Blind deblurring means recovering a sharp photo and the unknown blur that degraded it, when all you observe is the blurry photo. It is an old problem with a well-known trap. The "explanation" that nothing was blurred, with the sharp image equal to the blurry one and the blur equal to a single spike, often scores better under standard image priors than the truth. Practical methods escape the trap with hand-built scaffolding: coarse-to-fine image pyramids, centered kernel initializations, and schedules that anneal regularizer weights.
A separate body of theory shows that for idealized versions of the same problem, the scaffolding is unnecessary. For sparse spike trains blurred by random filters, gradient descent on the sphere of unit-norm filters starts from a random point and still reaches the true blur, because every local minimum is the truth up to shift and sign. This note proposes to bring that theory to single photographs with real camera shake. We would identify which of its assumptions break, repair the two that matter most, and test whether a single-scale, randomly initialized method can match pyramid-based pipelines on the standard benchmarks.
Core Claim
The claim has two parts. First, the image gradients of a blurry photo behave much like the "short-and-sparse" signals of the theory. The unit-norm kernel constraint that practitioners already use to avoid the spike trap is the same sphere constraint the theory analyzes. So the failures of random-start descent on real photos are not arbitrary. We predict they fall into two recognizable families.
The first family is shift-truncations: kernels that are a cropped, shifted piece of the true blur. These are predicted by Structured Local Minima in Sparse Blind Deconvolution, which proves that on the sphere all local minima of the short-and-sparse problem are such truncations and all saddles are strict. Strict saddles are points with at least one downhill direction, so descent does not get stuck there. The same form of minimum appears in On the Global Geometry of Sphere-Constrained Sparse Blind Deconvolution, which shows that replacing a sum-to-one kernel constraint with a unit-length one pushes descent away from the spike solution.
The second family is spectral-null failures. Real motion and defocus blurs erase some frequencies entirely. The whitening step that all these proofs rely on then becomes undefined, and descent drifts inside the erased subspace.
Second, both families can be removed with two concrete changes. We replace exact whitening with damped whitening. We replace a fixed kernel support with a lifted, oversized support that is re-centered during descent. With these changes, one randomly initialized run should match coarse-to-fine methods without any pyramid or annealing.
Concretely, following the inverse-filter formulation of A Nonconvex Approach for Exact and Efficient Multichannel Sparse Blind Deconvolution, we solve
$$\min_{q \in S^{n-1}} \; \sum_j h_\mu\!\left( \big[\, g \circledast P_\varepsilon q \,\big]_j \right), \qquad P_\varepsilon = (C_g + \varepsilon I)^{-1/2}.$$
The terms are defined as follows. $g$ is a derivative image of the blurry photo. $q$ is a candidate inverse filter on the unit sphere $S^{n-1}$ of length-$n$ filters. $\circledast$ is convolution. $h_\mu$ is the Huber function, a smooth stand-in for sparsity with smoothing width $\mu$. $C_g$ is the circulant autocorrelation of $g$. $\varepsilon$ is a damping level.
When $\varepsilon = 0$ this is exactly the preconditioned landscape of Qu et al., and it fails when $C_g$ is singular. With $\varepsilon > 0$ the preconditioner behaves like a Wiener filter on the erased frequencies. The kernel is then read off from $q$ and refined by projected descent on the $\ell_2$ kernel sphere, as in Normalized Blind Deconvolution.
Research Program
The program has three components.
Measuring the landscape. We run the objective above, and separately the Post-Normalization (PN) descent of Jin et al., from hundreds of random kernel initializations per image. We classify every endpoint as one of four outcomes: true kernel up to shift, shift-truncation, spike solution, or other. This population count of endpoints, which we call a landscape census, directly tests the prediction that shift-truncations and null-subspace drift account for nearly all failures.
Jin et al. prove only that the spike is no longer the global minimum under $\ell_2$ normalization. They leave open whether spurious minima exist and whether random starts work, and the census answers that empirically.
Analyzing a simplified model. We treat a one-dimensional model in which the sparse signal is Bernoulli–Gaussian, meaning each entry is zero with some probability and Gaussian otherwise, and the kernel is low-pass with exact spectral nulls. For this model we aim to prove that, with damping $\varepsilon$, every local minimum lies within a distance of a shift-truncation that grows with the fraction of kernel energy lost at the nulls. We also aim to prove that saddles remain strict.
This extends Global Geometry of Multichannel Sparse Blind Deconvolution on the Sphere, which requires an invertible filter. We treat that paper, along with Qu et al. and Zhang–Kuo–Wright, as supplying the method. The trajectory-tracking style of Gradient Descent with Random Initialization: Fast Global Convergence for Nonconvex Phase Retrieval is a fallback if uniform landscape bounds prove too loose. The file itself records how loose they are: the multichannel proof requires on the order of $n^4$ channels, while experiments show polylogarithmic requirements suffice.
Removing shift-truncations by lifting. We run descent on a kernel support twice the assumed size and periodically re-center the kernel's energy. This is meant to replace the continuation schedule on the sparsity weight that Zhang et al. need in their second stage.
We are candid about one constraint. How Diffusion Prior Landscapes Shape the Posterior in Blind Deconvolution shows that even learned diffusion priors grow monotonically more favorable to smoother images, which tilts the whole landscape toward the spike solution. Our bet does not contradict this. We do not rely on the image prior to reject blur; the sphere geometry does that work. If the census shows the spike solution still captures many random starts under the sphere constraint, the bet is wrong.
