Blind Image Deblurring in Dynamic Scenes Using Physics-Informed Neural Networks
TL;DR
PI-NAFDeblur is presented, a physics-informed neural network that utilizes a coupling of a lightweight NAFNet restoration backbone and a differentiable reblur operator that enforces forward-model consistency during training to demonstrate that physics-informed regularization can help improve cross-dataset robustness under resource-constrained training.
Abstract
Blind motion deblurring of dynamic scenes is difficult due to the spatially-varying nature of the blur kernels, and the inverse nature of the problem. In this work, we present PI-NAFDeblur, a physics-informed neural network that utilizes a coupling of a lightweight NAFNet restoration backbone and a differentiable reblur operator that enforces forward-model consistency during training. The predicted per-pixel displacement field and blur resynthesis in the reblur branch provide a self-supervised regularization signal and yield an inferred latent blur parameterization as a by-product of restoration; because the reblur operator is sign-symmetric, this field is not validated against ground-truth motion and is therefore not claimed to be a physical motion estimate. PI-NAFDeblur, trained on a single 4 GB GPU, reaches 30.20 dB PSNR and 0.9041 SSIM on the full 1111-image GoPro test set with only 29.4 M parameters, which is 3.6–4.4 dB below recent 2024–2025 methods such as MISC Filter, EVSSM, MLWNet, XYScanNet, and AdaRevD; the contribution of this work is accordingly the feasibility of physics-informed deblurring under a 4 GB training budget together with a controlled analysis of the reblur prior, and not benchmark-leading accuracy. A controlled 50-epoch matched-seed ablation reveals that the physics-informed loss is neutral in-distribution (+0.004 dB) and yields small improvements on two of three out-of-distribution benchmarks (HIDE +0.092 dB, t = 20.14, p < 0.001; RealBlur-R +0.098 dB, t = 6.88, p < 0.001) that remain significant under a Bonferroni-corrected threshold of α = 0.0167, whereas the RealBlur-J difference (+0.022 dB, t = 2.36, p = 0.018) does not and is reported as non-significant. All statistics come from a single matched-seed pair, so they characterise consistency across test images rather than robustness across independent training runs. All settings are reported in a fully transparent fashion, and we demonstrate that physics-informed regularization can help improve cross-dataset robustness under resource-constrained training; a full reproducibility package (code, configurations, seeds, checkpoints, and per-image metrics) accompanies the paper.