Blind Motion Deblurring

Updated 7 April 2026

Blind motion deblurring is a process that restores a sharp latent image from blurred observations by estimating unknown motion blur and noise effects.
Traditional pipelines use kernel estimation and deconvolution with hand-crafted priors, while modern methods employ CNNs, GANs, and transformers for end-to-end learning.
Benchmark datasets like GoPro and RealBlur, along with metrics such as PSNR and SSIM, are used to measure the progress and challenges in addressing spatially varying and dynamic blur.

Blind motion deblurring refers to the restoration of a sharp image from an observation degraded by unknown motion blur and, optionally, unknown noise, with no prior knowledge of the blur kernel. This involves recovering both the latent sharp image and the blur (point spread function, PSF), which may be spatially invariant or spatially varying across the image domain. The problem is highly ill-posed due to the non-uniqueness and high-dimensionality of possible solutions, particularly for spatially non-uniform and dynamic motion blurs encountered in real-world photography and imaging systems (Xiang et al., 2024).

1. Mathematical Formulation and Blur Models

In the classical imaging model, a blurry observation $B \in \mathbb{R}^{H\times W}$ is represented as

$B = I \otimes k + n,$

where:

$I \in \mathbb{R}^{H\times W}$ is the unknown sharp (latent) image,
$k$ is the blur kernel (spatially uniform or varying),
$\otimes$ denotes convolution, and
$n$ is additive noise (sensor or environmental) (Xiang et al., 2024).

Types of Motion Blur:

Linear or Global Blur: Uniform camera shake, simple object motion; a single global PSF $k$ .
Non-uniform Blur: Spatially varying due to local object motion, parallax, or camera rotation; requires local PSFs $k(x,y)$ .
Dynamic Blur: Both scene and camera undergo complex motion, leading to spatiotemporally varying blur kernels.

The estimation problem is more ill-posed for non-uniform and dynamic blurs, which cannot be addressed by standard single-kernel inversion or hand-crafted priors (Xiang et al., 2024).

2. Classical Approaches and Limitations

Traditional blind deblurring methods operate in two stages:

Kernel Estimation: Via MAP or EM frameworks, employing hand-crafted image priors (e.g., sparse gradients, dark channel, patch priors, $L_0$ or $L_2$ regularizers) (Shao et al., 2014, Anger et al., 2019).
Non-blind Deconvolution: Using the estimated $B = I \otimes k + n,$ 0, reconstruct $B = I \otimes k + n,$ 1 via inverse or regularized deconvolution (e.g., Richardson–Lucy, Wiener filter).

These classical methods are limited by:

Instability and ambiguity in kernel estimation for high-dimensional, real-world blur.
Insufficient modeling power of hand-tuned priors, often resulting in ringing artifacts, over-smoothing, and poor generalization to spatially varying blur (Xiang et al., 2024).

End-to-end learning approaches have overtaken kernel-inference-based pipelines, leveraging large-scale paired data to learn both image and blur priors implicitly (Xiang et al., 2024). Key architectural categories include:

3.1. Convolutional Neural Networks (CNNs)

Early CNNs: Two-stage (kernel estimation + fixed deconvolution) networks, e.g., Schuler et al., Sun et al.
Direct end-to-end CNNs: Modern U-Net variants (Nah et al., MPRNet, HINet) predict $B = I \otimes k + n,$ 2 directly from $B = I \otimes k + n,$ 3, often with multi-scale or multi-patch hierarchies (DMPHN) and iterative refinement (SRN, Park).
Losses: Combination of $B = I \otimes k + n,$ 4 pixel loss, perceptual loss (e.g., VGG), edge-aware, or frequency-domain losses.

3.2. GAN-Driven Approaches

Models such as DeblurGAN and CycleGAN-based approaches treat deblurring as a domain translation, using adversarial losses (Wasserstein or least-squares GAN), cycle-consistency, and perceptual feature supervision (Kupyn et al., 2017, Yuan et al., 2019, Saqlain et al., 2021).
Unsupervised cycle-consistent architectures: Learn from unpaired data by enforcing reversibility between blurred and sharp domains (Yuan et al., 2019, Saqlain et al., 2021).

3.3. Spatially-Adaptive and Deformable Models

Deformable convolution and aggregation: Modules such as ASPDC (Huo et al., 2021), MISC filtering (Liu et al., 2024), and CDCN (Tang et al., 2022) learn spatially variant sampling offsets and weights for direct motion compensation, enabling explicit modeling of non-uniform, dynamic motion blur.
Models like CDGNet (Wang et al., 2021) employ attention-guided two-branch architectures to separate and specialize processing for large- and small-blur components.

3.4. Recurrent and Transformer Architectures

Recurrent neural networks (ConvLSTM, multi-step refinement) achieve parameter-sharing and adaptive inference over spatial or temporal domains; e.g., SRN, Park, RNN-MBP.
Transformer-based deblurring: Global self-attention modules (Restormer, U-Former, FFTFormer) leverage long-range dependencies to reconstruct large kernels and fine textures, achieving top performance on synthetic and real datasets (Xiang et al., 2024).

