Proceedings Article | 4 April 2023
KEYWORDS: Speckle pattern, Spectroscopy, Diffraction, Neural networks, Diffraction gratings, Speckle, Multimode fibers, Convolution, CCD cameras, Single mode fibers
The development of traditional spectrometer has been hindered by a variety of problems, such as mutual-limitations among instrument size, resolution and luminous flux. In recent years, computational optics has been developing vigorously, which overcomes the limitations of hardware with more and more abundant computing resources. The computational spectrometer utilizes the correlation between the speckle of the random scattering medium and the incident wavelength to restore the incident spectrum through the reconstruction algorithm. It has the advantages of simple structure, low requirements for alignment and adjustment, as well as easy miniaturization. However, in existing computational spectrometer schemes, if the incident light-field changes, the speckle pattern will no longer correspond to the incident wavelength uniquely. Therefore, most studies adopt single-mode optical fiber instead of multi-mode fiber as the light-guiding component, so as to control the incident light field. However, the single-mode fiber has a low optical energy transmission efficiency, which will degrade the sensitivity and other properties of the computational spectrometer. In this paper, we designed and fabricated a computational spectrometer based on random pixelated grating, and used a multimode fiber as the light guiding component. In order to eliminate the influence of incident light-field variation on spectral reconstruction, a dual-input deep-learning neural network model was proposed and implemented. The 1st-order and the 0th-order diffraction speckle patterns of the random pixelated grating were collected synchronously, which then were input into the neural network for network training. As a result, a spectral reconstruction model was obtained, which was insensitive to the incident light field. The experimental results show that the dual-input deep-learning neural network model can effectively suppress the influence of incident light field variation on spectral reconstruction. For the restricted spectrum, the spectral peak position accuracy can reach 99.4%, the spectral peak amplitude error range is 2.36% to about 7.69%, and the reconstruction speed is 90.19fps. This study could lay the foundation for the practical application of computational spectrometer.
