imperfections or noise patterns. Moreover, this paper might inspire further adaptation of similar message-passing and LBP techniques to other domains requiring efficienty inference in large-scale latent variable models. Future developments could also include the exploration of alternative structure-exploitation models or priors that further improve the tradeoff between computational efficiency and reconstruction accuracy in compressive imaging tasks.