Researchers have developed a method called GridPSF, which integrates a Zernike-parameterized forward regularizer into a video restoration network to mitigate atmospheric turbulence. Published in Applied Physics B, the study introduces an inspectable optical forward constraint to regularize training of the restoration network. The approach is designed to improve the restoration of video clips distorted by atmospheric refractive-index fluctuations, which cause time-dependent displacement and higher-order blur.
The core innovation of GridPSF is its use of a latent phase-distortion representation from the MambaTM host network. From the observed clip and the host’s latent field, the new branch predicts an 8×8 grid of 35 piston-free Zernike coefficients for each frame. These coefficients are used to render a field of unit-mass point-spread functions (PSFs) through a scalar Fraunhofer model. By blending these local convolutions, the method synthesizes a degraded version of the clean target video and compares it to the observed turbulent input, thereby enforcing physical consistency during training.
The study was evaluated using the ATSyn-dynamic benchmark, a synthetic dataset that simulates horizontal terrestrial imaging under various turbulence conditions. The dataset uses a Kolmogorov/Noll construction for Zernike statistics and defines turbulence severity based on a combination of aperture-to-Fried-parameter ratio, blur support, and geometric displacement. The researchers trained the model on 400 clips and tested it on a separate set of 100 videos, ensuring that the evaluation was performed on held-out data to measure generalization performance.
In terms of forward consistency, the fitted operator achieved a peak signal-to-noise ratio (PSNR) of 29.49 dB and a structural similarity index (SSIM) of 0.8882 on the held-out test set. This performance is notably higher than the identity floor of 25.54 dB, which represents the score obtained when the clean target is used directly as the prediction. The results indicate that the model successfully learns a forward action that reproduces the turbulent observation, although the authors caution that this does not necessarily identify the exact generating atmospheric wavefront.
Analysis of the fitted PSF field revealed that the mean second-moment widths spanned 4.20 to 4.25 pixels, which is close to the renderer’s zero-aberration width of 4.121 pixels. The principal-axis second-moment ratios remained near unity, suggesting that the fitted kernels are relatively isotropic. However, the singular-value participation rank of the PSF field was low, with one component retaining at least 95% of the squared energy. This concentration is attributed to a common renderer core rather than an intrinsically low-rank atmospheric field, highlighting the difference between the implemented operator and the physical phenomenon it approximates.
When comparing restoration performance, the validation-selected matched-host checkpoint pair showed a PSNR gain of +0.419 dB and an SSIM improvement of +0.0022 over the unmodified MambaTM host. However, the learned perceptual image patch similarity (LPIPS) increased by 0.0017, indicating a slight perceptual regression. The authors emphasize that this single-pair result is checkpoint-sensitive; in a broader sweep of 54 matched checkpoints, only 22 pairs showed a positive PSNR difference, and the unweighted mean difference was negative. This suggests that the observed gain is not stable across the entire training trajectory.
To assess whether the optical parameterization provides a specific advantage, the team conducted a five-seed comparison using a same-information convolutional neural network (CNN) control. The CNN received the same inputs as GridPSF but predicted the degraded sequence directly without the Zernike-based forward constraint. The seed-level PSNR differences averaged +0.0107 dB, with a 95% confidence interval that included zero. Similarly, comparisons with the unmodified host showed no detected average restoration gain. These controls suggest that the benefits of the GridPSF branch are not clearly distinguishable from those of a standard neural network with equivalent information access.
In a separate diagnostic test using replayed synthetic turbulence, the fitted field for each clip matched simulator-defined displacement better than fields from other clips or shuffled cells. However, the displacement magnitudes were underestimated, and the agreement in blur-width ranking and PSF shape remained poor. The researchers note that while the branch is inspectable, its coefficients are not calibrated atmospheric-wavefront estimates. The forward fit does not identify the generating wavefront, and the method should be viewed as an effective image-formation regularizer rather than a calibrated reconstruction of the complete turbulent PSF.
The study also highlights the dimensionless nature of the implemented pupil-to-image scale. The forward model does not include a separate geometric-warp stage, and the phase multiplier is a fixed value rather than a calibrated atmospheric scale. This means that the PSF widths are reported in renderer pixels without metric or angular calibration. The authors acknowledge that a calibrated extension would require known physical parameters such as wavelength, aperture diameter, and focal length, along with measured turbulence video to test transfer across different scenes.
Qualitative assessments on static printed-chart patterns from the Open Turbulent Image Set (OTIS) showed that the GridPSF and MambaTM outputs looked similar, with no clear visual advantage for the new method. The binary nature of the chart targets and the lack of moving natural scenes in this test limit its utility for evaluating real-world video restoration. The authors conclude that while GridPSF provides a physically structured forward constraint, its impact on restoration performance is modest and checkpoint-dependent. The method offers a step toward more interpretable atmospheric turbulence mitigation, but further work is needed to establish its advantages over standard deep learning approaches in broader benchmarks.
