PaperlayerView exact revision on arXiv

Source-linked AI summary

MuLoG, or How to apply Gaussian denoisers to multi-channel SAR speckle reduction?

Charles-Alban Deledalle, Loïc Denis, Sonia Tabti, Florence Tupin

arXiv:1704.05335v1math.STstat.AP

TL;DR

Multi-channel SAR speckle corrupts polarimetric and interferometric information, while most denoisers target additive Gaussian noise. The paper proposes MuLoG, which combines a matrix logarithm with user-selected Gaussian denoisers; it extends available speckle-reduction methods and supports comparison of their artifacts.

  • Problem

    Multi-channel SAR speckle makes physical parameters highly variable, while a generic method for applying additive-Gaussian denoisers to such data is lacking.

  • Method

    MuLoG uses a matrix logarithm to approximately stabilize speckle variance and produce nearly independent channels before applying a user-defined Gaussian denoiser.

  • Results

    MuLoG produces a family of speckle-reduction algorithms applicable to single- and multi-channel SAR images, with user-selectable Gaussian denoisers and no parameter tuning beyond the denoiser.

  • Takeaways & Limitations

    Different Gaussian denoisers yield multiple reduced-speckle images whose method-specific artifacts can be compared, supporting artifact rejection and future benchmarking.

Abstract

from arXiv · show

Speckle reduction is a longstanding topic in synthetic aperture radar (SAR) imaging. Since most current and planned SAR imaging satellites operate in polarimetric, interferometric or tomographic modes, SAR images are multi-channel and speckle reduction techniques must jointly process all channels to recover polarimetric and interferometric information. The distinctive nature of SAR signal (complex-valued, corrupted by multiplicative fluctuations) calls for the development of specialized methods for speckle reduction. Image denoising is a very active topic in image processing with a wide variety of approaches and many denoising algorithms available, almost always designed for additive Gaussian noise suppression. This paper proposes a general scheme, called MuLoG (MUlti-channel LOgarithm with Gaussian denoising), to include such Gaussian denoisers within a multi-channel SAR speckle reduction technique. A new family of speckle reduction algorithms can thus be obtained, benefiting from the ongoing progress in Gaussian denoising, and offering several speckle reduction results often displaying method-specific artifacts that can be dismissed by comparison between results.

I. INTRODUCTION

SAR speckle creates severe fluctuations that make direct estimation of reflectivity and multi-channel physical parameters unreliable. MuLoG addresses the lack of a generic way to apply Gaussian denoisers to multi-channel SAR data.

  • Speckle fluctuations make direct estimation of reflectivity, interferometric phase, and polarimetric properties unusable because of their high variance.
  • Multi-channel SAR data provide interferometric, polarimetric, and tomographic information for elevation, displacement, biomass, and related Earth-observation tasks.
  • Many existing speckle-reduction methods do not generalize well to multi-channel images because similarity criteria and estimators must handle multivariate data.
  • MuLoG provides a generic method for embedding Gaussian denoisers into speckle-reduction filters for both single-channel and multi-channel SAR images.
  • The method targets multiplicative, signal-dependent speckle by using a logarithm that converts it into approximately additive, signal-independent fluctuations requiring bias correction.

B. Homomorphic approach

The homomorphic approach log-transforms intensity data and applies Gaussian denoising, but its Gaussian approximation can mismatch the Fisher-Tippett statistics, especially at low numbers of looks. Direct gamma-domain alternatives retain the original statistical model but create constrained, nonconvex optimization problems.

  • The homomorphic procedure applies a Gaussian image denoiser to log-transformed intensities and exponentiates the result to estimate reflectivity.
  • For low numbers of looks, the Fisher-Tippett distribution differs significantly from a Gaussian and has a heavier left tail, often leaving dark stains.
  • Direct gamma-domain estimation enforces positivity and uses a nonconvex objective, making solutions dependent on initialization and solver parameters.
  • Convexifying the gamma-domain objective can lose statistical accuracy and leave bright outliers because the gamma right tail is not captured.

D. Variational approach on log-transformed data

Log-domain variational processing makes the data-fidelity term convex and permits Gaussian-denoising subproblems within an unconstrained optimization. The resulting variance stabilization supports uniform smoothing across signal levels, while preserving contrast better in the illustrated bright structures.

