Light spot mapping algorithm for optical fiber based on Siamese network

The integration of multiple optical fibers for directing starlight into telescope spectrographs is pivotal for spectral analysis in expansive multitarget spectroscopic surveys. Exceptional fiber positioning accuracy is needed for the attainment of high precision in stellar spectral acquisition. Current leading optical fiber positioning solutions often utilize back-illumination illuminating fibers and rely on photogrammetry systems to provide precise position feedback in a closed-loop configuration. While these photogrammetry-based methods are effective, they capture images lacking features essential for differentiating between the fiber positioner (FP) and individual light spots; this especially applies to the differentiation between the fiducial fiber positioner (FFP) and scientific fiber positioner (SFP) against a uniformly dark background. In this study, a one-dimensional convolutional neural network is used as a feature extractor to analyze the back-illuminated light spot images and FP operation data for the development a Siamese network model. By comparing the feature similarities between the back-illuminated light spot images and the operational target data within the Siamese network, a reliable mapping is established between each light spot and its corresponding FP. When applied to Large Sky Area Multi-Object Fiber Spectroscopy Telescope (LAMOST) observations, this methodology overcomes the pivotal challenges of mapping FFP light spots affected by focal plane rotation. This method further achieves complete mapping of SFP light spots in a heavily cluttered environment of scientific fibers. Our approach provides an important reference for fiber light spot mapping methodologies in prospective multitarget spectroscopic survey instruments.