Related Experiment Video
Updated: Aug 15, 2026
13:31
High Speed Sub-GHz Spectrometer for Brillouin Scattering Analysis
Published on: December 22, 2015
Model-guided deep unfolding network for broadband reconstructive spectrometers with non-orthogonal aliasing
Optics Express
|August 14, 2026
Summary
A new model-guided deep unfolding network (MGDUN) improves reconstructive spectrometers by combining physical models with learned priors. This method enhances spectral reconstruction accuracy and stability, even with significant noise and aliasing.
Area of Science:
- Spectroscopy
- Optical Engineering
- Computational Imaging
Background:
- Reconstructive spectrometers offer high resolution in compact devices.
- Practical deployment is hindered by ill-posed inverse problems and non-orthogonal aliasing.
Purpose of the Study:
- To develop a novel method for accurate spectral reconstruction in compact devices.
- To overcome limitations of existing methods in handling aliasing and noise.
Main Methods:
- Proposed a model-guided deep unfolding network (MGDUN).
- Combined a forward physical model with learned priors for spectral reconstruction.
- Evaluated performance on various spectral profiles with non-orthogonal aliasing and noise.
Main Results:
- MGDUN demonstrated high performance across diverse spectral profiles, outperforming transmission matrix (TM) and deep learning (DL) baselines.
- Achieved accurate broadband reconstruction (20 nm window) with low error metrics (MAE/RMSE/SAM/APS).
- Maintained stable reconstruction under Gaussian, Poisson, and drift perturbations.
Conclusions:
- MGDUN offers a robust solution for continuous spectrum reconstruction under bandwidth constraints.
- The approach represents a new paradigm for designing deep learning-based algorithms in spectral reconstruction.
- Moves beyond black-box approaches by integrating physical models with data-driven priors.