3.5. Hybrid and Explicit Kernel Modeling

Kernel prediction networks (KPN): Estimate dense, pixel-wise motion blur kernels followed by learned non-blind deconvolution (Carbajal et al., 2023). Basis-based kernel coding and ADMM unrolling enable explicit model-driven restoration.
Generative latent kernel priors: Approaches such as GLKM (Ding et al., 12 Jul 2025) learn a generative kernel manifold via GAN pretraining, embedding kernel estimation in a low-dimensional latent space to address the initialization sensitivity and non-convexity of classical variational objectives.
Pixel discretization frameworks: Decompose the regression into a classification over "blur classes" then lightweight continuous correction for compute efficiency (Kim et al., 2024).

4. Evaluation Benchmarks and Metrics

Standardized datasets for quantitative and qualitative comparison include:

GoPro: Synthetic motion blur, 2,100 train / 1,110 test pairs. Network training/testing standard.
HIDE: Human-centered motion scenes.
RealBlur-J/R: Real camera shake, ∼4,500 pairs; crucial for real-world generalization assessment.
DVD: Dynamic video deblurring.

Metrics:

PSNR (dB): Peak signal-to-noise ratio.
SSIM: Structural similarity (range $B = I \otimes k + n,$ 5).
LPIPS: Learned perceptual similarity.

Typical state-of-the-art:

U-Net/Transformer models: PSNR $B = I \otimes k + n,$ 632.6–34.2 dB, SSIM $B = I \otimes k + n,$ 70.959–0.969 on GoPro.
Deformable/deconvolutional networks and MISC Filter set new state-of-the-art on RealBlur-R and RealBlur-J (e.g., PSNR 41.23 dB / SSIM 0.978 on RealBlur-R by MISC) (Liu et al., 2024).
GAN-based approaches yield visually sharper outputs but may slightly lag in PSNR due to adversarial loss bias (Kupyn et al., 2017, Yuan et al., 2019).

Benchmark Table (selected):

Method	GoPro PSNR	GoPro SSIM	RealBlur-J PSNR	RealBlur-J SSIM
FFTFormer	34.21	0.969	32.62	0.932
MISC Filter	34.10	0.969	33.88	0.938
DeblurGAN-v2	29.6	0.934	-	-

5. Strengths, Failure Modes, and Limitations

Strengths of deep blind motion deblurring:

Spatial adaptivity: Deep models can learn spatially variant features, directly handling real-world, dynamic blur (Xiang et al., 2024).
End-to-end optimization: Implicitly encode scene and motion statistics, surpassing classical priors especially for complex blur.
Transformer models and deformable modules excel at large-kernel, global-to-local restoration (Xiang et al., 2024, Liu et al., 2024).
Plug-and-play kernel priors: Generative latent manifolds improve initialization and stability over classical MAP solvers (Ding et al., 12 Jul 2025).

Limitations:

Domain generalization gaps: Models trained on synthetic blur yield 2–3 dB PSNR and $B = I \otimes k + n,$ 80.02 SSIM drop on real blur without dataset-specific fine-tuning (Xiang et al., 2024).
Resource trade-offs: Top-performing transformers, three-stage encoders (e.g., MCMS (Qiao et al., 2024)) or deformable models (CDCN (Tang et al., 2022)) impose high memory and compute costs, hindering mobile deployment.
Limits of explicit kernel modeling: Maximal recoverable blur is constrained by kernel support; extremely large, nonlinear, or saturated blur remains challenging (Carbajal et al., 2023, Ding et al., 12 Jul 2025).
Unpaired/unsupervised training: Remains underexplored but promising (Zhang et al., 2022); e.g., NeurMAP trains on unpaired blurry/sharp sets using a reblurring loop and adversarial priors.

6. Emerging Directions and Open Problems

Current research themes:

Unsupervised and domain-adaptive learning to bridge synthetic/real data gaps, including cycle-consistency (Yuan et al., 2019, Saqlain et al., 2021, Zhang et al., 2022).
Real-time and efficient architectures: Lightweight models (e.g., Ghost-DeblurGAN, SegDeblur-S (Kim et al., 2024)) leverage discretization and compact backbones for on-device deployment.
Diffusion priors: Probabilistic generative models for deblurring (e.g., DPS-GLKM (Ding et al., 12 Jul 2025)).
Hybrid physics-driven networks: Integrate explicit degradations (motion flow, light field modeling (Srinivasan et al., 2017), local kernels) with learning-based priors to increase interpretability and robustness.
Scene/semantic integration: Networks that jointly solve deblurring and high-level recognition/segmentation, or learn to leverage temporal structure for videos (Niu et al., 2021, Wieschollek et al., 2017).

Unresolved challenges include the generalization to truly unstructured real blur, robust handling of extreme spatial non-uniformity, reduction of model complexity, and accurate per-pixel uncertainty quantification for downstream vision tasks (Xiang et al., 2024, Ding et al., 12 Jul 2025).