Research on an automatic focusing method for the Large Sky Area Multi-Object Fiber Spectroscopic Telescope closed-loop positioning system based on deep learning

The positioning accuracy of optical fibers directly affects the observation efficiency of spectroscopic survey telescopes. The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST) adopts an optical fiber closed-loop positioning system based on a photogrammetric measurement system. A photogrammetric measurement system requires accurate identification of the optical spot position to achieve higher optical fiber positioning accuracy. However, the temperature difference within 1 day of the barrel of LAMOST will reach 40°C during the observation period. The change of air refractive index will lead to nonlinear deviation of the light propagation path of the optical spot, which will lead to the dispersion of the optical spot and cause a huge deviation in the position detection of the optical spot. To address this issue, we propose incorporating an automatic focusing system into the optical fiber closed-loop positioning system. Because each LAMOST’s fiber view camera (FVC) covers a field of view with hundreds of 320 μm-diameter optical fibers, fast and efficient focusing control for a large number of tiny optical fiber targets is highly challenging. Our solution implements a passive autofocus architecture combining an efficient image clarity evaluation method with a rapid focus search algorithm. Existing image sharpness evaluation methods are incompatible with FVC raw optical spot images, and traditional hill-climbing-based focusing algorithms require dozens of adjustments, significantly reducing efficiency. Our approach introduces an adaptive method for evaluating high-resolution optical spot image definition using the mean variance of the optical spot region, significantly reducing computational cost while improving threshold adaptability. Furthermore, we developed a focus search algorithm based on a hybrid convolutional neural network and recurrent neural network architecture. The algorithm predicts defocus distance from optical spot state sequence, achieving 95% single-shot focusing accuracy within the depth of field. It has been successfully applied to the LAMOST’s FVCs and also offers new methods for other focusing search methods based on passive autofocus for cameras.