Overview
Research introduces a training-free framework designed for the denoising of speckle in dynamic coherent imaging. This method integrates a spatiotemporal implicit neural representation with an aperture-aware maximum-likelihood formulation. The objective is to recover dynamic, speckle-free imagery directly from noisy observations.
Research Context
Speckle noise is a fundamental limitation in coherent imaging. It manifests as multiplicative, spatially correlated noise that can obscure scene structure. The challenge of removing speckle noise is particularly pronounced in dynamic scenes, which do not benefit from conventional speckle averaging techniques. Traditional methods for speckle removal often struggle with dynamic scenarios due to the inherent motion and changing scene properties.
Approach
The proposed framework utilizes an implicit neural representation, which is spatiotemporal in nature. This representation is combined with a maximum-likelihood formulation that is aperture-aware. This formulation explicitly models the aperture-dependent spatial covariance of speckle. A key aspect of this design is its ability to adapt to arbitrary pupil geometries without requiring retraining.
The implementation of the framework incorporates several features to ensure practicality for realistic image sizes:
- A matrix-free approach.
- FFT-accelerated operators for computational efficiency.
- Stochastic approximations.
- Conjugate gradients for optimization.
Furthermore, the framework integrates a blind holdout criterion, which facilitates automatic early stopping during the optimization process. This criterion operates without the need for clean reference data.
Findings
Evaluations conducted using both simulated and laboratory results indicated several improvements:
- Improved spatial fidelity in the recovered imagery.
- Enhanced temporal consistency across dynamic scenes.
- Increased robustness to varying speckle statistics.
These improvements were observed in comparison to classical, unsupervised, and supervised baseline methods.
Why This Matters
The development of a training-free, implicit neural framework that effectively denoises speckle in dynamic coherent imaging addresses a significant limitation in the field. By explicitly modeling aperture-dependent speckle covariance and adapting to various pupil geometries without retraining, the approach offers a flexible solution for acquiring clearer images from noisy observations in dynamic scenarios where conventional methods are insufficient.