GridPSF: an inspectable Zernike-parameterized forward regularizer for synthetic turbulence video restoration

GridPSF adds a Zernike-based forward constraint to a video restoration network. From the observed clip and the host’s latent field, it predicts an 8× 8 grid of 35 piston-free Zernike coefficients. A dimensionless Fraunhofer renderer blends unit-mass point-spread functions (PSFs) to fit the turbulent observation from the clean target. On held-out ATSyn-dynamic videos, forward consistency reaches 29.49 dB peak signal-to-noise ratio (PSNR) and 0.8882 structural similarity (SSIM). The averaged PSF field has low singular-value participation rank and nearly equal principal second moments. The zero-aberration renderer itself has a 4.121-pixel second-moment width. One validation-selected matched-host checkpoint pair gives + 0.419 dB PSNR and + 0.0022 SSIM, while learned perceptual image patch similarity (LPIPS) increases by 0.0017. This pair is checkpoint-sensitive. A five-seed comparison with a same-information CNN and the unmodified host finds no detectable average restoration gain. In replayed synthetic turbulence, each clip’s field better matches simulator-defined displacement than wrong-clip and shuffled-cell controls, but displacement is underestimated and blur-width and shape agreement remain poor. Forward fit does not identify the generating wavefront. The fitted branch is inspectable, but its coefficients are not calibrated atmospheric-wavefront estimates.