Compressive Sensing Spectroscopy Using a Residual Convolutional Neural Network.

Area of Science:

• Spectroscopy
• Signal Processing
• Machine Learning

Background:

• Compressive sensing (CS) spectroscopy enables compact spectrometer development.
• Spectrum reconstruction is a key challenge in CS spectroscopy.
• Existing methods like sparse recovery and standard CNNs have limitations.

Purpose of the Study:

• To propose a novel Residual Convolutional Neural Network (ResCNN) for spectrum reconstruction from compressed measurements.
• To evaluate the performance of ResCNN against existing spectral recovery techniques.
• To demonstrate the advantages of ResCNN in terms of accuracy, stability, and speed.

Main Methods:

• Developed a ResCNN architecture with learnable layers and residual connections.
• Trained the ResCNN using both synthetic and measured spectral datasets.
• Compared ResCNN performance against sparse recovery methods and standard Convolutional Neural Networks (CNNs) using RMSE and PSNR metrics.

Main Results:

• ResCNN demonstrated superior spectral recovery performance compared to existing methods.
• Achieved better average root mean squared errors (RMSEs) and peak signal to noise ratios (PSNRs).
• ResCNN showed stable reconstructions even under noisy conditions and faster convergence than standard CNNs.

Conclusions:

• ResCNN offers a powerful and efficient approach for spectral reconstruction in compressive sensing spectroscopy.
• The method eliminates the need for prior knowledge of sparsifying bases or spectral features.
• ResCNN advances the development of compact and high-performance spectrometers.