Universal generalization and quantitative assessment of deep learning for imaging through scattering media

Deep learning (DL) provides a powerful data-driven approach to solving ill-posed inverse problems in optical imaging by directly learning a mapping from input patterns to target images. However, its performance is often hampered by a fundamental data dependency, which can bias the learning mapping and limit its generalization to unseen target types. Here, we overcome this limitation by introducing a spacetime-homogeneous speckle pattern dataset for training, which enables universal generalization across diverse datasets. The inherent complexity of the scattering media imaging mapping, coupled with the controllable feature scale of input speckle patterns, provides an ideal benchmark to systematically dissect the learning-to-simulate capabilities of neural networks. We rigorously investigate the influence of key factors, including the number of independent spatial modes, training data size, and network capacity (depth and number of convolutional filters of a U-Net), on the network’s simulation performance. Our findings reveal that the network structure defines the upper bound of its learning-to-simulate capability, while the training data size determines how effectively this capacity is utilized. Furthermore, we propose and demonstrate a novel, quantitative framework to assess a network’s learning-to-simulate capability based on the complexity of the targets it can successfully reconstruct. This work not only addresses the critical challenge of cross-dataset generalization but also provides a fundamental principle for designing and evaluating DL models in computational imaging and beyond.