Deep Learning Super-Resolution Spectrometer Based on Fiber Random Laser With Ultrahigh Spectral Purity.

The super-resolution technique based on random laser (RL) achieve the spectra that break through the frequency resolution limit of the original spectrometer. However, the speed of super-resolution spectrometer methods based on RL is limited by the time-consuming need to record many thousands of sub-resolution sparse spectral frames. Here, we propose a deep learning super-resolution spectrometer based on fiber random laser with ultrahigh spectral purity, obtaining super-resolution images from up to an 80% reduction in reconstruction time compared with what is usually needed. By coupling to a nested fiber microcavity, the decay rates of RL quasi-modes broaden, resulting in an excellent micro-nano light source for a super-resolution spectrometer showing high spectral purity, good directivity, and a miniature size. Based on this micro-nano light source, the sparse frames for reconstructing super-resolution spectra decreased threefold compared with that reported before. Furthermore, a convolutional neural network is demonstrated to recover the super-resolution spectra from an 80% smaller number of raw frames or an 80% smaller density of localizations. The drastic reduction in the acquisition time of the super-resolution spectrometer promotes the development of integrated, low-cost, high-resolution spectroscopy with a small footprint.