Spatially Varying Motion Deblurring

97 concepts83 papers1 proposal

Imagine taking a photograph from the window of a moving train. The distant mountains appear perfectly crisp, but the guardrail right next to the tracks is reduced to a smeared, unrecognizable streak. If you attempt to fix this image using a traditional deblurring algorithm, you will quickly run into a fundamental mathematical wall.

Traditional image restoration assumes that blur is spatially invariant (uniform). It models degradation as a single global operation: the entire image is convolved with one uniform blur kernel. However, real-world motion blur is almost always spatially varying (non-uniform). When a scene contains independent moving objects, camera rotation, or variations in depth, every pixel effectively undergoes a different degradation process. Applying a single global correction kernel to such an image creates a destructive mismatch: sharpening the blurred foreground will violently distort the already-sharp background, leaving behind ugly "ringing" artifacts.

To resolve this, modern computational photography has shifted toward architectures and models that can adapt to localized, heterogeneous degradation.

The Roots of Spatial Variance

To understand why uniform deblurring fails, we have to look at how non-uniform blur is generated in a camera sensor. Spatially varying blur typically arises from three distinct physical phenomena:

  1. Varying Scene Depth: As shown by Liyuan Pan et al. in Single Image Deblurring and Camera Motion Estimation with Depth Map, when a handheld camera shakes, the geometric displacement on the sensor depends heavily on how far an object is from the lens. Closer objects span a larger pixel displacement than distant ones during the same exposure window.
  2. Camera Rotation: Even in a completely static scene, if a camera rotates or pans during exposure, pixels do not move uniformly. As modeled by Yunpeng Li et al. in Generating sharp panoramas from motion-blurred videos, pixels further from the rotation axis trace longer trajectories across the sensor than those near the center.
  3. Independent Object Motion: In dynamic environments, multiple subjects move at different speeds and directions relative to the camera. Tae Hyun Kim et al. demonstrated in Dynamic Scene Deblurring that this creates distinct localized blur patterns that require separate, isolated restoration paths for the moving foreground and the static background.

From Explicit Kernels to Motion Flow

Early attempts to solve non-uniform deblurring tried to explicitly estimate a unique mathematical blur kernel—or Point Spread Function (PSF)—for different regions of the image. For example, Wenqi Ren in Video Deblurring via Semantic Segmentation and Pixel-Wise Non-linear Kernel utilized semantic segmentation to identify object boundaries, applying distinct non-linear kernels to different semantic regions. Similarly, Tae Hyun Kim in Generalized video deblurring for dynamic scenes proposed jointly estimating pixel-level optical flow and latent sharp frames to track complex, non-linear pixel trajectories.

While mathematically rigorous, calculating unique 2D matrices for millions of pixels is computationally exhausting and highly sensitive to estimation errors. A major conceptual breakthrough occurred when researchers realized they did not need to calculate abstract mathematical kernels; instead, they could predict the physical pixel movement itself.

Dong Gong et al. introduced this intuition in From Motion Blur to Motion Flow: A Deep Learning Solution for Removing Heterogeneous Motion Blur. Instead of outputting a sharp image directly, their network estimates a dense motion flow field—a per-pixel vector map representing the direction and magnitude of the motion that caused the blur. By understanding the underlying physics of the motion, the model can invert the blur far more reliably, turning a highly ill-posed blind deblurring problem into a structured, non-blind restoration task.

Kernel-Free and Multi-Scale Processing

As deep learning matured, the field transitioned away from explicit motion estimation entirely, moving toward kernel-free end-to-end learning. The foundational challenge here is scale: a network needs to be able to fix massive, sweeping blurs (low-frequency degradation) while simultaneously preserving fine, sharp textures (high-frequency details).

To address this, Seungjun Nah et al. proposed a multi-scale architecture in Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring.

Blurry Input Image 
       │
       ├─► Downsample (Coarse Scale) ──► CNN Restoration ──┐
       │                                                   ▼
       ├─► Downsample (Medium Scale) ──► CNN + Coarse Feature Fusion ──┐
       │                                                               ▼
       └───────────────────────────────► Fine Scale CNN + Fusion ──► Sharp Output

By operating on an image pyramid, the coarser scales of the network possess a wide receptive field capable of resolving large-magnitude blurs. The outputs of these coarse layers are then passed as hints to finer scales, which gradually restore sharp edges and minute details. This multi-scale strategy allows standard, spatially invariant convolutional layers to effectively handle non-uniform blur by breaking the degradation down across different resolutions.

Dynamic and Spatially Adaptive Networks

While multi-scale networks are powerful, they are computationally heavy because they process every pixel with the same intensity, regardless of whether a region is severely blurred or already sharp. The latest generation of architectures solves this by using spatially adaptive layers that dynamically alter their behavior based on local image content.

One approach is to generate custom convolutional filters on the fly. Hyeonjun Sim et al. developed this concept in A Deep Motion Deblurring Network Based on Per-Pixel Adaptive Kernels With Residual Down-Up and Up-Down Modules, where the network outputs a unique set of local filter weights for every single coordinate in the image. If a pixel is located in a static background, the network generates an identity filter; if it sits on a fast-moving object, it generates an elongated filter aligned with the local blur trajectory.

To achieve this efficiency without the massive memory overhead of per-pixel filter generation, researchers have turned to attention mechanisms and deformable operations:

  • Deformable Convolutions: Kuldeep Purohit in Region-Adaptive Dense Network for Efficient Motion Deblurring demonstrated that instead of changing filter weights, a network can use deformable convolutions to dynamically shift the sampling offsets of its receptive field. This allows the network's internal filters to physically stretch and bend to match the non-uniform shape of the local blur.
  • Hybrid Attention: Maitreya Suin et al. designed a dual-mechanism approach in Spatially-Attentive Patch-Hierarchical Network for Adaptive Motion Deblurring. They paired a lightweight global self-attention module—which captures the broader context of camera shake across the entire frame—with pixel-adaptive local filtering to handle independent object motion.

By allowing the network architecture itself to be fluid and spatially variant, these modern systems mimic the complex, heterogeneous nature of real-world motion, restoring sharp details to dynamic scenes with remarkable precision and speed.

Key Foundational Work

An Adaptive and Heterodyne Filtering Procedure for the Imaging of Moving Objects (1987) — contributed 2 concept elements to this direction.
Single Image Motion Deblurring Using Transparency (2007) — 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_non_uniform_motion_blur_ec343a98,
  title = {non-uniform motion blur},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-uniform-motion-blur/rdr_ec343a98233d4e1eb97bbb7ceb880614},
  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_non_uniform_motion_blur_ec343a98_v1,
  title = {non-uniform motion blur},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-uniform-motion-blur/rdr_ec343a98233d4e1eb97bbb7ceb880614/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