Research map / Image generation / Image restoration
Cross-Scale Self-Supervised Blind Image Deblurring via Implicit Neural Representation
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 self-supervised blind image deblurring method that uses INR-based kernel and image generators, a cross-scale consistency loss derived from a QMF-based dyadic identity, and progressive coarse-to-fine training, without ground-truth images.
Problem
Blind image deblurring (estimating latent sharp image and blur kernel from a blurred image) without ground-truth images, where training of NN generators tends to overfit or converge to trivial solutions.
Approach
Re-parameterize the kernel and latent image with resolution-free INRs (SIREN kernel MLP, U-Net-style image generator with high-frequency encoding), add a cross-scale consistency loss based on an exact QMF-based relation between consecutive dyadic scales, and train progressively from the coarsest scale to the finest.
Key results
- On the Lai et al. synthetic dataset, the method reaches 25.16 dB / 0.879 average PSNR/SSIM, the best among all methods including supervised ones (MCEM 24.55 dB, VDIP 25.12 dB for self-supervised).
- On Köhler et al. dataset, reports 30.69 PSNR / 0.942 MSSIM, above the next best (Li et al. 30.26 PSNR, MCEM 0.940 MSSIM).
- Ablation: removing the cross-scale loss costs about 0.72 dB average PSNR (24.44 vs 25.16); training at a single scale costs about 1 dB; joint multi-scale loss instead of progressive training costs about 3 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cross-scale consistency loss (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 24.44 -> 25.16 | Helped |
| Use multi-scale training rather than training only at the original scale (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 24.14 -> 25.16 | Helped |
| Use progressive coarse-to-fine training rather than joint multi-scale loss (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 22.14 -> 25.16 | Helped |
| Use INR rather than CNN for the image generator (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 23.01 -> 25.16 | Helped |
| Use INR rather than MLP for the kernel generator (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 19.17 -> 25.16 | Helped |
| Use INR generators rather than MLP/CNN image-to-image generators (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 18.19 -> 25.16 | Helped |
| Generate multi-scale representations with INR grids rather than direct downsampling (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 20.89 -> 25.16 | Helped |
| Use four scales instead of three scales (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 25.16 -> 24.21 | Hurt |
| Use three scales instead of two scales (ablation) | INR-based blind deblurring · Lai et al. dataset | PSNR: 24.78 -> 25.16 | Helped |
| Use the proposed method rather than Restormer | Blind deblurring · RealBlur-J | PSNR: 28.96 -> 28.31 | Hurt |
| Use the proposed method rather than Restormer | Blind deblurring · RealBlur-R | PSNR: 36.19 -> 35.62 | Hurt |
| Use the proposed method rather than INIKNet | Microscopic deconvolution · Gaussian PSFs | PSNR: 37.32 -> 36.65 | Hurt |
Surprising findings
- Direct multi-scale training with a sum of losses across scales performed much worse (22.14 dB) than the progressive scheme, indicating the progressive schedule is critical.
- Supervised methods pretrained on GoPro perform poorly on the Lai et al. dataset (e.g., MIMO-UNet 16.36 dB), suggesting weak generalization.
Limitations
- Requires per-image training from scratch, so computation cost is high for large image collections.
- Only handles uniform blur because it relies on the convolution model; non-uniform blur is not addressed.
- Gap to supervised methods on real-world RealBlur dataset remains.
Experimental setup
Datasets: Lai et al. dataset Köhler et al. dataset Lai et al. real dataset RealBlur Microscopic datasets Levin et al. dataset GoPro (for pretrained baselines) BSD-D (for retraining baselines)
Benchmarks: Lai et al. dataset Köhler et al. dataset Lai et al. real dataset RealBlur-J RealBlur-R Microscopic deconvolution dataset Levin et al. dataset
Metrics: PSNR SSIM MSSIM Error Ratio running time memory usage
Baselines: SelfDeblur MCEM VDIP DEBID BlindDPS MPRNet Restormer MIMO-UNet DeblurGAN-v2 Kaufman & Fattal
Builds on: deep image prior SIREN SelfDeblur Fourier feature positional encoding
Closest papers by meaning
- Unsupervised Blind Image Deblurring Based on Self-Enhancement (CVPR 2024 · similarity 0.90)
- Neural Blind Deconvolution Using Deep Priors (CVPR 2020 · similarity 0.89)
- Deblurring Using Analysis-Synthesis Networks Pair (CVPR 2020 · similarity 0.88)
- Learning a Discriminative Prior for Blind Image Deblurring (CVPR 2018 · similarity 0.88)
- Uncertainty-Aware Unsupervised Image Deblurring With Deep Residual Prior (CVPR 2023 · similarity 0.87)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.86)
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.