Blind Noisy Image Deblurring Using Residual Guidance Strategy¶

Conference: ICCV 2025 arXiv: N/A (CVF Open Access only) CVF: Paper Link | PDF Code: N/A Authors: Heyan Liu, Jianing Sun, Jun Liu (Corresponding), Xi-Le Zhao, Tingting Wu, Tieyong Zeng Affiliations: Northeast Normal University, University of Electronic Science and Technology of China, Nanjing University of Posts and Telecommunications, BNU-HKBU United International College Area: Image Restoration Keywords: Blind Deblurring, Noise Robustness, Residual Guidance, Image Pyramid, Blur Kernel Estimation

TL;DR¶

This paper proposes a Residual Guidance Strategy (RGS) for coarse-to-fine blind image deblurring within an image pyramid framework. At each scale transition, the convolution residual from the adjacent coarser scale is denoised via a guided filter and used to correct the blurred input at the current scale. This approach significantly improves kernel estimation accuracy and restoration quality under high noise levels (σ=0.1), surpassing multiple deep learning methods without requiring any training.

Background & Motivation¶

Blind image deblurring is a classical ill-posed inverse problem: simultaneously recovering a sharp image and a blur kernel from a single blurry observation. The degradation model is given by $B = K \otimes L + N$, where $K$ is the blur kernel, $L$ is the latent sharp image, and $N$ is noise.

In practice, long-exposure photography introduces not only motion blur but also substantial noise. Existing methods—both traditional and learning-based—perform reasonably well under noise-free or mild-noise conditions, but suffer severe degradation in kernel estimation accuracy when noise increases (e.g., Gaussian noise with σ≥0.05). The root cause lies in a fundamental tension: deblurring requires preserving high-frequency information (edges, textures), whereas denoising requires suppressing it. Striking a balance between the two is the central challenge.

Through empirical observation, the authors identify an interesting phenomenon in coarse-to-fine pyramid frameworks: coarser-scale kernel estimates are more accurate (due to weakened noise and preserved dominant structures after downsampling), yet lack fine detail; finer-scale estimates are more precise but increasingly corrupted by noise. This observation directly motivates the proposed RGS.

Core Problem¶

How can coarser-scale, more reliable estimates be leveraged within a coarse-to-fine blind deblurring framework to guide finer-scale kernel estimation and suppress noise-induced degradation?

Method¶

Overall Architecture¶

The proposed method is a training-free traditional optimization approach based on an alternating iterative scheme grounded in a physical degradation model:

Input: A blurry and noisy image $B$
Image Pyramid Construction: Progressive downsampling produces multi-scale images $B_1, B_2, \ldots, B_n$ ($B_1$ finest, $B_n$ coarsest)
Coarse-to-Fine Iteration: Starting from the coarsest scale $B_n$, alternately optimize the intermediate sharp image $L$ and blur kernel $K$
Residual Guidance Correction: At each cross-scale transition, RGS corrects the input to the next finer scale to suppress noise
Non-blind Restoration: Given the final estimated kernel $K_1$, a non-blind method (Zhong et al.) is applied for the final deblurring

Key Designs¶

Alternating Optimization Subproblems:
- L-subproblem (Eq. 5): Given $K^t$, estimate the sharp image $L$ using the prior from Liu et al.—$\ell_0$ norm for salient edge selection plus image surface area regularization—solved via half-quadratic splitting.
- K-subproblem (Eq. 6): Given $L^{t+1}$, estimate the kernel $K$ using image gradients rather than raw pixel values (gradient domain is more stable), with $\ell_2$ regularization and efficient FFT-based solution.
Residual Guidance Strategy (RGS): The core contribution. At the transition from scale $i+1$ (coarse) to scale $i$ (fine):
- Compute the residual: $R_i = B_i - \text{Up}(L_{i+1} \otimes K_{i+1})$, i.e., the difference between the current-scale blurry image and the upsampled estimate from the coarser scale.
- The residual primarily contains noise along with some structural information.
- Apply a guided filter to the residual, using its Gaussian-smoothed version as the guidance image, yielding a denoised residual $\tilde{R}_i = g(R_i)$.
- Correct the current-scale input: $\tilde{B}_i = \text{Up}(L_{i+1} \otimes K_{i+1}) + \tilde{R}_i$.
- Substitute $\tilde{B}_i$ for the original $B_i$ in subsequent alternating optimization at this scale.
Comparison with Naive Guidance Strategy (NGS): An intuitive baseline directly applies a guided filter to the upsampled estimate. Experiments show that NGS is inferior to RGS, as direct filtering of the blurry image may destroy important structural details. In contrast, RGS filters only the residual, preserving the dominant structural content.

Loss & Training¶

No training is required. The optimization objective is: $$\min_{K,L} \|K \otimes L - B\|_2^2 + \lambda P_l(L) + \mu P_k(K)$$

where $P_l(L) = \|\nabla L\|_0 + \gamma \sum_{i,j}\sqrt{1 + |\nabla_{i,j}L|^2}$ is the sharp image prior, and $P_k(K) = \|K\|_2^2$ is the kernel regularization term.

