Damped SDB inversion robustly recovers images from noisy rotated copies

ninamiolane · x · 2026-11-21

The companion paper introduces a damped inversion for MRA that regularizes to avoid division by zero, and derives the expectation and variance of the SDB under i.i.d. Gaussian pixel noise to define the regularization's noise floor. Experiments with N=20/50/100 noisy rotated copies show stable relative ℓ2 error for MNIST, Fashion-MNIST and cell images, outperforming jasp and cytd baselines.

Original post →

More from Research

Research channel →