Two-dimensional multifibre spectral image correction based on machine learning techniques

Abstract

ABSTRACT Owing to the limited size and imperfections of the optical components in a spectrometer, aberrations inevitably make their way into 2D multifibre spectral images in the Large Sky Area Multi-Object Fiber Spectroscopy Telescope (LAMOST), which leads to obvious spatial variation of the point spread functions (PSFs). However, if spatially variant PSFs are estimated directly, the large storage and intensive computational requirements result in the deconvolution spectrum extraction method becoming intractable. In this paper, we propose a novel method to solve the problem of spatial variation of the PSFs through image aberration correction. When CCD image aberrations are corrected, the convolution kernel can be approximated by only one spatially invariant PSF. Specifically, a novel method based on machine learning is proposed to calibrate the distorted spectral images. The method includes many techniques, such as total least squares (TLS) algorithm, self-supervised learning and multilayer feed-forward neural networksnetworks, and it makes use of a special training set sampling scheme combining 2D distortion features in a flat-field spectrum and calibration lamp spectrum. The calibration experiments on the LAMOST CCD images show that the proposed method is feasible. Furthermore, the spectrum extraction results before and after calibration are compared, and the experimental results show that the characteristics of the extracted 1D waveform are closer to those of an ideal optics system after image correction, and that the PSF of the corrected object spectrum estimated by the blind deconvolution method is nearly centrosymmetric, which indicates that our proposed method can significantly reduce the complexity of spectrum extraction and improve extraction accuracy.