Alternating Optimization for Blind Image Deconvolution

34 concepts32 papers1 proposal

Imagine you are handed a blurry photograph $y$. You know it was formed by convolving a sharp, latent image $x$ with an unknown blur kernel $k$ (plus some sensor noise $n$):

$$y = x * k + n$$

Your goal is to recover both $x$ and $k$. This is the classic problem of blind image deconvolution, and it is fundamentally ill-posed. Because both the image and the kernel are unknown, infinitely many pairs of $(x, k)$ can perfectly explain the observed blurry image $y$.

The most frustrating manifestation of this ambiguity is the Dirac trap: a trivial mathematical solution where the estimated image is simply the blurry input ($x = y$) and the estimated kernel is a single, perfectly sharp pixel ($k = \delta$). To prevent an optimization algorithm from falling into this trap and instead guide it toward a physically meaningful, sharp image, researchers rely on a powerful algorithmic strategy: alternating optimization.

The Core Intuition: Exploiting Bilinear Symmetry

Why does alternating optimization work so well for blind deconvolution? The key lies in the algebraic structure of the blur equation. While optimizing for $x$ and $k$ simultaneously is a highly non-convex, non-linear nightmare, the system is bilinear (Sivic et al.).

This means the problem possesses a helpful symmetry:

  1. Fix the kernel $k$, and finding the sharp image $x$ becomes a standard, well-behaved non-blind deconvolution problem.
  2. Fix the image $x$, and finding the blur kernel $k$ simplifies to a linear least-squares problem, which can be solved efficiently in the gradient domain (Anger et al.).

Alternating optimization exploits this symmetry by breaking the complex joint problem into two simpler, decoupled sub-problems solved sequentially (Suo et al., Chen et al.). By taking a step to improve the image, freezing that result, and then taking a step to improve the kernel, the algorithm navigates the complex optimization landscape via coordinate descent (Shearer et al., Bredell et al.).

┌──────────────────────────────┐
│  Initial Blurry Image (y)    │
└──────────────┬───────────────┘
               │
               ▼
 ┌────────────────────────────┐
 │ Fix Kernel, Update Image   │◄────────────────┐
 └─────────────┬──────────────┘                 │
               │                                │ Iterative
               ▼                                │ Alternating
 ┌────────────────────────────┐                 │ Loop
 │ Fix Image, Update Kernel   │─────────────────┘
 └─────────────┬──────────────┘
               │
               ▼
┌──────────────────────────────┐
│ Sharp Image & Correct Kernel │
└──────────────────────────────┘

Escaping the Dirac Trap via Strategic Asymmetry

If both sub-problems are solved perfectly, why doesn't the algorithm still slide down into the trivial Dirac solution? To prevent this, modern frameworks introduce deliberate asymmetries and structural constraints into the alternating loop.

Update Imbalance

Nguyen et al. demonstrate that while the Dirac trap is a global minimum, a favorable "sharp" solution still exists as a stable local minimum. To bias the optimization toward this sharp minimum, they propose an imbalanced alternating scheme: updating the image estimate at a much higher frequency than the blur kernel parameters. This ensures that image reconstruction remains the primary driver of the optimization process, stabilizing the recovery of sharp features.

Kernel Elimination

Another strategy is to remove the kernel as an independent optimization variable entirely. Lee et al. derive the blur kernel analytically as a direct mathematical function of the current image estimate. This collapses the search space of the ill-posed problem and guarantees that the estimated kernel remains physically consistent with the recovered image structures at every step of the iteration.

Proxies and Support Constraints

Instead of optimizing on the raw image directly, some frameworks alternate using a simplified proxy. For example, Bai et al. estimate a "skeleton image" proxy to isolate strong edges for kernel estimation. Others focus on rapidly refining the spatial support (the bounding box) of the kernel first, which drastically restricts the feasible solution space and accelerates convergence (Anger et al.).

Why Joint Alternation Beats Sequential Pipelines

One might wonder: why not just use a deep network to estimate the kernel first, and then run a standard deblurring algorithm? The danger of such two-stage pipelines is error propagation. If the initial kernel estimate is even slightly wrong, those errors are amplified during image reconstruction, resulting in severe ringing artifacts.

As Heinimann et al. and Tang et al. argue, simultaneous or alternating estimation forces mutual consistency. An updated, sharper image estimate provides better edge information, which in turn allows for a more accurate kernel refinement in the next iteration (Ma et al.). This feedback loop corrects errors dynamically rather than compounding them.

The Modern Evolution: Alternating Deep Priors

Classical alternating minimization relied heavily on handcrafted mathematical priors, like Total Variation or $\ell_0$ gradients (Anger et al.). While effective for simple textures, these priors often struggle with complex real-world scenes or sensor non-linearities like overexposed, saturated pixels (the authors of Blind Deblurring for Saturated Images).

Modern research has revitalized alternating optimization by replacing these rigid mathematical regularizers with neural network architectures acting as implicit priors.

In the Deep Image and Kernel Priors (DIKP) framework introduced by Wang et al., two separate, randomly initialized neural networks are used to represent the latent sharp image and the degradation kernel. The system alternates updates to the weights of these networks to minimize the reconstruction error. Because convolutional networks inherently resist generating noise or trivial Dirac deltas, the network structure itself acts as a powerful regularizer.

Similarly, Tran et al. and Zhang et al. alternate optimization within the learned latent spaces of generative models. By constraining the kernel and image updates to valid manifolds, they ensure that both components remain statistically and physically plausible throughout the entire iterative descent.

Key Foundational Work

Motion Deblurring and Super-resolution from an Image Sequence (1996) — contributed 1 concept elements to this direction.
Satellite and aerial image deconvolution using an EM method with complex wavelets (2002) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_alternating_optimization_for_blind_decon_c7d4e74f,
  title = {alternating optimization for blind deconvolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/alternating-optimization-for-blind-deconvolution/rdr_c7d4e74f3867415d81d0119aa2960c82},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_alternating_optimization_for_blind_decon_c7d4e74f_v1,
  title = {alternating optimization for blind deconvolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/alternating-optimization-for-blind-deconvolution/rdr_c7d4e74f3867415d81d0119aa2960c82/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake