Smart Source Detection Framework for Lobster Eye Telescopes

Abstract Lobster eye telescopes are designed to capture X-ray emissions from celestial objects across a broad field of view. However, the unique optical system of these telescopes leads to a cruciform pattern on the detector plane, complicating the detection of celestial objects. While machine learning has driven the development of algorithms to address this, real applications still face challenges such as observation condition variations, often requiring manual adjustments of parameters in algorithms. Relying solely on 2D morphological features further restricts detection efficiency, as energy distribution and temporal variations are not considered. This paper introduces a novel source detection framework enhanced by deep neural networks, presenting three advancements: first, the use of a Swin Transformer to efficiently detect bright sources; second, a neural network to identify ordinary sources from a 3D data cube that integrates spatial and energy information; and third, an adaptive neural network branch within the 3D framework that automatically adjusts thresholds to classify true sources, accommodating observation condition variations. Applied to simulated data, our framework demonstrated a precision of over 95% and a recall rate exceeding 60% for targets brighter than 0.5 mCrab. Testing on 13,469 real observation images from the Wide-field X-ray Telescope on board the Einstein Probe resulted in an improvement in accuracy from 79% to 82% and a 27% increase in the number of detected sources compared to the current data-processing pipeline. The source detection framework and catalog obtained by the framework in a selected sky area are made accessible to the community for further applications.