TL;DR
This paper introduces CLEAR, a novel framework for "re-fitting" the outputs of image restoration algorithms to reduce systematic errors (bias) introduced by regularization techniques like Total Variation. By preserving the Jacobian of the original estimator, CLEAR offers a more robust and accurate way to refine images compared to existing methods, particularly for ill-posed problems.
What Problem Does This Paper Address?
Many powerful image restoration algorithms, especially those using regularization techniques like $\ell_1$ norm or Total Variation (TV), introduce a systematic bias. This bias arises from the mathematical approximations made to solve complex, ill-posed inverse problems. For instance, $\ell_1$ regularization tends to shrink coefficients, leading to a loss of contrast in images, while TV regularization can smooth out important details. While post-hoc "re-fitting" methods exist to correct this bias, they often struggle with preserving the desired properties of the original solution or require precise identification of features like non-zero coefficients or jump locations, which can be difficult in practice. This paper addresses the challenge of developing a general, robust re-fitting strategy that effectively reduces bias without sacrificing the structural integrity and desired properties of the original restored image.
Methods
The core contribution is the Covariant LEAst-square Re-fitting (CLEAR) framework. Unlike previous "invariant" re-fitting methods that focus on preserving the "model subspace" (essentially, the linear behavior of the estimator), CLEAR emphasizes preserving the Jacobian of the original estimator. The Jacobian captures how the output of the estimator changes with respect to small perturbations in the input signal.
CLEAR is formulated as a two-step process:
- Guess-based Re-fitting: An intermediate estimator, $D\hat{i}_{z}$, is defined. This estimator is designed to be close to a "guess" $z$ (which is later set to the observed data $y$) and preserves specific properties of the original estimator $\hat{i}$:
- Affine Map: It's locally affine.
- Covariant Preserving: It preserves a scaled version of the original estimator's Jacobian.
- Coherent Map: If the input is the output of the original estimator at $z$, the re-fitting should not change it.
- Covariant Re-fitting (CLEAR): This is achieved by setting the guess $z$ to the observed data $y$. The resulting estimator, $R_{\hat{i}}(y)$, adds a local affine transformation of the residual term ($y - \Phi\hat{i}(y)$) to the original estimate $\hat{i}(y)$. The scaling factor for this affine transformation, $p$, is derived to ensure optimal bias reduction while maintaining the Jacobian's structure.
The paper also details how to compute the Jacobian efficiently using algorithmic differentiation or finite differences, making CLEAR applicable to a wide range of existing algorithms.
Datasets
Not discussed in the paper. The paper focuses on theoretical development and numerical simulations using synthetic images and standard image processing tasks.
Key Insights and Ideas
- Jacobian Preservation is Key: The paper argues that preserving the Jacobian of the original estimator is crucial for a robust re-fitting process. The Jacobian captures essential "covariant" information about how the estimator's output relates to input variations, which is lost in simpler "invariant" re-fitting methods.
- Beyond Model Subspace: CLEAR moves beyond preserving just the "model subspace" (which captures linear invariances) to also account for the estimator's first-order behavior, including its singularities and smoothness.
- "Twicing" Flavor: The re-fitting adds a term that is a scaled version of the directional derivative of the estimator along the residual. This has a "twicing" or "boosting" flavor, re-injecting information from the residual but in a controlled, covariant manner.
- General Applicability: The framework is designed to be general and can be applied to a broad family of estimators, including Lasso, Total Variation (aniso-TV, iso-TV), Tikhonov regularization, and non-local means.
- Bias Reduction without Support Identification: A significant advantage is that CLEAR does not require explicit identification of features like non-zero coefficients (for Lasso) or jump locations (for TV), which are often unstable or difficult to obtain accurately.
Key Results
- Improved Image Quality: Numerical experiments demonstrate that CLEAR effectively reduces the bias introduced by regularization techniques, leading to sharper images with better contrast and detail recovery, especially in cases where the original estimator's model is well-suited to the data.
- Preservation of Structure: CLEAR successfully refines the amplitudes of features without altering their locations or introducing spurious structures, unlike some other re-fitting or boosting methods.
- Efficiency: The computational overhead for CLEAR is at most twice that of the original algorithm when using algorithmic differentiation, and can be as low as twice the complexity of a single iteration for certain estimators.
- Robustness: CLEAR is shown to be more stable than methods relying on explicit support identification, which can be prone to errors.
- Bias-Variance Trade-off: While CLEAR generally reduces bias, it can increase variance. The paper shows that the re-fitting is most beneficial when the original estimator is well-chosen for the underlying image structure.
Limitations and Future Work
- MSE Not Always Improved: The paper acknowledges that while bias is reduced, the increase in residual variance means that Mean Squared Error (MSE) is not always improved, especially when the original estimator is already nearly unbiased or when the underlying image structure is not well-captured by the model.
- Parameter Tuning: While CLEAR itself has a parameter $p$ that is derived, the original estimators still require their own parameters (e.g., $\lambda$ for TV). The optimal choice of these parameters can still influence the effectiveness of CLEAR.
- Future Work: The authors suggest exploring the application of Jacobian-based re-fitting to boosting strategies and investigating its potential for other types of problems beyond image restoration. They also note that characterizing the specific class of images for which re-fitting is most beneficial remains an open challenge.
Summary updated