A Convolution Neural Network Based Algorithm for More Accurate Spectrum Reconstruction of Miniaturized Spectrometers

To address the optimization challenges related to device count and responsivities in miniaturized spectrometers, we introduce an optimized strategy based on convolution neural networks (CNNs). This approach employs nonlinear techniques to mitigate issues arising from excessive similarity within the response matrix. Concurrently, we incorporate a matrix similarity computation method, superseding the conventional approach of resolving underdetermined equation sets, which enhances the precision of spectral reconstruction. The experiments demonstrate that under identical conditions, our method achieves an average central wavelength deviation of 1.2 nm for spectra with a single peak, which is lower than that of compressed sensing (CS) at 1.9 nm and CNN at 2.1 nm. The mean squared error (MSE) for reconstructing spectra with double peaks is 4.89e-3, representing an order of magnitude reduction compared to other algorithms, exhibiting unique advantages in both image quality and parameters.