Coarse-to-Fine Curvature Steering in Shift-Invariant Shearlet Subbands

Denoisers that follow image geometry need to know how edges bend, not just where they are. A plain edge-preserving filter treats a tight loop or a sharp corner as noise and straightens or rounds it off. Curvature-adaptive diffusion avoids this by steering the smoothing along arcs, using second derivatives of the image. The catch is that second derivatives are exactly what noise destroys. At the fine scales where the steering matters most, the curvature estimate is the least trustworthy. Shift-invariant multi-scale transforms have the opposite profile. These are filter banks, such as the non-subsampled shearlet transform, that split an image into frequency octaves and orientations without downsampling, so every subband stays pixel-aligned with the input. They isolate noise into fine bands very well, but they then only threshold or average coefficients and never let geometry steer the result.

This note proposes running curvature-steered diffusion directly on the directional subbands of a non-subsampled shearlet transform, treated as a multi-scale orientation volume. The key move is that the curvature used to steer each fine scale is measured at the next coarser scale and passed down. Because the transform is undecimated, coarse and fine subbands are co-registered pixel for pixel, so a curvature field computed where the signal-to-noise ratio is high can be applied directly where it is low. The claim to test is specific. Coarse-to-fine curvature transport preserves high-curvature structure (loops, hairpins, corners, crossings) under heavy noise better than two simpler options: estimating curvature from a Gaussian-smoothed image at a single, tuned scale, or estimating it independently within each scale. It should also cost no more than existing lifted-domain PDE methods.

What Exists and What Is Missing

The method side comes from the orientation-score line of work. Crossing-Preserving Coherence-Enhancing Diffusion on Invertible Orientation Scores and Nonlinear Diffusion on the 2D Euclidean Motion Group lift an image into a position-and-orientation volume with directional filters. They estimate curvature from a Hessian in that volume and diffuse along circular arcs, which keeps crossing fibers separate. Total Variation and Mean Curvature PDEs on the Homogeneous Space of Positions and Orientations extends this with total-variation and mean-curvature flows. Its flows are steered by "gauge frames," local coordinate frames obtained from a singular value decomposition of the lifted Hessian.

All three share two properties. They lift at a single scale, and they estimate curvature from the noisy data at that same scale, with only Gaussian regularization. The authors themselves flag the cost of the lifted computation and the lack of a stopping rule. We reuse their steering idea but change where the curvature comes from and how many scales it acts on.

The representation side comes from directional multi-scale transforms. The shearlet edge analysis in Edge analysis and identification using the continuous shearlet transform shows that shearlet coefficients locate edges and their orientation. Contourlets (Contourlets and sparse image expansions) and curvelets (Recovering edges in ill-posed inverse problems: optimality of curvelet frames) establish why multi-scale anisotropic elements represent smooth edges sparsely. Synthetic-Aperture Radar Image Despeckling based on Improved Non-Local Means and Non-Subsampled Shearlet Transform is the closest working system on the multi-scale side. It decomposes radar images with the non-subsampled shearlet transform and filters the subbands with a sparsity-weighted non-local means, but the filtering ignores geometry.

Bendlet Transform Based Adaptive Denoising Method for Microsection Images comes nearest to putting curvature inside a multi-scale frame. It uses bent, second-order shearlet atoms, but its curvatures are a fixed dictionary and its denoising is coefficient shrinkage. We differ from both in the same way: the subband coefficients evolve under a flow whose direction and bending are read off the data at a coarser scale.

Two results constrain the design. First, A Level Set Filter for Speckle Reduction in SAR Images uses curvature to kill speckle, because speckle looks like high-curvature spikes. Sun et al. instead argue that speckle corrupts pointwise derivatives so badly that transform-domain averaging is needed. Our design takes both seriously. Fine-scale curvature is not trusted; coarse-scale curvature is.

Second, Different Faces of the Shearlet Group shows that the shearlet group is structurally different from the rotation-and-translation group on which the orientation-score PDEs are defined. We therefore do not claim a group-invariant flow. Instead, we map each shear index to its true angle and diffuse in that angle-space, and we report how much rotation-equivariance is lost.

