Learning from better simulation: creating highly realistic synthetic data for deep learning in scattering media

Obtaining the ground truth for imaging through the scattering objects is always a challenging task. Furthermore, the scattering process caused by complex media is too intricate to be accurately modeled by either traditional physical models or neural networks. To address this issue, we present a learning from better simulation (LBS) method. Utilizing the physical information from a single experimentally captured image through an optimization-based approach, the LBS method bypasses the multiple-scattering process and directly creates highly realistic synthetic data. The data can then be used to train downstream models. As a proof of concept, we train a simple U-Net solely on the synthetic data and demonstrate that it generalizes well to experimental data without requiring any manual labeling. 3D holographic particle field monitoring is chosen as the testing bed, and simulation and experimental results are presented to demonstrate the effectiveness and robustness of the proposed technique for imaging of complex scattering media. The proposed method lays the groundwork for reliable particle field imaging in high concentration. The concept of utilizing realistic synthetic data for training can be significantly beneficial in various deep learning-based imaging tasks, especially those involving complex scattering media.