TL;DR
This paper introduces a novel "Score-Based Turbo Message Passing" (STMP) algorithm for compressive image recovery. By integrating powerful score-based generative models into a message-passing framework, STMP achieves state-of-the-art image reconstruction quality with significantly fewer computational resources than existing methods.
What Problem Does This Paper Address?
Compressive imaging aims to reconstruct high-resolution images from a limited number of measurements, a task that is inherently ill-posed. Traditional methods often rely on hand-crafted image priors or generic denoisers, which struggle to capture the complex distributions of real-world images, especially in highly underdetermined scenarios (when the number of measurements is much smaller than the image dimensions). While score-based generative models excel at learning these complex image priors, directly applying them to inverse problems can be computationally prohibitive due to the need for many sampling steps. This paper bridges this gap by developing a message-passing algorithm that efficiently leverages the power of score-based models for compressive image recovery.
Methods
The core innovation is the development of the Score-Based Turbo Message Passing (STMP) algorithm. STMP builds upon the Turbo-CS framework, a message-passing algorithm that iteratively refines image estimates using two modules: a linear MMSE estimator and an MMSE denoiser. The novelty lies in replacing the traditional denoiser in Module B with a score-based MMSE denoiser.
This score-based denoiser is derived by exploiting the theoretical connection between score functions (gradients of log-probability densities) and MMSE estimation via Tweedie's formula. The paper proposes learning both the first-order score function (for the mean of the denoised estimate) and the second-order score function (for the variance) using a denoising score matching objective. This allows for per-iteration MMSE performance with a single call to the learned score network. To handle the computational cost of training, they employ a unified objective for training both first- and second-order score models. The algorithm also incorporates damping to stabilize convergence, particularly in low compression scenarios.
Datasets
- FFHQ dataset: Used for evaluating the performance of the proposed STMP method. The authors used images from the validation set for testing.
- No new benchmarks were introduced.
Key Insights and Ideas
- Bridging the gap between message passing and score-based models: The paper successfully integrates the fast convergence of message passing with the powerful prior modeling capabilities of score-based generative models, leading to improved reconstruction quality.
- Tweedie's formula as a theoretical foundation: The authors leverage Tweedie's formula to establish a direct link between score functions and MMSE denoising, enabling the efficient use of learned score models.
- Learning second-order score functions for variance estimation: Crucially, the paper highlights the importance of learning the second-order score function to accurately estimate the posterior variance, which is essential for effective message passing updates.
- Significant reduction in computational cost: By using score functions for per-iteration MMSE estimation, STMP achieves comparable or better results with substantially fewer neural function evaluations (NFEs) compared to other score-based sampling methods.
- Practical implications for underdetermined inverse problems: The method demonstrates superior performance in challenging scenarios with low compression ratios, where traditional methods often fail.
Key Results
- State-of-the-art performance: STMP achieves significantly better performance-complexity tradeoffs compared to conventional message passing, regularized linear regression, and score-based posterior sampling baselines on the FFHQ dataset.
- Fast convergence: The algorithm typically converges in fewer than 20 neural function evaluations (NFEs).
- Improved image quality: Qualitative results show STMP producing more faithful and higher-quality reconstructions, especially in challenging low-compression and noisy scenarios, outperforming methods like PnP-ADMM and D-Turbo-CS.
- Quantitative improvements: Table I shows STMP achieving higher PSNR and SSIM, and lower FID and LPIPS scores compared to baselines, particularly for the more challenging δ₀ = 0.05 case.
Limitations and Future Work
- Computational complexity of LMMSE estimator: The paper notes that the LMMSE estimator in Module A requires a matrix inverse at each iteration, which can be computationally prohibitive. While they mention avoiding this with SVD for general matrices and diagonal matrices for DCT-based sensing, it remains a point of consideration.
- Availability of second-order score models: The authors mention the lack of publicly available second-order score models and had to train one for their experiments. This suggests that developing more accessible and pre-trained second-order score models could further facilitate the adoption of such methods.
- Future work: The paper implicitly suggests that further exploration of score-based generative models within message-passing frameworks could lead to even more advanced recovery techniques. The success of STMP also opens avenues for applying this approach to other inverse problems beyond image recovery.
Summary updated