Core Claim

Let $w_{j,k}(x)$ be the non-subsampled shearlet coefficient at scale $j$ (larger $j$ is finer), shear direction $k$ (mapped to angle $\theta_k$), and pixel $x$. Let $\kappa_{j-1}(x)$ and $c_{j-1}(x)$ be curvature and orientation confidence. They are computed from the Hessian of the magnitude volume $|w_{j-1,\cdot}|$ at the next coarser scale, using the gauge-frame construction of Smets et al. Each fine directional subband stack then evolves as

$$\partial_t w_{j,\cdot} = \nabla_{(x,\theta)} \cdot \big( D(\kappa_{j-1}, c_{j-1}) \, \nabla_{(x,\theta)} w_{j,\cdot} \big).$$

Here $D$ is a tensor in position-and-angle that diffuses strongly along the arc of curvature $\kappa_{j-1}$ through each point. It shuts diffusion off where confidence $c_{j-1}$ is low, which is the case at junctions and corners, following the freezing logic of Feature-Preserving Flows: A Stochastic Differential Equation's View. The coarsest band is steered by its own curvature, since it is the cleanest.

Each scale stops at its own time $t_j$, chosen by a per-band discrepancy rule:

$$\|w_{j,\cdot}(t_j) - w_{j,\cdot}(0)\|^2 = \sigma^2 \|g_j\|^2 N.$$

Here $\sigma^2$ is the noise variance, $\|g_j\|^2$ is the known energy of the scale-$j$ analysis filters, and $N$ is the number of coefficients. The rule says to remove roughly the noise energy expected in that band and no more. This addresses the missing stopping rule noted by Smets et al. Its form follows from white noise passing through a linear filter bank.

The prediction is that the gain should concentrate where the local radius of curvature is comparable to the scale-$j$ atom length. In that regime single-scale curvature from noisy data is unreliable, while the coarser estimate is still valid. If the gain is instead uniform, or absent in exactly that regime, the mechanism is not what we claim.

Research Program

We will first build the pipeline in 2D. It has four parts: a non-subsampled shearlet decomposition with 4 to 5 scales and 8 to 16 directions; a mapping from shear indices to angles with the correct non-uniform spacing; an explicit finite-difference scheme for the steered diffusion; and inverse reconstruction. For radar data we will add the log-transform used by Sun et al.

Second, we will characterize the curvature estimator alone. On synthetic images with known curvature (concentric rings of decreasing radius, spirals, polygons with known corner angles, and X- and Y-crossings) we will measure curvature-estimation error against noise level. We compare three sources: coarse-scale shearlet curvature passed to the finer scale, same-scale shearlet curvature, and Gaussian-presmoothed image curvature at swept integration scales. This isolates whether the transport idea helps before any denoising is involved.

Third, we will test the full denoiser on real and benchmark data, reporting curvature-resolved error alongside global scores.

Fourth, we will measure runtime and memory against the single-scale lifted PDE. The shearlet redundancy of $\sum_j K_j + 1$ subbands is comparable to a single-scale lift with a similar total number of orientations, so we expect similar cost. We will report it rather than assume it.

Baselines

The naive reference points are the noisy input and a plain Gaussian or isotropic heat-equation filter tuned for PSNR (peak signal-to-noise ratio). Every method must clear these by a margin that grows with curvature.

The strongest existing method is the gauge-frame total-variation flow of Total Variation and Mean Curvature PDEs on the Homogeneous Space of Positions and Orientations. It is the strongest competitor because it already does curvature-steered flow in an orientation-lifted space, which is exactly our mechanism minus the scale pyramid and the cross-scale curvature transport. On the noisy Mona Lisa image it reports 26.1 dB PSNR against 23.9 dB for BM3D. We will run it with its own number of orientations and tune its stopping time by oracle PSNR, which is generous to it, while our method uses the automatic per-band rule. We will match total wall-clock budgets and use identical noise realizations. Beating it means at least matching 26.1 dB on Mona Lisa under the same noise protocol. More importantly, it means a clear reduction in error on structures whose radius of curvature is a few pixels, both on the synthetic phantoms and on the high-curvature fraction of real images.

BM3D (block-matching 3D, which groups similar patches and filters them jointly) is the standard patch-based reference. Its Mona Lisa number above comes from Smets et al. Bendlet Transform Based Adaptive Denoising Method for Microsection Images reports that bendlets beat it on brain MRI, with 41.41 dB and SSIM (structural similarity) above 0.99, and that BM3D was the slowest method. We will run the standard implementation with its noise level given.

