High-fidelity image reconstruction from dynamic nonlinear speckles mediated by chlorophyll solution using deep learning

Abstract

The information recovery of speckles is crucial in applications such as underwater exploration, satellite remote sensing, and biomedical imaging. The interplay of nonlinearity and scattering breaks the wavefronts and destroys the image formation. Previous studies have primarily focused on linear scattering models, leaving the reconstruction of images distorted by nonlinear perturbations largely unexplored. Here, we employ a convolutional neural network named the Attention U-Net to reconstruct images modulated by a chlorophyll solution, exhibiting self-defocusing nonlinearity. The proposed approach exhibits strong generalization capability for unseen object categories under nonlinear perturbation. Furthermore, by comparing the performance of models trained with single-power and multi-power mixed datasets, we reveal the relationship between reconstruction ability and nonlinear optical response. Our study establishes an effective connection between green-material-induced nonlinear modulation and deep learning-based optical decoding, offering promising potential for applications in optical information encryption and intelligent imaging systems.