Partially Linear Estimation with Application to Image Deblurring Using Blurred/Noisy Image Pairs

Tomer Michaeli, Daniel Sigalov, Yonina C. EldarLVA/ICA 2012 · DOI

TL;DR

This paper introduces a "Partially Linear Minimum Mean Squared Error" (PLMMSE) estimator that offers a computationally efficient way to estimate a signal from multiple noisy and/or blurred measurements. It achieves performance close to more complex methods for image deblurring and denoising, but at a significantly faster speed.

What Problem Does This Paper Address?

Estimating an unknown signal (like an image) from noisy or degraded measurements is a fundamental problem in many fields. The ideal solution often involves minimizing the Mean Squared Error (MSE), which typically leads to a conditional expectation estimator. However, calculating this "true" MMSE estimator can be computationally prohibitive, especially when dealing with complex signal models (like sparse representations) or multiple measurement sets. This paper tackles the challenge of finding a good estimator that is "good enough" in terms of performance but significantly faster to compute, particularly when we have multiple types of measurements (e.g., a noisy version and a blurred version of the same signal).

Methods

The core contribution is the Partially Linear MMSE (PLMMSE) estimator. Instead of assuming the entire estimator must be linear (like the Linear MMSE or LMMSE estimator), PLMMSE assumes the estimator is linear with respect to one set of measurements (say, Y) and can be a general function of the other set of measurements (say, Z). This is expressed as $\hat{X} = AY + b(Z)$, where A is a matrix and b(Z) is a function.

The paper derives the form of this PLMMSE estimator, showing it can be expressed as $\hat{X} = \hat{X}_{X|Z} + \Gamma_{XW|Z} \Gamma_{WW|Z}^{-1} W$, where $W$ is an "innovation" term related to the difference between the actual measurement Y and its expected value given Z. Crucially, this form requires only second-order statistics of X and Y, and the conditional expectation of X and Y given Z.

A key application is demonstrated for sparse signal recovery. When a signal X is sparse in a known dictionary (meaning it can be represented by a few non-zero coefficients), and we have noisy (Z) and blurred (Y) observations, the PLMMSE estimator offers a practical solution. The paper shows how to compute the necessary components, particularly when the signal is sparse in a wavelet basis.

Datasets

The paper uses the "Gold-hill" image for experimental evaluation. It does not introduce new datasets but rather uses a standard image to demonstrate the effectiveness of the proposed method.

Key Insights and Ideas

  • The power of partial linearity: By allowing linearity in only part of the measurements, the estimator strikes a balance between the optimality of MMSE and the computational tractability of LMMSE.
  • Bridging the gap: PLMMSE can significantly outperform estimators that only use one type of measurement (e.g., denoising alone or deblurring alone) when multiple measurements are available.
  • Efficiency for sparse signals: For signals with sparse representations (common in image processing), the PLMMSE estimator can be computed very efficiently, especially when combined with wavelet transforms.
  • Practical image enhancement: The method provides a fast and effective way to improve image quality from blurred and noisy pairs, outperforming many existing methods in speed.

Key Results

  • For image deblurring/denoising using blurred/noisy image pairs, the PLMMSE estimator achieved PSNR values comparable to or slightly better than state-of-the-art methods like BM3D.
  • Crucially, the PLMMSE method was significantly faster, running an order of magnitude faster than BM3D deblurring and denoising, and also faster than the RD (Residual Deconvolution) method. For example, it was 11 times faster than BM3D deblurring and 18 times faster than RD on average.
  • Qualitative results showed that while the individual denoised or deblurred components might be weaker than specialized methods, the joint PLMMSE estimate provided a superior overall recovery.

Limitations and Future Work

  • Approximation of terms: While the PLMMSE estimator is derived, some of its components (like the scalar shrinkage functions in the sparse case) might require numerical computation, especially when parameters vary.
  • Assumptions for speed: The significant speedup in the image deblurring application relies on simplifying assumptions, such as all wavelet coefficients being independent and identically distributed, and the blur kernel being known.
  • Future work: The paper suggests that the PLMMSE framework is broadly applicable to various scenarios where multiple measurement types are available and computational efficiency is a concern. Further exploration of its application in other signal processing domains is implied.

Summary updated

BibTeX
@inproceedings{michaeli2012partially,
  author = {Tomer Michaeli and Daniel Sigalov and Yonina C. Eldar},
  title = {{Partially Linear Estimation with Application to Image Deblurring Using Blurred/Noisy Image Pairs}},
  booktitle = {LVA/ICA},
  year = {2012},
  doi = {10.1007/978-3-642-28551-6\_2},
}
Figure