Physics-Guided Dataset Homogeneity Enables Universal Deep Learning Generalization in Scattering Media Imaging

Abstract Deep learning (DL) is widely used in computational imaging for problems lacking analytic solutions, yet its generalization across datasets remains a critical challenge. Conventional wisdom attributes this limitation to feature-prior mismatches, but through physics-guided analysis of imaging through scattering media, we reveal a more fundamental cause: DL networks learn approximations of the system’s true physical mapping (\(\:{T}⁻¹\)), constrained by the spatial-intensity distribution of training data. We demonstrate that enforcing the spacetime homogeneity—ensuring every point in the region of interest is equally and sufficiently trained—bridges the gap between learned mappings (\(\:M\)) and \(\:{T}⁻¹\). By optimizing training datasets (e.g., transforming MNIST digits into grayscale-augmented variants), we achieve unprecedented cross-dataset generalization: networks trained on simple digits successfully reconstruct complex face images. This physics-guided framework not only overcomes generalization barriers in scattering imaging but also establishes a universal principle for designing robust DL architectures. Our work repositions DL from data-driven approximation to physics-simulating computation, unlocking reliable deployment in real-world applications.