Color imaging through thick scattering media with hybrid deep learning

Imaging objects obscured by optically thick scattering media remains a major challenge in biomedical imaging, industrial nondestructive evaluation, and remote sensing. This challenge becomes even more severe when accurate color recovery is required. This work introduces a hybrid deep learning framework for high-fidelity color imaging through strongly scattering media with optical thicknesses up to 33 transport mean free paths. The approach combines a convolutional neural network with an analytical model, enabling accurate reconstruction of an object in terms of shape, size, and intrinsic color. To overcome the usual requirement for very large training datasets, the model is trained efficiently using only 500 simulated three-channel measurements by incorporating a specialized preprocessing transform that promotes stable, rapid learning. The method is experimentally validated as a non-invasive, real-time imaging scheme under white-light illumination and is shown to be robust in dense, dynamic scattering environments. The proposed framework substantially outperforms the widely used UNet architecture, yielding approximately two orders of magnitude lower mean squared error, more than 2.5 times higher peak signal to noise ratio, a fourfold improvement in Pearson correlation, and more than 1.5 times enhancement in structural similarity index measure. These results demonstrate a robust, accurate, and generalizable solution for color imaging through highly scattering media, with strong potential for advanced biomedical imaging, environmental monitoring, and remote sensing applications.