UMD's training-free speckle denoising AI clears coherent images

Speckle noise—a granular, multiplicative interference pattern—has long plagued coherent imaging systems like laser radar, optical coherence tomography, and synthetic aperture radar. Traditional denoising relies on averaging multiple frames, which fails for dynamic scenes. Now, a team from the University of Maryland led by Christopher Metzler has unveiled arXiv:2608.06574, "Implicit Neural Speckle Denoising," which removes speckle directly from noisy observations without any training data or clean references.

The method blends a spatiotemporal implicit neural representation with an aperture-aware maximum-likelihood formulation, explicitly modeling the spatial covariance of speckle tied to the imaging system's pupil. This enables it to adapt to arbitrary aperture geometries—circular, annular, Golay—simply by swapping the covariance model. To make optimization tractable, the team implemented a matrix-free pipeline using FFT-accelerated operators, stochastic approximations, and conjugate gradients. A blind holdout criterion automatically determines when to stop training, eliminating the need for ground-truth images. In simulated and lab experiments (including dynamic bars and toy scenes), the approach outperformed classical, unsupervised, and supervised baselines in spatial fidelity, temporal consistency, and robustness to varying speckle statistics. The paper includes nine supplementary videos demonstrating the results across different apertures and scenes.