Research map / Probabilistic and scientific modeling / Differential equations and PDEs

A Physics-Informed Blur Learning Framework for Imaging Systems

CVPR 2025 · Probabilistic and scientific modeling › Differential equations and PDEs · approach: Physics-informed operator learning · 1 citations (OpenAlex)

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 physics-informed PSF learning framework that fits a wavefront-based PSF model (with a per-direction basis and curriculum learning from center to edge) to measured SFR, then uses the estimated PSFs to train deblurring networks.

Problem

Accurate, spatially varying PSF estimation for real imaging systems is hard: optical simulation needs proprietary lens designs, and existing learned or parametric models are either inaccurate or tied to specific systems.

Approach

A modified wavefront (Seidel-like) basis where each term affects SFR along only one direction (cos or sin terms), optimized via MLP surrogates with narrow field-of-view intervals and center-to-edge curriculum learning, plus a two-stage chromatic PSF shift estimation driven by chromatic area differences; requires no lens parameters.

Key results

• On Edmund lens #63762 at H=1 without noise, the proposed method reaches 50.156 dB PSNR versus 44.920 (Degradation Transfer) and 44.993 (Fast Two-step) in PSF accuracy.
• Deblurring with Restormer trained on estimated PSFs reaches 31.506 dB / 0.894 SSIM without noise versus 30.691 dB for Degradation Transfer and 30.340 dB for Fast Two-step.
• Ablation shows removing the wavefront basis, curriculum learning, or narrow-FoV optimization each lowers PSNR, most notably at larger field heights.

Measured effects reported in the paper

Surprising findings

• The Fast Two-step method produces consistent outputs across runs, while the optimization-based methods (including the proposed one) give variable results.

Limitations

• Chromatic aberrations in wide field-of-view images are not fully corrected.
• Recovering a PSF is inherently ill-posed due to information loss during image formation.
• Evaluation is on one lens/sensor setup in simulation and a custom-built device for real captures.

Experimental setup

Datasets: Flickr2K FiveK synthetic checkerboard patterns from Zemax PSF

Benchmarks: Flickr2K test split (100 images) synthetic checkerboard PSF accuracy (Edmund lenses #63762, #89752) real captures

Metrics: PSNR SSIM MUSIQ MANIQA

Baselines: Degradation Transfer Fast Two-step MPRNet Restormer FFTFormer

Builds on: Seidel aberration model Degradation Transfer Fast Two-step PSF estimation curriculum learning

Closest papers by meaning

• BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.84)
• Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos (NeurIPS 2025 · similarity 0.84)
• DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal (CVPR 2025 · similarity 0.84)
• BLURD: Benchmarking and Learning using a Unified Rendering and Diffusion Model (NeurIPS 2024 · similarity 0.84)
• BUFF: Bayesian Uncertainty Guided Diffusion Probabilistic Model for Single Image Super-Resolution (AAAI 2025 · similarity 0.83)
• Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.83)