Speckle-free self-supervised learning for scalable imaging through scattering media with unseen condition changes

Recently, deep learning (DL) methods have been extensively developed for imaging through scattering media. However, most learning methods rely on training with pairs of target-speckle data and lack integration with the physical imaging process. Here, we report a speckle-free self-supervised learning method that could be applied for scalable imaging through unknown random diffusers with unseen condition changes. Unlike traditional learning-based methods, our approach optimizes speckle reconstruction by leveraging the physical process of scattering imaging instead of fitting to "speckle-label" pairs. Our method models the scattered light field across potential scattering conditions to generate speckle patterns and extracts their correlation properties for model training and optimization. This eliminates the need for any pre-collected speckle patterns during network training. Our speckle-free method exhibits high reconstruction performance for imaging in unseen scattering conditions. We validate its performance across 440 unseen scattering conditions, including plane displacements, rotations, and combinations. Our method outperforms physics-informed learning approaches regarding reconstruction performance, consistency, and generalization ability in scalable imaging scenarios. This approach addresses the challenges of model generalization and extensive data collection for training, demonstrating its feasibility and superiority for imaging through unknown scattering media in novel scenarios.