Research map / Image generation / Image restoration
A Physics-informed Low-rank Deep Neural Network for Blind and Universal Lens Aberration Correction
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 blind, kernel-calibration-free lens aberration correction method that warps annular rings into stripes, factorizes PSFs with orthogonal nonnegative matrix factorization, and uses a decoupling network with pre-trained feature-domain Wiener deconvolution modules.
Problem
Blind, universal correction of non-stationary optical aberrations from diverse low-end camera lenses without PSF calibration or lens-specific training.
Approach
Combines rotational-symmetry annular warping (non-uniform blur to uniform stripes) with a low-rank ONMF basis of PSFs; a U-Net decomposes the blurry input into basis-degraded components, each restored by a pre-trained adaptive feature-domain Wiener deconvolution module, and fused with confidence-guided attention.
Key results
- Ours achieves 30.96 dB PSNR / 0.952 SSIM on synthetic data, surpassing blind baselines DeblurGANv2 (23.04 dB), MPRNet (28.67 dB) and Eboli's (29.42 dB).
- Non-blind DPIR (31.86 dB) and DWDN (31.78 dB) remain higher than the proposed method, which is described as comparable to non-blind approaches.
- Annular partitioning with 9 bases gives 30.16 dB versus 27.77 dB for grid partitioning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use annular partitioning instead of grid partitioning with 9 bases (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 27.77 -> 30.16 | Helped |
| Use 5 annular rings instead of 3 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 29.12 -> 30.96 | Helped |
| Increase annular rings from 5 to 7 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 30.96 -> 30.99 | No clear effect |
| Increase annular rings from 7 to 9 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 30.99 -> 30.91 | Hurt |
| Increase PSF bases from 5 to 7 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 28.14 -> 29.12 | Helped |
| Increase PSF bases from 7 to 9 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 29.12 -> 30.96 | Helped |
| Increase PSF bases from 9 to 11 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 30.96 -> 31.01 | No clear effect |
| Increase PSF bases from 11 to 13 (ablation) | Ours · Synthetic lens aberration dataset | PSNR (dB): 31.01 -> 31.00 | No clear effect |
Limitations
- Assumes sufficient exposure and high pixel count; does not model realistic noise or downsampling.
- Performance trails the best non-blind methods (DPIR, DWDN) on synthetic benchmark.
- More bases increase training time and decomposition ill-posedness.
Experimental setup
Datasets: Flickr2K Zemax-simulated PSF dataset from ZEBASE lenses
Benchmarks: Flickr2K-based synthetic test set real captured images from compact cameras
Metrics: PSNR SSIM
Baselines: DeblurGANv2 MPRNet Eboli's DPIR DWDN
Builds on: Orthogonal Non-negative Matrix Factorization DWDN Yue et al. annular warping
Closest papers by meaning
- A Physics-Informed Blur Learning Framework for Imaging Systems (CVPR 2025 · similarity 0.83)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.82)
- Exploiting Deblurring Networks for Radiance Fields (CVPR 2025 · similarity 0.82)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.82)
- Adaptive Camera Sensor for Vision Models (ICLR 2025 · similarity 0.81)
- All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising (CVPR 2025 · similarity 0.81)
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.