Beyond its core architectural definition, the training pipeline for GridPSF couples the auxiliary forward operator to the restoration network through a multi-term objective balanced by a scaling factor of 0.1. While the host objective incorporates Charbonnier reconstruction, perceptual similarity, ReBlurNet, and Kullback–Leibler constraints, the physical branch optimizes five distinct regularizers. The dominant term enforces forward consistency with unit weight, comparing the synthesized degraded frame directly against the turbulent input. To constrain the inferred coefficients, the objective applies a modal penalty weighted at 0.005, which penalizes modes inversely according to their theoretical radial-order variance using an exponent of eleven-thirds. Additional terms penalize temporal frame-to-frame coefficient shifts with a weight of 0.0005, total-variation spatial differences across grid cells with a weight of 0.0001, and four-neighbour point-spread function variations with a weight of 0.01. During joint training, the clean video target remains detached, meaning that gradient updates from this composite forward loss propagate exclusively into the host through its four-channel latent phase-distortion representation. Under this explicit penalty, the fitted coefficient energy exhibited a radial-order decay slope of -5.991, which is notably steeper than the -3.667 reference slope derived from atmospheric Kolmogorov theory.
The internal implementation of the coefficient prediction network also introduced several architectural characteristics that shaped the learned operator. The network models temporal evolution through three gated temporal blocks employing left-padded, length-seven depthwise convolutions, providing a nineteen-frame receptive field. However, training-time batch normalization pooled statistics across all sequence frames, preventing the resulting coefficients from remaining strictly causal. Furthermore, the tensor projection reshaped features directly without permuting the channel and cell axes, inadvertently mixing features across distinct grid cells. Spatial image degradation was executed through a gathering fusion model, where local convolved patches were blended across output coordinates using weights generated by a graph neural network lattice. In the primary experimental model, a global Fourier shift applied during convolution shifted each point-spread function four grid positions away in both horizontal and vertical directions with wraparound. This offset altered the local spatial correspondence between grid cells and their applied blur, meaning the resulting kernel maps reflect this structural shift rather than direct localized pupil properties.
Computational profiling and sensitivity diagnostics further clarified the behavior of the trained forward operator. Although GridPSF and the convolutional control had similar parameter totals at 157,831 and 156,377 parameters respectively, GridPSF incurred greater training latency. Its steady-state training interval averaged 0.444 seconds per iteration across three seeds, compared with 0.386 seconds for the convolutional branch and 0.338 seconds for the unaugmented host. When the fixed 29,000-step operator was tested across a frozen twenty-four-clip validation panel, changing its operational configuration produced notable variations in forward fit. Replacing circular boundary handling with reflect padding slightly increased forward consistency from 28.762 dB to 28.800 dB, while switching to a source-weighted scattering formulation reduced it to 26.963 dB. Correcting the four-cell spatial shift on these frozen weights caused forward performance to drop sharply to 25.844 dB. Modifying the phase multiplier across values from four pi to twelve pi shifted forward scores between 27.176 dB and 28.984 dB, peaking at ten pi.
To determine whether spatial variation or modal weighting drove restoration changes, the researchers also trained specific architectural variants. A layout-revised model correcting both the temporal reshape and grid-cell assignment yielded five-seed average differences of -0.0032 dB compared to the convolutional control and -0.0255 dB compared to the unmodified host, failing to show an advantage. Similarly, single-seed ablations comparing the standard architecture against a single-cell shared-kernel model and an unpenalized zero-modal-weight variant produced small differences of +0.0010 dB and -0.0022 dB respectively, with both confidence intervals spanning zero. These additional controls confirmed that the complex spatially varying grid and Kolmogorov modal regularization provided no detectable, isolated benefit to video restoration.
Subject of Research: Development of an inspectable Zernike-parameterized forward regularizer for synthetic atmospheric turbulence video restoration.
Article Title: GridPSF: an inspectable Zernike-parameterized forward regularizer for synthetic turbulence video restoration
Article References: Uppuluri, P., Saxena, N., Jape, S., Jacob, Z., & Aggarwal, V. (2026). GridPSF: an inspectable Zernike-parameterized forward regularizer for synthetic turbulence video restoration. Applied Physics B, 132(12), Article 119. https://doi.org/10.1007/s00340-026-08736-8
Image Credits: AI Generated
DOI: 10.1007/s00340-026-08736-8
Keywords: GridPSF, Zernike polynomials, atmospheric turbulence, video restoration, point-spread function, MambaTM, ATSyn-dynamic, forward regularizer, Fraunhofer diffraction, latent phase-distortion, peak signal-to-noise ratio, structural similarity, Kolmogorov/Noll model, anisoplanatic imaging, neural network regularization
News Source: Denise Maddox. (November 14, 2026). Zernike-parameterized forward regularizer yields no average synthetic turbulence restoration gain. Scienmag.