Baselines
The naive reference is the blurry input itself, equivalently the spike solution. Alongside it we run plain $\ell_1$-normalized total-variation descent from random initialization, which we expect to collapse to the spike. This baseline shows how bad random starts are without the sphere.
The strongest competitor on this exact question is Post-Normalization from Normalized Blind Deconvolution. It uses the same $\ell_2$ kernel sphere with a total-variation prior and centered initialization. On Levin it reports a normalized $L_2$ error of 3.788, against 5.452 for an earlier $\ell_2$ normalization approach, and on the SUN dataset it reports a mean error ratio of 2.299. We run it twice. The first run uses its published centered initialization with its pyramid, the configuration that produced those numbers. The second run uses random initialization at a single scale, so that the value of the scaffolding is measured directly.
The strongest competitor on real camera shake is automatic gradient activation, Blind Image Deconvolution by Automatic Gradient Activation. It reports 28.75 dB PSNR (peak signal-to-noise ratio) and 0.8768 SSIM (structural similarity) on Köhler. Beating our method would mean matching these within noise while using one scale and random starts. Clearly exceeding them is not required for the central claim.
We also include the large-to-small kernel initialization heuristic of Nguyen et al., transplanted onto the $\ell_2$-sphere objective. It is the most direct competing answer to "how do you start without a warm start."
All methods receive the same non-blind final deconvolution step, the same noise levels, and the same wall-clock budget per image. Where a method has regularization weights, it gets the same tuning grid on a held-out subset of BSDS.
Ablations isolate the two ingredients. We compare damping $\varepsilon = 0$ with $\varepsilon > 0$, and fixed support with lifted support. We also swap the Huber loss for total variation, which shows whether the sparsity surrogate or the sphere is doing the work.
Why not just…?
Why not just keep Jin et al.'s centered initialization and pyramid, since it already works? It may well work as well on benchmark accuracy. If so, the benchmark comparison alone will show it, and the proposal's contribution shrinks to explanation. The census and the one-dimensional analysis would still show why the $\ell_2$ sphere works and where it fails, which Jin et al. do not provide. The practical payoff of removing the pyramid is that the method no longer depends on hand-tuned schedules, and that becomes measurable only if the single-scale random-start variant matches it. We report that comparison regardless of which way it goes.
Why not just run Post-Normalization from many random starts and keep the lowest-objective result? This is the most serious threat. If the landscape is as structured as we predict, multi-restart should also succeed, and it needs no new objective. We therefore include it at equal compute. Our method wins only if it needs markedly fewer restarts to reach the true basin, or if restarts keep landing on spectral-null failures that damping fixes. The census measures exactly this. If multi-restart matches our method at equal budget, we will say the damping and lifting are unnecessary on these benchmarks.
Why not just use spectral initialization? Global Geometry of Multichannel Sparse Blind Deconvolution on the Sphere reports that truncated power iteration, a classical spectral warm start, fails where randomly initialized manifold descent succeeds when activations are dense or off-grid. Dense, off-grid activations are the normal situation for gradients of natural images. We include a spectral warm start of the same objective as an extra baseline to confirm this on photographs.
Why not just train an end-to-end deblurring network? Such networks target a different regime. Scale-Recurrent Network for Deep Image Deblurring reaches 30.26 dB on GOPRO but requires paired training data and does not output a kernel. Our target settings are kernel estimation benchmarks and scientific imaging, where no paired data exist. We report it on Köhler only for context.
Proposed Experiments
The first decisive check is synthetic. We use one-dimensional Bernoulli–Gaussian signals blurred by kernels with a controlled number of spectral nulls. On this data we measure the recovery rate from random starts as a function of sparsity, kernel length, and null count, with and without damping. If damping does not move the phase transition, meaning the boundary between success and failure, the second part of the core claim fails.
The second check is the landscape census on Levin, with 32 images and 8 measured kernels. For each image we use 200 random starts and report the fraction of endpoints in each failure class. The prediction is that shift-truncations and null-subspace drift together account for more than 90% of failures, and that lifting plus damping lowers the total failure rate below that of single-scale Post-Normalization.
The third check is benchmark accuracy. We report normalized kernel error on Levin, the error-ratio distribution on the 640 SUN images, PSNR and SSIM on the 48 Köhler images, and PSNR under 1–5% Gaussian noise on BSDS, all against the baselines above at matched compute. Noise is where How Diffusion Prior Landscapes Shape the Posterior in Blind Deconvolution observed collapse of the basin around the true solution. We therefore track how the success rate from random starts degrades as noise grows.
The fourth check tests whether the approach transfers beyond camera shake. We apply it to the scanning tunneling and fluorescence microscopy settings used by Zhang et al. and Qu et al. There, no competing pyramid-based pipeline was designed, and random-start guarantees matter most. No reported numbers for these microscopy settings were found on Lacuna, so they will be reported qualitatively against those papers' published reconstructions.
References
- Structured Local Minima in Sparse Blind Deconvolution
- On the Global Geometry of Sphere-Constrained Sparse Blind Deconvolution
- A Nonconvex Approach for Exact and Efficient Multichannel Sparse Blind Deconvolution
- Normalized Blind Deconvolution
- Global Geometry of Multichannel Sparse Blind Deconvolution on the Sphere
- Gradient descent with random initialization: fast global convergence for nonconvex phase retrieval
- How Diffusion Prior Landscapes Shape the Posterior in Blind Deconvolution
- Blind Image Deconvolution by Automatic Gradient Activation
- Scale-Recurrent Network for Deep Image Deblurring