Computational snapshot multi-spectral imaging using a Fourier imager neural network

We demonstrate a snapshot multi-spectral imaging framework that operates with a standard monochrome image sensor and a deep learning-based reconstruction model, eliminating the need for spectral filters or dispersive optics. By leveraging wavelength-dependent defocusing as a natural encoding mechanism, spectral information is computationally reconstructed using the multi-spectral Fourier Imager Network (mFIN). Our approach achieves 92.98% accuracy in spectral channel prediction and high-fidelity reconstructions from single-shot images. Future integration with diffractive optical elements may further enhance spectral encoding. This scalable and cost-effective technique offers a compelling alternative for biomedical imaging, remote sensing, and industrial quality control.