Moving Beyond Spatially Invariant Blur Assumptions

39 concepts42 papers1 proposal

In classical image restoration, the process of blurring is typically modeled as a convolution: a sharp image is convolved with a single "blur kernel" (or Point Spread Function) to produce the blurry observation. This model relies on the spatially invariant blur assumption, which posits that every pixel in the image has been degraded by the exact same motion path or defocus level.

While mathematically convenient, this assumption is a simplification that frequently fails in real-world photography. Modern research is increasingly focused on breaking this constraint to handle the complex, non-uniform degradations found in dynamic scenes.

The Mathematical Convenience of Uniformity

The primary reason the spatially invariant assumption persists is computational efficiency. As Farsiu et al. note, assuming a space-invariant blur allows for fast mathematical decompositions, particularly in the frequency domain. When the blur is uniform, restoration can be performed using Fast Fourier Transforms (FFTs), which are significantly faster than processing pixels individually.

Many foundational techniques rely on this speed. For instance, Chen et al. utilize the assumption to develop efficient recursive filters for blind restoration. Similarly, the authors of Deep Wiener Deconvolution and FDN leverage the efficiency of Wiener deconvolution and FFTs to achieve high-speed restoration. However, as the authors of FDN acknowledge, these frequency-domain techniques struggle to generalize to complex degradations like rotational motion or depth-of-field effects.

Why the Assumption Fails

In practice, blur is rarely uniform across a frame. Several physical factors introduce spatial variance:

  • Camera Rotation: If a camera rotates (panning or tilting) during exposure, pixels further from the axis of rotation will blur more than those near it. Shao et al. and Yang et al. point out that standard models fail to account for this rotational motion, which produces varying blur patterns across the image.
  • Scene Depth: Objects at different distances from the lens experience different levels of defocus. Zhu et al. and Gong et al. emphasize that a single global kernel cannot accurately model scenes where depth-of-field variations are present.
  • Independent Object Motion: In dynamic scenes, a moving car will have a different blur kernel than the stationary background. Li et al. note that while patch-based modeling can help, capturing the complex blending at the boundaries of these moving objects remains a significant challenge.
  • Lens Aberrations: Even in static scenes, real-world lenses often produce blur that changes from the center to the edges of the sensor. Delbracio et al. observe that global PSF estimation fails to capture these edge-dependent aberrations.

Strategies for Non-Uniform Restoration

To move beyond the uniform blur limitation, researchers have proposed several strategies that trade mathematical simplicity for physical accuracy.

Mixture Models and Basis Kernels

One approach is to model the spatially varying blur as a combination of simpler, known kernels. Quan et al. address spatially varying defocus by modeling kernels as a linear mixture of Gaussians. This allows the model to adapt to different blur radii across the image. However, the authors note that these methods can still struggle when the actual blur exceeds the range of the predefined basis kernels.

Patch-Based and Local Estimation

Instead of one global kernel, some methods estimate kernels locally. Li et al. use patch-based deep learning to handle non-uniform motion. By breaking the image into smaller regions, the model can approximate local invariance, though this introduces difficulties in maintaining consistency across patch boundaries.

Physics-Informed Priors

In specialized environments like long-range imaging, the blur is caused by atmospheric turbulence. Mao et al. describe this as "anisoplanatic" conditions, where the blur is inherently non-uniform. They address this by integrating physics-informed priors that account for the space-time averaging of turbulent cells. Similarly, Sanghvi et al. rethink kernel modeling for photon-limited (low-light) conditions, where the uniform assumption is particularly damaging to image structure.

The Path Forward

The transition from spatially invariant to spatially varying models represents a shift toward "blind" deconvolution that is robust to the messiness of the real world. While the uniform assumption allowed the field to mature through efficient FFT-based solvers, the authors of RefDeblur and VBDeblurNet argue that the underlying physics-based estimation remains fundamentally limited until it can fully embrace the non-uniform nature of light and motion. Future work, as suggested by Pan et al., involves extending powerful priors—like $L_0$ regularization—directly into these non-uniform frameworks.

Key Foundational Work

Video-to-Video Dynamic Super-Resolution for Grayscale and Color Sequences (2006) — contributed 1 concept elements to this direction.
Efficient Recursive Multichannel Blind Image Restoration (2007) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 37 papers
  2. Version 1 · 2026-10-04 · 31 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_spatially_invariant_blur_assumption_58463368,
  title = {spatially invariant blur assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-invariant-blur-assumption/rdr_584633689b7446c8ab73478f8fbe5fce},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_spatially_invariant_blur_assumption_58463368_v2,
  title = {spatially invariant blur assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spatially-invariant-blur-assumption/rdr_584633689b7446c8ab73478f8fbe5fce/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake