Spectral Reconstruction Algorithm for Tunable Detectors: A Learning Based Approach

Hyperspectral imaging, crucial for detecting and classifying earth’s surface elements, often requires complex optical equipment susceptible to environmental perturbations. A learning-based approach utilizing smart spectral sensors that considerably reduces hardware complexity while maintaining good resolution in spectral regions of interest is proposed in this paper. This method capitalizes on the spectral redundancy inherent in spectral sensors that have poor spectral resolution albeit the wavelength at which their responsivity curve peaks is continuously tunable. Central to our approach is the optimized superposition of photocurrents, corresponding to a family of responsivities that are tuned judiciously to appropriate wavelengths, along with their pairwise correlations, to render improved resolution spectrometry and enable continuously tunable sensing in spectral domain. The weights for this superposition are learnable parameters derived through a learning process utilizing a comprehensive library of spectra corresponding to various materials and filters. Through learning from this spectra library, the algorithm finds an optimal set of weights that can reconstruct any realistic spectrum as similar as possible. By employing a stochastic gradient-descent method to minimize mean-square error, the weights are generated to enable high-resolution estimation of the spectrum of the object being sensed. The method was applied to quantum-dot-in-a-well electrically tunable infrared photodetectors to estimate various spectra from the ECOSTRESS Spectral Library after being trained with 1,855,000 sample spectra. The method demonstrated a 17-fold improvement in target spectral estimation accuracy compared to prior methods. Additionally, when integrated with an Artificial Neural Network, the method achieved 96.38% accuracy in material classification. This innovation promises significant advances in high-resolution spectrometry and accurate material classification in various practical applications.