A Neural Network Computational Spectrometer Trained by a Small Dataset with High-Correlation Optical Filters

Area of Science:

• Spectroscopy
• Optical Engineering
• Machine Learning

Background:

• Computational spectrometers offer portable in situ analysis but face challenges in filter design and reconstruction accuracy.
• Current methods require non-correlated filters and efficient algorithms for real-time applications.

Purpose of the Study:

• To develop a neural network computational spectrometer (NNCS) that performs accurately with a small dataset and high-correlation optical filters.
• To challenge the conventional reliance on large datasets and non-correlated filters in NNCS.

Main Methods:

• A novel distribution law was identified for large datasets to extract a small, representative training set.
• A fully connected neural network was designed for spectral reconstruction.
• Thin film filters were employed as the encoding layer.

Main Results:

• The NNCS demonstrated high performance in simulations and experiments using a small dataset and high-correlation filters.
• The trained neural network successfully reconstructed spectra, validating the approach.

Conclusions:

• This work presents a viable method for computational spectrometers utilizing high-correlation filters and limited training data.
• The findings offer a new paradigm for designing efficient and accurate portable spectrometers.