Automating the detection of strong gravitational lenses in large-scale surveys using deep learning

This work presents advancements in automated strong gravitational lens detection using novel CNN-based architectures and techniques, focusing on reducing false positives - a major challenge in large-scale surveys. The DenseNet-121 ensemble outperforms previous ResNet models, achieving higher true-positive rates with fewer parameters. A novel Information Content (IC) metric is introduced for ranking lens candidates, improving classification performance. Integration with a U-Net segmentation algorithm (U-Denselens) enhances detection by filtering false positives based on segmentation scores, achieving substantial reductions in false positives while retaining true lenses, as demonstrated on KiDS and Euclid datasets.
To further enhance training, Denoising Diffusion GANs (DDGANs) generate large datasets of mock lenses. A balanced combination of real and synthetic data improves CNN's performance, addressing data scarcity in upcoming surveys. The U-DenseLens model applied to Euclid Early Release Observations (ERO) achieves lower false-positive rates and high detection efficiency, identifying 46 lens candidates across 16 fields, with contamination rates significantly lower than for KiDS data. Scaling projections suggest discovering over 5,500 strong lenses across the Euclid survey area.
These results underscore the potential of advanced CNN architectures, generative models, and segmentation algorithms to automate lens detection. By integrating multiple metrics, this framework offers a promising pathway to optimizing lens searches in large-scale astronomical datasets.