Parameter settings: $\lambda = \gamma = 0.004$, maximum outer iterations $M=5$. The guided filter window size $w$ and smoothing parameter $s$ are also specified. The non-blind restoration stage incorporates NLM-based denoising enhancement for improved performance under high-noise conditions.

Key Experimental Results¶

All experiments are conducted under Gaussian noise with σ=0.1, which represents a challenging noise level.

Dataset	Metric	Ours	Prev. SOTA	Gain
Lai et al.	PSNR/SSIM/LPIPS	21.41/0.75/0.18	18.48/0.53/0.34 (Li)	+2.93dB
Zhang et al. (Face)	PSNR/SSIM/LPIPS	24.73/0.66/0.46	23.37/0.60/0.47 (Li)	+1.36dB
Levin et al.	PSNR/SSIM	Best	—	Significant margin
RealBlur (Real)	PSNR/SSIM/LPIPS	25.07/0.72/0.20	24.06/0.67/0.23 (Anger)	+1.01dB

RGS as a plug-in module for other methods (Table 2, σ=0.1):

Method	Original PSNR (Lai)	+RGS PSNR (Lai)	Gain
Dong et al.	17.88	19.45	+1.57dB

Ablation Study¶

NGS vs. RGS (Table 3): On the Lai dataset, RGS achieves PSNR 19.45 vs. NGS 18.96 (+0.49dB), and kernel accuracy MNC 0.66 vs. 0.60, with a clear advantage for RGS.
Stability: Across 100 independent runs, PSNR variance is only $5.1\times10^{-3}$ and SSIM variance is $6.93\times10^{-6}$, demonstrating high stability.
Mixed Noise (Gaussian-Poisson mixture, σ=0.05, λ=20): Residual-based denoising in RGS naturally adapts to different noise types.
Generality of RGS: Can be integrated into other blind deblurring methods (e.g., Dong et al.), yielding consistent PSNR and SSIM improvements across four datasets.

Highlights & Insights¶

Minimal yet effective: A training-free, purely traditional optimization approach that outperforms all deep learning methods under high noise, demonstrating that physical modeling combined with principled strategy design still holds considerable potential.
Precise core insight: The observation that coarser-scale kernel estimates are more accurate but coarser, while finer-scale estimates are more detailed but noise-corrupted, directly drives the method design.
Residual rather than image: Filtering the residual instead of the image directly avoids destroying structural information—a simple but critical design choice.
Universal plug-in: RGS can serve as a drop-in module to enhance any blind deblurring method built on a coarse-to-fine framework.
Strong robustness: Effective across unknown noise types (Gaussian, Poisson, mixed), as residual filtering is inherently insensitive to noise distribution.

Limitations & Future Work¶

No handling of dynamic scene blur: The method assumes a linear convolution degradation model (spatially invariant kernel) and is not applicable to dynamic scenes.
Fixed guided filter: The authors acknowledge that the guided filter could be replaced by a more powerful filter, representing a potential direction for improvement.
Replaceable non-blind restoration: The current non-blind restoration module could be substituted with more advanced alternatives.
Computational efficiency not discussed: Traditional iterative methods are generally slower than end-to-end deep learning approaches.
Only spatially invariant kernels evaluated: Real-world blur is often spatially varying.

vs. Traditional methods (Dong et al., Zhong et al., Anger et al.): These methods are effective under low noise but suffer catastrophic kernel estimation failure at high noise levels. RGS substantially improves robustness.
vs. DIP/VDIP/WDIP (deep image prior family): These unsupervised methods exploit network structural priors for blind deblurring but tend to overfit to noise under high-noise conditions. The proposed method does not suffer from this issue.
vs. End-to-end methods (DeblurGAN-v2, Zhang et al.): These methods rely on large-scale training data, generalize poorly to high-noise scenarios, and cannot produce explicit kernel estimates. The proposed method offers stronger interpretability.
vs. Lee et al. (ECCV 2024): While also targeting noise-robust blind deblurring, the proposed method achieves superior performance.

Further Inspirations:

Learnable residual filtering: Replacing the guided filter with a lightweight learnable network for adaptive filtering strength could be a valuable improvement direction.
Cross-task transfer: The residual guidance concept—using coarse-scale information to correct fine-scale inputs—can potentially generalize to other coarse-to-fine image restoration tasks such as super-resolution and dehazing.
Integration with diffusion models: Analogous cross-step residual guidance could be incorporated into the denoising process of diffusion-based restoration models.

Rating¶

Novelty: ⭐⭐⭐⭐ — The residual guidance concept is concise and elegant; while technically straightforward, the underlying insight is sharp.
Experimental Thoroughness: ⭐⭐⭐⭐ — Four datasets, multiple noise types, complete ablations, stability verification, and plug-in experiments; computational efficiency comparison is absent.
Writing Quality: ⭐⭐⭐⭐ — Motivation is clear, logic is coherent, and figures and tables are persuasive.
Value: ⭐⭐⭐⭐ — A training-free traditional method outperforming deep learning approaches in high-noise scenarios, with a generalizable plug-in module offering high practical utility.