Edge-Gated Global Propagation from Non-Convex Priors

Image restoration networks face two opposite demands. They need a wide view, because recovering one pixel often depends on evidence far away: a streak in a sparse-view CT scan or a smeared spectral band spans the whole image. They also need to keep edges crisp. The usual ways to widen the view, such as self-attention or non-local blocks, average over many pixels, and averaging blurs edges. Classical non-convex regularizers point to a different way to reconcile the two. These are penalties on image gradients whose cost stops growing once a gradient is large enough to be a real edge. When such a penalty is minimized by the standard reweighting trick, each gradient receives a weight that is close to zero at strong edges. Solving the reweighted smoothing problem exactly along an image row means inverting a banded matrix, and that inverse is dense. Every pixel then influences every other pixel on the row, except that the influence is cut off wherever a strong edge sits in between.

We propose to build an image restoration model's long-range operator from this fact rather than from attention. Inside an unrolled half-quadratic splitting solver, where each stage of an iterative optimization algorithm becomes a network layer, we replace the local learned denoiser with exact row and column solves of a learned non-convex prior's reweighted quadratic subproblem. Each pass costs time linear in the number of pixels. Two passes, one over rows and one over columns, reach the whole image, much as criss-cross attention does. Information does not leak across edges, because the prior itself sets the cutoff. The claim we will test is that this gives the global context of non-local attention without its edge-blurring bias, and does so while keeping an energy the solver provably decreases.

Core Claim

Existing work treats "non-convex edge preservation" and "escaping the local receptive field" as separate modules that get mixed together. DNU unrolls half-quadratic splitting for snapshot spectral imaging and runs a convolutional branch alongside a non-local attention branch, blended by one learned weight. Its prior is a black box. The authors note that the non-local branch's memory cost limits resolution, and nothing prevents that branch from averaging across edges. Butterworth as Attention supplies the motivating diagnosis: dot-product self-attention behaves as a low-pass filter and erodes high-frequency detail. Going the other way, IHQS$_p$-Net and DECUN keep an explicit non-convex or convergent structure, but they update the prior with pixel-separable thresholding or local convolutions, so their receptive field stays local. Our claim is that the two can be merged into a single operator. The global receptive field comes from exactly solving the non-convex prior's linearized subproblem, so where information can travel is decided by the edge-preservation criterion itself.

Concretely, with observation $y$, known forward operator $A$ (a blur, a CT projector, or a coded-aperture mask), and learned penalty $\Phi$ with a redescending slope, we unroll the multiplicative half-quadratic form of the objective:

$$E(x, z) = \|Ax - y\|^2 + \lambda \sum_i \big( z_i\, (\nabla x)_i^2 + \psi(z_i) \big),$$

Here $z_i \ge 0$ is a per-gradient weight and $\psi$ is the conjugate function that makes minimizing over $z$ recover $\Phi$. The optimal weight is

$$z_i = \frac{\Phi'(|g_i|)}{2|g_i|}, \qquad g_i = (\nabla x)_i.$$

For hyper-Laplacian, minimax concave, or Tukey-type penalties, $z_i$ falls toward zero on large gradients. The $x$-update with $z$ held fixed is a spatially varying screened-Poisson problem. We do not solve it with an FFT or conjugate gradients. Instead we split it into alternating one-dimensional problems along rows and columns. Each one is tridiagonal and is solved exactly with the Thomas algorithm, or with cyclic reduction on GPU. The entries of the inverse decay as products of the $z$ weights along the path, which is why edges act as barriers. Each stage learns only the penalty's shape parameters (in the spirit of the learned $p^{(t)}, \lambda^{(t)}$ of Zuo et al.), a small convolutional feature map that supplies the gradients $g$ on which weights are computed, and the step constants. Gradients during training are obtained by implicit differentiation, which solves the transposed tridiagonal system.

Research Program

The work has three parts. The first is the operator. We parameterize $\Phi$ as a monotone spline with a concavity constraint, so that $p$-norms, minimax concave penalty, and Tukey-like shapes are all reachable. We implement batched GPU tridiagonal solves with implicit backward passes. To limit the axis bias of row and column scans, we add two diagonal passes as an option. The bias concern is the same one that Bui et al. address with the anisotropic-minus-isotropic TV penalty, and that EccoMamba patches heuristically with axial shifts after flattening images for sequence models.

The second part is convergence. When the penalty is shared across stages, each weight update and each exact line solve is a block-coordinate descent step on $E$, so $E$ decreases monotonically. We will prove this for the row/column alternation. We then allow stage-specific parameters only as vanishing perturbations around shared values, following DECUN's construction. We will check empirically that running the model for more stages than it was trained on converges instead of diverging, a property DNU's unconstrained attention branch does not provide.

The third part is analysis. We will measure effective receptive fields and cross-edge leakage. Leakage is the fraction of an output pixel's sensitivity that comes from across a ground-truth boundary. We compare it between our operator, DNU-style non-local attention, and criss-cross attention (CCNet).

