AI assisted astronomical spectral processing based on distortion correction
TL;DR
An AI-assisted processing framework is presented that integrates subpixel distortion correction, deep learning-based noise identification, two-dimensional fiber-spot reconstruction, and spectral de blending in fiber-fed multi-object astronomical spectroscopy.
Abstract
In fiber-fed multi-object astronomical spectroscopy, each spectrograph records two-dimensional spectral images from hundreds of fibers. We present an AI-assisted processing framework that integrates subpixel distortion correction, deep learning-based noise identification, two-dimensional fiber-spot reconstruction, and spectral de blending. For spectral images acquired with the Fiber Arrayed Solar Optical Telescope (FASOT), a maximum Keystone displacement of 24.79 pixels was corrected to subpixel accuracy, with 87.4% of residuals below 0.10 pixels. The Convolutional Neural Network (CNN) classifier achieved 99.2% accuracy, and the You Only Look Once version 8-Small (YOLOv8-S) model reached 98.1% mAP@0.5 for pixel-level noise localization. Among 250 fibers from the Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST), 95.6% exhibited smooth single-peaked spots, while 4.4% showed central depressions indicative of coupling defects. For overlapped spectra, the recovered spectra closely matched the reference spectra, demonstrating effective suppression of inter-fiber crosstalk. A Flexible Image Transport System (FITS)-based processing platform was developed and successfully applied to LAMOST and FASOT data.