Research map / Image generation / Image restoration
Robust Blind Text Image Deblurring via Maximum Consensus Framework
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a maximum consensus framework combined with ADMM and half-quadratic splitting to jointly deblur and denoise noisy blurry text images via sparsity priors on latent image and kernel.
Problem
Blind deblurring of text images when the blurry observation is additionally corrupted by substantial outlier noise, which existing sparsity-based text deblurring methods ignore.
Approach
Reformulates text deblurring as maximum consensus (inlier maximization) with a slack variable for outliers, and solves the resulting l0-constrained problem with ADMM outer loop and HQS inner updates, plus residual deconvolution refinement in a coarse-to-fine pyramid.
Key results
- Outperforms sparsity-based methods (Pan et al. 2016a, Li et al. 2023) and noise-handling method Zhong et al. 2013 in latent PSNR and kernel similarity on synthetic noisy text images.
- Beats Chen et al. 2020 notably when outlier ratio is under 10%; comparable on the 50% outlier case in visual results.
- Convergence shown empirically over 120 instances; intensity-for-latent and gradient-for-kernel data terms performed best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the maximum-consensus framework with ADMM and HQS for noisy text image deblurring | Proposed method · Synthetic text images with added noise | PSNR | Helped |
| Use the maximum-consensus framework with ADMM and HQS for noisy text image deblurring | Proposed method · Synthetic text images with added noise | kernel similarity | Helped |
| Use the proposed method instead of sparsity-based text deblurring methods | Noisy text images · Representative text blurry images | visual deblurring quality | Helped |
| Use the proposed method instead of Chen et al. (2020) | Noisy text images with a low proportion of outliers · Representative text blurry images, 10% outliers | visual deblurring quality | Helped |
| Use the proposed method instead of Chen et al. (2020) | Noisy text images with a high proportion of outliers · Representative text blurry images, 50% outliers | visual deblurring quality | No clear effect |
| Use image intensity for the data-fidelity term when estimating the latent image and image gradient when estimating the kernel (ablation) | Proposed method · Synthetic noisy text images | kernel similarity | Helped |
| Vary the data-fidelity formulation between intensity and gradient in latent-image and kernel estimation (ablation) | Proposed method · Synthetic noisy text images | objective energy | Mixed |
| Vary γL across the tested range (ablation) | Proposed method · Synthetic noisy text images | kernel similarity: 10−5 to 0.01 | No clear effect |
| Vary γK across the tested range (ablation) | Proposed method · Synthetic noisy text images | kernel similarity: 0.02 to 5 | No clear effect |
| Vary σ across the tested range (ablation) | Proposed method · Synthetic noisy text images | kernel similarity: 0 to 2 | No clear effect |
Surprising findings
- Chen et al. 2020, a noise-handling method, was reported as comparable to the proposed method at high outlier ratios but weaker at low outlier ratios, which the authors attribute to its design for natural images.
Limitations
- Evaluated only on synthetically corrupted text images with Gaussian outliers.
- Exhaustive maximum consensus is intractable; the approximate solver offers only local convergence guarantees.
- Does not address saturated images, natural scenes, or non-uniform deblurring.
Experimental setup
Datasets: Levin et al. 2009 blur kernels 15 text images from Cho, Wang, and Lee 2012 (visual examples)
Benchmarks: synthetic noisy blurry text images (15 images x 8 kernels, 120 instances) Cho et al. 2012 text examples (visual)
Metrics: PSNR kernel similarity
Baselines: Zhong et al. 2013 Pan et al. 2016a Chen et al. 2020 Li et al. 2023
Builds on: Pan et al. 2016 l0-regularized intensity and gradient prior ADMM half-quadratic splitting maximum consensus
Closest papers by meaning
- Noise-Blind Image Deblurring (CVPR 2017 · similarity 0.87)
- Blind Image Deblurring Using Dark Channel Prior (CVPR 2016 · similarity 0.87)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.84)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.83)
- Unsupervised Blind Image Deblurring Based on Self-Enhancement (CVPR 2024 · similarity 0.82)
- Improving Image Restoration through Removing Degradations in Textual Representations (CVPR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.