  • D. Variational approach on log-transformed data: Log-domain variational estimation uses the Fisher-Tippett likelihood rather than replacing it with a Gaussian approximation.
  • D. Variational approach on log-transformed data: Changing variables to log-reflectivities makes optimization unconstrained and the data fidelity convex, reducing dependence on initialization and solver parameters.
  • D. Variational approach on log-transformed data: The regularization subproblem can be solved by a dedicated regularized least-squares solver corresponding to a Gaussian denoiser.
  • D. Variational approach on log-transformed data: MuLoG generalizes log-transform variational approaches to multi-channel SAR images while embedding a Gaussian denoising step.
  • D. Variational approach on log-transformed data: Because noise variance is signal-independent, convergence is practically uniform and produces similar smoothing in dark and bright regions.
  • D. Variational approach on log-transformed data: Log-domain total-variation minimization produces less contrast loss in bright structures and reduces speckle equally across regions.

III. EXTENSION TO MULTI-VARIATE SAR IMAGERY

Multi-variate SAR data are represented by complex covariance matrices whose Wishart-distributed speckle couples intra- and inter-channel fluctuations. Direct variational estimation is difficult because it is constrained, nonconvex, and can lose robustness or uniform suppression across regions.

  • Complex covariance matrices gather polarimetric and interferometric information at each multi-variate SAR pixel.
  • Wishart modeling captures signal-dependent fluctuations whose off-diagonal variance depends on diagonal covariance terms.
  • Direct variational estimation optimizes on the cone of Hermitian positive definite matrices and involves a highly nonconvex negative log-likelihood.
  • Convex-envelope approximations can underestimate the distribution’s right tail and reduce robustness to bright outliers.
  • Signal-dependent noise can make speckle suppression uneven across regions, with performance also dropping at high input dynamic range.
  • The proposed alternative extends the univariate log-transform strategy through the matrix logarithm for multi-channel SAR images.

C. The Wishart-Fisher-Tippett distribution

The matrix logarithm transforms positive-definite covariance matrices into Hermitian matrices while inducing the Wishart-Fisher-Tippett distribution. Its trace has additive, signal-independent bias and variance terms, so the transform supports approximate stabilization but not direct bias-corrected inversion.

  • The matrix logarithm maps a positive-definite covariance matrix into a complex Hermitian matrix by transforming its eigenvalues logarithmically.
  • Applying the transform to covariance data yields the Wishart-Fisher-Tippett distribution, whose normalization includes the exponential-map Jacobian.
  • The trace of the transformed covariance equals the covariance log-determinant and has explicit expectation and variance expressions.
  • The trace contains additive non-zero-mean, signal-independent corruption, while the transformed channels can reasonably be treated as approximately variance-stabilized.
  • Because no bias-correction inversion formula is available, the matrix logarithm is used within a variational strategy rather than as a direct stabilization procedure.

D. Log-channel decomposition

The log-transformed covariance is re-parameterized as a real vector with decorrelation and variance-scaling transforms. This produces channels that are approximately independent and similarly normalized for subsequent processing.

  • The log-transformed covariance matrix is represented as a vector of D^2 real coefficients.
  • A unitary transform maps real vectors to Hermitian matrices, while affine whitening and diagonal scaling decorrelate and balance channel variances.
  • The likelihood’s two-dimensional sections are examined with respect to channel pairs (x1, x2), (x1, x3), and (x3, x4).
  • The inverse re-parameterization maps log-transformed observations to real-valued data y used by the algorithm.
  • The affine transform is chosen so the re-parameterized channels are approximately independent, and scaling sets each channel variance to 1.

E. Proposed variational approach

MuLoG formulates multi-channel SAR estimation as an unconstrained variational problem in the re-parameterized log domain and solves it with ADMM and quasi-Newton updates. Numerical integration can be approximated cheaply while retaining small optimization error.

  • The MAP formulation extends scalar channel penalties to images of n real-valued vectors with D^2 channels.
  • The final covariance estimate is recovered by applying the matrix exponential to the optimized re-parameterized vector.
  • Optimization occurs unconstrained in R^D^2 rather than on the cone of positive-definite matrices, although the likelihood remains nonconvex.
  • ADMM separates the channel updates, while approximate Hessian updates yield a quasi-Newton iteration.
  • A single regularization parameter can be used across channels because the transformed noise variance is approximately stabilized near 1.

F. Adaptation to advanced filters

MuLoG replaces the Gaussian-denoising subproblem in ADMM with a plug-and-play Gaussian denoiser, enabling denoiser-based multi-channel SAR speckle reduction. Under a boundedness condition, the resulting algorithm is proven to converge.

