Deep learning-enhanced spectral ghost imaging with accelerated and high-fidelity reconstruction.

Abstract

Ghost imaging is an indirect imaging method that utilizes the correlation properties of light to reconstruct the real-space image of an object. While originally developed in the spatial and temporal domains, its principles can be extended into the spectral domain by spatially dispersing broadband light and pseudo-randomly modulating its spectral components. In this work, we present a proof-of-concept implementation of computational spectral ghost imaging, combined with a deep learning framework to dramatically improve reconstruction fidelity and reduce measurement acquisition time. We introduce Spectral Ghost Imaging using Convolutional Neural Network (SGICNN), an encoder-decoder model trained exclusively on simulated data. Remarkably, SGICNN achieves high-fidelity image reconstruction and effective denoising of rudimentary spectral ghost images generated from as few as 8000 realizations, surpassing the accuracy of images constructed with 100,000 measurements. This corresponds to more than 10× reduction in acquisition time without compromising image quality. Our proposed approach is robust, straightforward, and holds strong potential for remote spectral sensing and high-resolution integrated spectrometers.