INCREASING RESOLUTION OF A NOISED GAS-MIXTURE ABSORPTION SPECTRA BY NEURAL NETWORKS AND DATA PREPROCESSING METHODS
For improving low-resolution of absorption spectra, we have previously researched and developed several machine learning models based on different network architectures. All of them coped with the task with varying degrees of accuracy, but they all were learned and tested on modeled noise-free spectra. In the present work, we research the problem of improving low-resolution noised spectra in different ways. The first way is to build and learn models that will work with noised spectra as input data. The second way is to develop preprocessing algorithms that use Fast Fourier Transformation and Gaussian filtering methods that can decrease high-frequency noises. Also, convolutional neural networks with different architectures were trained and tested in solving such problems. New sequential ensemble architecture was proposed to improve the quality of results for spectral super-resolution.
The research was carried out with the support of a grant under the Decree of the Government of the Russian Federation No. 220 of 09 April 2010 (Agreement No. 075-15-2021-615 of 04 June 2021)
File with abstract
Tomsk State University, Leninа Ave., 36, 634050, Tomsk, Russian Federation
File with report