In an article published in Coatings, researchers proposed a novel spectral reconstruction method that combined deep learning (DL) and the maximum a posteriori (MAP) technique to improve the degraded Raman spectrum.
The spectral reconstruction technique employed the MAP approach to reconstitute the spectra measured by Raman spectroscopy and to provide an initial estimate of the spectra. A convolutional neural network (CNN) was subsequently trained using the real Raman spectra and the preliminary estimated Raman spectra was measured by Raman spectroscopy to understand the mapping from both spectra. The technique helped to accomplish a better spectral reconstruction impact than merely utilizing the MAP method or a convolutional neural network.
Raman spectroscopy is usually hindered by instrument response function and noise, which increases measurement error and reduces the precision of substance identification. The novel spectral reconstruction method and other conventional techniques were employed to reconstitute the simulated and observed spectra measured by Raman spectroscopy. The findings indicated that the estimated Raman spectra reconstituted by the proposed approach were more accurate than those reconstructed using conventional techniques.
Eliminating Spectral Distortion in Raman Spectroscopy
A spectrometer is a simple optical detection tool that can gather the spectral data of the measured object. Spectrometers are currently some of the most crucial optical detection tools utilized widely in industries such as color measurement, gas composition analysis, agricultural medicine, and food safety.
The Raman spectrometer, which consists of a probe and a spectrometer, has recently undergone significant development due to its numerous advantages. It can qualitatively evaluate and distinguish between different material types and molecular structures with only a few samples. Raman spectrometers have been employed extensively in analytical chemistry and biology in the past ten years due to their non-destructive, quick, and real-time detection capabilities.
However, due to the effect of the instrument response function, often referred to as the bandwidth function, the obtained spectra measured by Raman spectroscopy frequently contain spectrum distortion. Thus, it causes considerable measurement error and further compromises the precision of substance identification when combined with noise. The problem can be fixed effectively by processing the measured Raman spectra using spectral reconstruction techniques.
Traditional spectral reconstruction techniques can be challenging to employ for recovering the narrowband spectra with accuracy in the spectra obtained by Raman spectroscopy because of the higher spectral distortion. Therefore, Raman spectroscopy spectral reconstruction presents more significant difficulties than conventional spectral deconvolution.
The present paper utilized the emerging technology of deep learning to achieve spectral reconstruction in the spectra obtained by Raman spectroscopy. Deep learning is an algorithmic program based on an artificial convolutional neural network to learn data features. With the help of the gathered datasets, deep learning trains a convolutional neural network. It creates a restoration design directly from input to output data without the need for complex mathematical calculations.
However, deep learning frequently necessitates several training datasets to accomplish better results. At present, deep learning is actively used in signal processing and classification, which has opened up new opportunities for developing spectral deconvolution techniques that can successfully recover Raman spectra.
The new spectral reconstruction technique restored the damaged Raman spectrum by combining the MAP technique and deep learning. First, the MAP approach was used to reconstitute the measured spectra obtained by Raman spectroscopy and to produce an initial estimate of the Raman spectra. Additionally, the mapping from the preliminary estimated spectra obtained by Raman spectroscopy to the real Raman spectra was learned utilizing the real Raman spectra and the primary estimated spectra obtained by Raman spectroscopy. Finally, a convolutional neural network was used to perform the training.
Designing the New Approach
The proposed spectral reconstruction system had three components: convolutional neural network architecture, training, and prediction. Eight learnable layers constituted the convolutional neural network architecture comprising three fully linked layers and five convolutional layers.
In a non-linear mapping between the primary estimated Raman spectra and the real spectra obtained by Raman spectroscopy, the convolution layers were used for feature extraction. For reconstructing the true Raman spectrum, the fully connected layers were used to synthesize the features that the convolution layers extracted.
Initially, a real Raman spectra dataset was established for the convolutional neural network training stage. The real Raman spectra dataset combined with the noise and instrument response function was used to replicate the measured Raman spectra dataset. Finally, the measured Raman spectra dataset was pre-processed using the MAP method and deep learning. The primary estimated Raman spectra were used as the convolutional neural network input dataset.
The proposed spectral reconstruction technique and some traditional spectral reconstruction techniques were employed to reconstitute the simulated and observed Raman spectra, respectively, to demonstrate the efficiency of the new spectral reconstruction method.
A tunable laser was used to obtain a linear spectrum that roughly corresponds to the Raman spectrometer's 1400 cm-1 position, and Gaussian function fitting was used to determine the bandwidth function of the Raman spectrometer. The recorded Raman spectra of caffeine were then reconstructed using three different approaches.
The reconstruction findings demonstrated an improvement in the resolution of the estimated spectra produced by the three approaches, particularly the ability to split the overlapping peak into multiple peaks. However, the reconstructed spectrum of the technique discussed in this paper was observed to be smooth in the flat region, in contrast to the estimated spectra produced by the two conventional methods, which showed apparent residual noise.
Improved Spectral Reconstruction Technique Proves Effective for Raman Spectroscopy
The present study demonstrated a novel spectral reconstruction technique that combined the MAP and deep learning approaches to restore the degraded Raman spectrum. The measured Raman spectra were first reconstructed using the MAP approach to obtain an initial estimated Raman spectrum.
Additionally, the mapping from the initial estimated Raman spectra to the real Raman spectra was learned by training a convolutional neural network with both the initial estimated Raman spectra and the real Raman spectra. The primary distinction between the new and traditional approaches was that the new approach created a mapping between the pre-processed and real spectra, resulting in a more effective spectral reconstruction performance than the conventional approaches or convolutional neural network alone.
The experimental findings demonstrated that the estimated Raman spectra reconstituted by the suggested method were more accurate than those reconstructed using conventional techniques.
Q. Zhou, Z. Zou, L. Han, 2022. Deep Learning-Based Spectrum Reconstruction Method for Raman Spectroscopy. Coatings. http://www.mdpi.com/2079-6412/12/8/1229