General Cost‐Effective Deep Learning Protocol for Reconstruction of Raman Spectrum in Handheld Spectrometers

ABSTRACT Raman spectroscopy is a non‐destructive optical detection technique providing fingerprints of molecular vibrations. For on‐site detection, the handheld spectrometers require high excitation laser power to overcome the Johnson noises in the detector, leading to deterioration of analyte molecules. In this work, a general deep learning algorithm of neural network integrating residual networks and U‐Net (i.e., ResUNet) is developed for reconstruction of Raman spectra from the low‐power handheld spectrometer. The deep learning model is trained by a predesigned Raman spectral dataset from only 30 substances covering the full detective wavenumbers of the spectrometer. Two criteria, termed as mean squared error (MSE) and structure similarity index measure (SSIM), are employed to evaluate the training performance. The denoising performance via the ResUNet model is superior to the traditional Savitzky–Golay filter and wavelet transform algorithms, with the MSE down to 0.003038 and the SSIM up to 0.7124, respectively. The F1‐score is further employed to evaluate the reconstructed Raman peaks with true positives, false positives, and false negatives, by which the optimized ResUnet model achieves 77.23%. The trained model is able to reduce the excitation laser power by 28‐fold in spectral acquisition and well‐reconstruct major Raman characteristic peaks, avoiding the potential damage of analyte molecules using handheld spectrometers. The present work paves a new way to upgrade the low‐cost handheld spectrometers achieving high sensitivity for on‐site detection.