Single-shot multi-wavelength imaging over scattering media enabled by deep learning

Deep learning enables high-fidelity speckle imaging over scattering media in a single-shot manner. However, extending it to multi-wavelength scenarios remains challenging because coherent interference and dispersion effects complicate the training of deep the neural network (DNN). Herein, we propose a single-shot multi-wavelength scattering imaging approach leveraging a trained spectrum vision speckle network (SVSnet) to retrieve multi-wavelength images that are experimentally generated by multi-wavelength holography. The SVSnet is optimized to accelerate the training process and retrieve a multi-wavelength image with high fidelity simultaneously. Experimental results on three datasets demonstrate high retrieval fidelity over different multi-wavelength images. This approach also offers flexible wavelength adaptation, which can be generalized to other combinations of wavelengths.