A lightweight neural network compatible with high-correlation broadband optical filters in computational spectroscopy

Abstract

Computational spectrometer is powerful for portable in situ applications. To achieve better signal to noise ratio (SNR), broadband optical filters are more commonly introduced. As for conventional sparse signals reconstruction algorithms like gradient projection, its performance relies on the low correlation property of optical filters, which poses severe challenge for filters design and fabrication. In this study, we propose a lightweight neural network(NN) named HBOF-Net that can be compatible with high-correlation broadband optical filters to renovate the situation. The spectra reconstruction performs very well under a group of high correlated optical filters and multiscale training datasets. Firstly, we propose a group of various structure design of photonic crystal(PC) slabs and select every 16 to construct a spectrometer megapixel randomly. Then we extract different percentage of spectra from CAVE and ICVL dataset to build up multiscale training dataset groups. Afterwards, the training of the network has been carried out and the testing curve indicate there is no over fitting. A total of 16 trainings have been carried out and the results have been compared and discussed. Lastly, the experiment demonstrates that the mean square error(MSE) of the reconstructed spectra by HBOF-Net can reach 1E-6 by maximum correlation coefficient of 0.93 under multi-scale datasets which is 90%, 10%, 5%, and 3% of the entire dataset. This study successfully prove that NN can be compatible with high correlated optical filters, which is a significant leap in the development of computational spectrometers. It provides a good reference for computational spectrometers based on broadband and high-correlation optical filters.