Optical data augmentation enhances out-of-distribution robustness of scattering medium image transmission

Deep learning has been successfully applied in imaging through scattering media, enabling direct recovery of the input light field from output speckle patterns. However, due to the high degrees of freedom in the light scattering process, current reconstruction methods can only work well in a trained data domain. Hence, achieving out-of-distribution (OOD) robustness in unseen scenes usually requires extensive experimental data collection. To overcome this limitation, we propose an optical image mixing (OIM) approach that introduces an efficient optical-domain data augmentation strategy to enhance model generalization under limited data conditions. By physically mixing optical images, OIM expands the effective training distribution without additional sample collection. We experimentally validate the proposed method on a multimode fiber (MMF) platform. With only 200 measured images and a 40-fold data augmentation by OIM, we achieve generalized reconstruction for 4096-pixel grayscale images. Compared to conventional models without OIM, we improve the image reconstruction fidelity by 26.2%. The results validate that OIM can serve as a plug-and-play module to enhance the generalization performance of existing reconstruction networks in computational imaging applications.