Baselines

The naive reference points are filtered backprojection on CT, which IHQS$_p$-Net reports at 23.64 dB PSNR on AAPM with 180 views, and the classical hyper-Laplacian half-quadratic solver of Krishnan and Fergus with hand-set continuation. Neither involves learning or global propagation. Any learned method should clear them comfortably.

On the spectral-imaging question, the strongest competitor is DNU. It is the only method in the file that already combines unrolled half-quadratic splitting with a non-local prior, and it reports 34.27 dB PSNR on ICVL CASSI reconstruction, about 2.5 dB above ISTA-Net. We will retrain DNU on the same splits with the same number of stages, the same optimizer budget, and hyperparameter searches of equal size. We will report both its published number and our retrained number. Beating it means matching or exceeding its PSNR at lower memory, together with clearly lower cross-edge leakage and better edge-restricted PSNR. PSNR parity alone would not support our claim.

On sparse-view CT, the strongest competitor is IHQS$_p$-Net, with 33.18 dB on AAPM at 180 views, just ahead of MetaInvNet at 33.01 dB. We will reuse its data protocol and swap only its prior update for our operator, so the comparison isolates the receptive-field question. Beating it means a gain above run-to-run variance across seeds, and no increase in the hallucinated "pseudo-connections" that paper emphasizes, which we will score on vessel and bone boundaries.

On non-blind deblurring, DECUN reports 33.24 dB on BSD linear kernels with 1,117 parameters. This is the tightest baseline for parameter efficiency, so we will compare at matched parameter counts.

Three ablations isolate the key ingredient. The first uses a convex-TV weight $z_i = 1/(2|g_i|)$, clipped, in place of the learned redescending weight. This tests whether non-convexity, and not just edge-aware propagation, matters. The second uses constant weights, which gives a global but edge-blind propagation. The third keeps the learned non-convex weights but truncates the line solve to a local window. This tests whether the global reach matters.

Why not just…?

Why not just put a non-local or transformer block inside the unrolled solver, as DNU does? That design is the direct competitor, and it could match us on PSNR. The file's evidence against it is structural rather than numerical. DNU's authors flag memory limits at high resolution. The Butterworth analysis (Wang et al.) shows softmax attention is low-pass, and a single learned mixing weight cannot stop that branch from averaging across edges. The leakage measurement and edge-restricted PSNR on ICVL will settle it. If DNU matches us on both, our central claim fails.

Why not just use a deeper or dilated local CNN denoiser in a DECUN or IHQS$_p$ style unrolling? Stacking local layers dilutes long-range context (ConvFormer). On CT, Guo et al. report that plain deep CNN reconstructions produce pseudo-connections. This alternative is still a real threat on deblurring, where kernels are small and local context may suffice. For that reason we will include a local denoiser matched in parameters and runtime, and we expect our advantage to appear mainly on CT streaks and coded-aperture smearing, where the artifacts are global.

Why not just solve the full 2D reweighted quadratic step with conjugate gradients instead of line solves? This is the exact half-quadratic update and is arguably more principled. Its costs are iteration counts that depend on how badly conditioned the problem gets near edges, and backward passes that are expensive. We will run it as an upper-bound reference at equal wall-clock time. If it is both cheaper and better, the row/column decomposition is not justified.

Why not just use convex TV weights, which give a unique minimizer? TV suffers shrinkage bias and staircasing, and on mixed noise it trails a hyper-Laplacian prior combined with non-local modeling by about 2 dB on average (Zhang et al.). Our convex-weight ablation tests this directly inside the learned model, where learned features might close the gap.

We also acknowledge that recursive edge-aware filters of a similar form exist in the graphics literature. The file does not cover them, and we will position against them directly. What we add is weights derived from a learned non-convex prior, embedded in a physics-aware unrolled solver with a descent guarantee.

Proposed Experiments

The decisive experiment is ICVL CASSI reconstruction. There we compare against retrained DNU on PSNR, on PSNR within a band of a few pixels around ground-truth edges, on cross-edge leakage, and on peak memory as resolution scales from 512 up to 2048 pixels on a side. The second experiment is AAPM sparse-view CT at 180 and fewer views, with zero-shot transfer to the COVID-19 lung CT set used by IHQS$_p$-Net. The third is BSD non-blind deblurring against DECUN at matched parameters, plus the classic mixed Gaussian-impulse images (for example Lena, where Zhang et al. report 33.00 dB) to test abrupt corruptions.

We will add two diagnostic checks. The first runs the model for twice the training stage count to confirm the energy keeps decreasing. The second evaluates on test images rotated by 30 and 45 degrees, with and without the diagonal passes, to measure scan-axis bias.

The proposal is supported if our operator matches DNU's accuracy with markedly lower leakage and memory, and beats IHQS$_p$-Net on CT with the non-convex ablation accounting for most of the gain. It is refuted if constant-weight global propagation, or a plain local denoiser, performs as well.