Pre-processed speckle-correlation learning for imaging through scattering media

Images through complex scattering media have wide applications such as biomedical imaging and remote sensing, but it still encounters many difficulties. Although deep learning has achieved significant progress in this field, the reconstruction effect is largely affected by the high quality real-world data, which restricts its use in resource-limited situations. In this paper, we present a pre-processed speckle-correlation learning technique that effectively decreases the differences in data distribution under various scattering conditions. This enables the network to acquire more reliable physical prior knowledge, thus greatly enhancing the model's generalization ability. The experimental results show that, compared with the conventional speckle correlation learning method, the proposed approach obtains a better structural similarity index and peak signal-to-noise ratio on the test data and exhibits superior stability against unknown scattering media with different statistical distributions. Particularly, our method can recover real-world data by using only simulated data for training, completely freeing us from the dependence on real-world data. This offers a robust and feasible solution for scattering imaging in a scarce-resources environment.