Lightweight computational spectrometer enabled by learned high-correlation optical filters.

Area of Science:

• Spectroscopy
• Computational Imaging
• Machine Learning

Background:

• Neural network (NN) spectrometers offer high accuracy but demand significant storage.
• Conventional gradient projection for sparse reconstruction (GPSR) algorithms require less storage but yield lower spectral accuracy.
• GPSR performance is limited by optical filter properties, complicating design and fabrication.

Purpose of the Study:

• To develop a computational spectrometer with both high spectral reconstruction accuracy and reduced storage requirements.
• To leverage neural networks for improved optical filter design in spectral reconstruction.

Main Methods:

• Implementation of a hybrid approach, NN-GPSR, combining NN-learned filters with an optimized GPSR algorithm.
• NN-learned filters act as an encoder, exploiting high correlation for efficient data representation.
• An optimized GPSR algorithm functions as the decoder, enabling precise spectral reconstruction.

Main Results:

• The NN-GPSR method demonstrates high-precision spectral reconstruction capabilities.
• NN-GPSR significantly reduces the storage requirements compared to traditional NN spectrometers.
• The approach effectively utilizes prior knowledge from extensive image datasets for optical filter design.

Conclusions:

• NN-GPSR offers a superior alternative for computational spectroscopy in embedded systems.
• This hybrid method balances accuracy and efficiency, overcoming limitations of existing techniques.
• NN-learned filters are key to achieving high performance with reduced computational load.