  • MuLoG applies a Gaussian denoiser within ADMM instead of explicitly minimizing the regularization subproblem.The resulting plug-and-play scheme is called MUlti-channel LOgarithm with Gaussian denoising (MuLoG).
  • The Gaussian denoiser processes each transformed channel with noise level β^-1/2.
  • Convergence is proven when the denoiser is bounded, a condition expected from most denoisers.The condition also requires the denoiser to approach the identity as the noise variance tends to zero.
  • The implementation requires a small update of β during iterations to satisfy the convergence result.

G. Calibration

MuLoG calibrates its log-channel representation using PCA-based whitening and variance scaling, producing approximately decorrelated channels with unit noise variance. Positive-definiteness issues when L < D are handled by diagonal loading and spatial re-scaling.

  • The calibration computes PCA-based whitening and variance scaling from vectors extracted from the matrix logarithm of the input image.
  • The affine transform decorrelates channels, while the diagonal scaling matrix makes each transformed channel’s noise variance approximately 1.Noise variances are estimated separately, using MAD in practice.
  • Figure 5 shows that log-channel fluctuation amplitudes are fairly constant across underlying signals and similar across channels, supporting approximate variance stabilization.Some regions remain slightly noisier, so the stabilization is not exact.
  • Six iterations with β = 1 + 2/L provide satisfying solutions across tested dimensions, looks, and embedded Gaussian denoisers.The initialization uses x̂ = y, ẑ_i = f_1(y_i), and d̂ = ẑ − x̂, with Q = 1.
  • When L < D, rank deficiency makes the matrix logarithm undefined, so diagonal loading and spatially varying off-diagonal re-scaling enforce positive definiteness.
  • Local kernel weighting introduces bias, although the authors report that it does not appear significant.

IV. EFFICIENT MATRIX LOG AND EXP TRANSFORMS

The paper accelerates matrix logarithm, exponential, and product operations through vectorized and closed-form implementations, then evaluates MuLoG with several Gaussian denoisers on simulated and real SAR data. Results show denoiser-specific artifacts and performance broadly comparable to NL-SAR in the reported amplitude experiment.

  • Numerical experiments: The simulated amplitude experiment compares NL-SAR, MuLoG with TV, DDID, or BM3D, and homomorphic variants using PSNR and SSIM across looks L.
  • Numerical experiments: Real single-look interferometric and polarimetric examples cover airborne and spaceborne acquisitions with D = 2 and D = 3.
  • Numerical experiments: MuLoG gives more relevant simulated solutions than the homomorphic approach, while overall results are on a par with NL-SAR and vary slightly by Gaussian denoiser.The homomorphic approach oversmooths bright targets and leaves residual dark stains.
  • Numerical experiments: TV produces piece-wise constant results that lose small details and low-contrast features, whereas DDID and BM3D restore details but show small oscillating artifacts.
  • Numerical experiments: NL-SAR is slightly over-smoothed but lacks the systematic oscillating artifacts associated with DDID and BM3D.

VI. CONCLUSION

MuLoG addresses the limited availability of multi-channel SAR speckle reducers by using a matrix-log transform to stabilize variance and enable user-selected Gaussian denoisers. The resulting family preserves the Wishart-based covariance structure, requires little tuning, and supports artifact comparison and future benchmarking.

  • MuLoG approximately stabilizes speckle variance and produces close-to-independent channels before Gaussian denoising.
  • Iterative processing enforces a good fit between the restored multi-channel image and the Wishart distribution of input covariance matrices.
  • The method requires no parameter tuning beyond possible settings inside the Gaussian denoiser, which can be tuned once for all.
  • Different embedded denoisers produce multiple reduced-speckle images that can be compared to discard artifacts and confirm weak structures.
  • The resulting method family can serve as a reference for benchmarking future specialized multi-channel SAR speckle reducers.

APPENDIX A GRADIENT OF THE NEG LOG LIKELIHOOD

The appendix derives the gradient component of the negative log-likelihood by exploiting the affine form of Ω(x), matrix-valued differentiation identities, and Hermitian structure.

  • The derivation begins from a real-valued differentiable function and applies the adjoint-based matrix chain rule to obtain the gradient structure.
  • Because Ω(x) = K(AΦx + b) is affine, its derivative simplifies the gradient calculation.
  • The appendix invokes identities for matrix-valued functions and applies them to an arbitrary matrix L in the gradient expression.
  • The integral term is Hermitian because it is formed by integrating Hermitian matrices.
Loading 1704.05335v1…