On radar data the strongest reported competitor is the non-subsampled shearlet plus sparsity-weighted non-local means of Synthetic-Aperture Radar Image Despeckling based on Improved Non-Local Means and Non-Subsampled Shearlet Transform. It reports an ENL (equivalent number of looks, a measure of smoothing in flat areas) of 1.42 and a DCV (an edge-detail score where lower is better) of 0.06, against 1.12 and 0.10 for plain non-local means. Because it shares our exact decomposition, it is the cleanest test of whether geometry-steered flow beats geometry-blind averaging inside the same subbands. Beating it means a higher ENL at equal or lower DCV, measured on the same kind of GF-3 scenes and the same homogeneous and edge regions.

The bendlet denoiser is the curvature-aware frame baseline. On our own MRI-style data we will compare against it at matched noise levels, rather than against its reported 41.41 dB, since its test images are not public in the file.

Four ablations isolate the key ingredient, each with an identical decomposition, solver, and stopping rule:

  • steering each scale with its own same-scale curvature;
  • steering with Gaussian-presmoothed image curvature at the best single integration scale, chosen by oracle;
  • setting curvature to zero, so the flow becomes straight-line coherence diffusion inside the subbands;
  • replacing the undecimated transform with a decimated one and interpolating the curvature across scales, which tests whether pixel alignment across scales is what makes transport work.

Why not just…?

Why not just run the single-scale gauge-frame flow, which already beats BM3D? It is the real threat, and on smooth photographs it may tie. The experiment that settles it is the curvature-resolved error on phantoms with mixed line thicknesses and tight bends at high noise. If the single-scale flow matches us there with oracle stopping, the multi-scale part adds nothing beyond automatic stopping, and we will say so.

Why not just smooth the image with a Gaussian before computing the Hessian, as structure-tensor methods already do? This is the most direct rival to our curvature estimator. Gaussian presmoothing at one scale shrinks small loops and rounds corners, which biases the very curvature it measures. Our transport instead uses a coarser estimate that is still pixel-aligned and orientation-resolved. The oracle-scale Gaussian ablation in the estimator study tests this directly. If it matches coarse-to-fine transport across the noise range, the core claim is wrong and the proposal reduces to a cheaper implementation of the existing flow.

Why not just threshold the shearlet coefficients, or threshold them and then apply a curvature flow? Fixed-direction thresholding scatters curved edges across wedges, as the bendlet work notes. Thresholding also cannot bridge interrupted fibers, as the comparison in Smets et al. indicates. We include a sequential "threshold then single-scale flow" variant at matched cost. If it matches us, the value lies in sequencing rather than coupling, and that would be worth reporting too.

Why not just use BM3D or non-local means? The reported numbers already favor geometric methods on Mona Lisa and favor shearlet-domain filtering over plain non-local means on radar data. Patch methods also have no mechanism for one-off structures such as a unique vessel tree. We include them because these reported comparisons come from the competing methods' own authors.

Proposed Experiments

The first decisive check is the estimator study on synthetic curvature phantoms. It plots curvature error against noise for transported, same-scale, and oracle-Gaussian curvature. We need transported curvature to win at moderate-to-high noise for radii near the fine atom length, or the project stops there.

The second is the curvature-resolved denoising study. It uses the same phantoms plus noisy Mona Lisa under the Smets et al. noise protocol, and reports PSNR overall and binned by local curvature, with corner-angle error and crossing separation measured as residual intensity between fibers.

The third is radar despeckling on GF-3 scenes, scored by ENL and DCV against the shearlet non-local means result. It probes whether coarse curvature survives multiplicative noise, which is where Huang et al. and Sun et al. disagree.

The fourth is a shift-consistency test. We shift inputs by one pixel and measure the output change, to confirm that the decimated ablation actually degrades.

Finally, on the REVIEW retinal database we will measure vessel-width error after denoising, as a downstream structural check. No reported result found on Lacuna gives numeric REVIEW scores for the orientation-score tracker in A Multi-Orientation Analysis Approach to Retinal Vessel Tracking, so we will rerun a common tracker on each denoiser's output rather than compare against a quoted figure.

The proposal succeeds if the gain over both the single-scale gauge flow and the oracle-Gaussian ablation is concentrated where the theory says it should be: small radii, high noise. Global PSNR is secondary.